id	sid	tid	token	lemma	pos
fcis-25702	1	1	frontiers	frontier	NOUN
fcis-25702	1	2	in	in	ADP
fcis-25702	1	3	computing	computing	NOUN
fcis-25702	1	4	and	and	CCONJ
fcis-25702	1	5	intelligent	intelligent	ADJ
fcis-25702	1	6	systems	system	NOUN
fcis-25702	1	7	issn	issn	VERB
fcis-25702	1	8	:	:	PUNCT
fcis-25702	1	9	2832	2832	NUM
fcis-25702	1	10	-	-	SYM
fcis-25702	1	11	6024	6024	NUM
fcis-25702	1	12	|	|	NOUN
fcis-25702	1	13	vol	vol	NOUN
fcis-25702	1	14	.	.	PROPN
fcis-25702	2	1	9	9	NUM
fcis-25702	2	2	,	,	PUNCT
fcis-25702	2	3	no	no	INTJ
fcis-25702	2	4	.	.	NOUN
fcis-25702	2	5	3	3	NUM
fcis-25702	2	6	,	,	PUNCT
fcis-25702	2	7	2024	2024	NUM
fcis-25702	2	8	1	1	NUM
fcis-25702	2	9	dynamic	dynamic	ADJ
fcis-25702	2	10	layer	layer	NOUN
fcis-25702	2	11	skipping	skip	VERB
fcis-25702	2	12	for	for	ADP
fcis-25702	2	13	large	large	ADJ
fcis-25702	2	14	language	language	NOUN
fcis-25702	2	15	models	model	NOUN
fcis-25702	2	16	on	on	ADP
fcis-25702	2	17	natural	natural	ADJ
fcis-25702	2	18	language	language	NOUN
fcis-25702	2	19	understanding	understanding	NOUN
fcis-25702	2	20	tasks	task	NOUN
fcis-25702	2	21	and	and	CCONJ
fcis-25702	2	22	machine	machine	NOUN
fcis-25702	2	23	translation	translation	NOUN
fcis-25702	2	24	using	use	VERB
fcis-25702	2	25	reinforcement	reinforcement	NOUN
fcis-25702	2	26	learning	learning	NOUN
fcis-25702	2	27	wei	wei	PROPN
fcis-25702	2	28	xu	xu	PROPN
fcis-25702	3	1	1	1	NUM
fcis-25702	3	2	,	,	PUNCT
fcis-25702	3	3	*	*	PROPN
fcis-25702	3	4	,	,	PUNCT
fcis-25702	3	5	xiaodong	xiaodong	PROPN
fcis-25702	3	6	jin	jin	NOUN
fcis-25702	3	7	2	2	NUM
fcis-25702	3	8	1	1	NUM
fcis-25702	3	9	school	school	NOUN
fcis-25702	3	10	of	of	ADP
fcis-25702	3	11	humanities	humanity	NOUN
fcis-25702	3	12	,	,	PUNCT
fcis-25702	3	13	college	college	NOUN
fcis-25702	3	14	of	of	ADP
fcis-25702	3	15	science	science	NOUN
fcis-25702	3	16	and	and	CCONJ
fcis-25702	3	17	technology	technology	NOUN
fcis-25702	3	18	,	,	PUNCT
fcis-25702	3	19	ningbo	ningbo	PROPN
fcis-25702	3	20	university	university	PROPN
fcis-25702	3	21	,	,	PUNCT
fcis-25702	3	22	ningbo	ningbo	PROPN
fcis-25702	3	23	,	,	PUNCT
fcis-25702	3	24	zhejiang	zhejiang	PROPN
fcis-25702	3	25	,	,	PUNCT
fcis-25702	3	26	china	china	PROPN
fcis-25702	3	27	2	2	NUM
fcis-25702	3	28	school	school	NOUN
fcis-25702	3	29	of	of	ADP
fcis-25702	3	30	foreign	foreign	ADJ
fcis-25702	3	31	languages	language	NOUN
fcis-25702	3	32	,	,	PUNCT
fcis-25702	3	33	ningbo	ningbo	PROPN
fcis-25702	3	34	university	university	PROPN
fcis-25702	3	35	of	of	ADP
fcis-25702	3	36	technology	technology	PROPN
fcis-25702	3	37	,	,	PUNCT
fcis-25702	3	38	ningbo	ningbo	PROPN
fcis-25702	3	39	,	,	PUNCT
fcis-25702	3	40	zhejiang	zhejiang	PROPN
fcis-25702	3	41	,	,	PUNCT
fcis-25702	3	42	china	china	PROPN
fcis-25702	3	43	*	*	PUNCT
fcis-25702	3	44	corresponding	correspond	VERB
fcis-25702	3	45	author	author	NOUN
fcis-25702	3	46	:	:	PUNCT
fcis-25702	3	47	wei	wei	PROPN
fcis-25702	3	48	xu	xu	PROPN
fcis-25702	4	1	(	(	PUNCT
fcis-25702	4	2	email	email	NOUN
fcis-25702	4	3	:	:	PUNCT
fcis-25702	4	4	xuwei2@nbu.edu.cn	xuwei2@nbu.edu.cn	NUM
fcis-25702	4	5	)	)	PUNCT
fcis-25702	4	6	abstract	abstract	NOUN
fcis-25702	4	7	:	:	PUNCT
fcis-25702	4	8	large	large	ADJ
fcis-25702	4	9	language	language	NOUN
fcis-25702	4	10	models	model	NOUN
fcis-25702	4	11	(	(	PUNCT
fcis-25702	4	12	llms	llm	NOUN
fcis-25702	4	13	)	)	PUNCT
fcis-25702	4	14	demonstrate	demonstrate	VERB
fcis-25702	4	15	remarkable	remarkable	ADJ
fcis-25702	4	16	proficiency	proficiency	NOUN
fcis-25702	4	17	in	in	ADP
fcis-25702	4	18	various	various	ADJ
fcis-25702	4	19	natural	natural	ADJ
fcis-25702	4	20	language	language	NOUN
fcis-25702	4	21	processing	processing	NOUN
fcis-25702	4	22	(	(	PUNCT
fcis-25702	4	23	nlp	nlp	NOUN
fcis-25702	4	24	)	)	PUNCT
fcis-25702	4	25	tasks	task	NOUN
fcis-25702	4	26	.	.	PUNCT
fcis-25702	5	1	however	however	ADV
fcis-25702	5	2	,	,	PUNCT
fcis-25702	5	3	their	their	PRON
fcis-25702	5	4	extensive	extensive	ADJ
fcis-25702	5	5	size	size	NOUN
fcis-25702	5	6	,	,	PUNCT
fcis-25702	5	7	resulting	result	VERB
fcis-25702	5	8	from	from	ADP
fcis-25702	5	9	the	the	DET
fcis-25702	5	10	inclusion	inclusion	NOUN
fcis-25702	5	11	of	of	ADP
fcis-25702	5	12	billions	billion	NOUN
fcis-25702	5	13	of	of	ADP
fcis-25702	5	14	parameters	parameter	NOUN
fcis-25702	5	15	across	across	ADP
fcis-25702	5	16	multiple	multiple	ADJ
fcis-25702	5	17	layers	layer	NOUN
fcis-25702	5	18	,	,	PUNCT
fcis-25702	5	19	presents	present	VERB
fcis-25702	5	20	significant	significant	ADJ
fcis-25702	5	21	challenges	challenge	NOUN
fcis-25702	5	22	regarding	regard	VERB
fcis-25702	5	23	storage	storage	NOUN
fcis-25702	5	24	,	,	PUNCT
fcis-25702	5	25	training	training	NOUN
fcis-25702	5	26	,	,	PUNCT
fcis-25702	5	27	and	and	CCONJ
fcis-25702	5	28	inference	inference	NOUN
fcis-25702	5	29	.	.	PUNCT
fcis-25702	6	1	traditional	traditional	ADJ
fcis-25702	6	2	methodologies	methodology	NOUN
fcis-25702	6	3	such	such	ADJ
fcis-25702	6	4	as	as	ADP
fcis-25702	6	5	model	model	NOUN
fcis-25702	6	6	pruning	pruning	NOUN
fcis-25702	6	7	and	and	CCONJ
fcis-25702	6	8	distillation	distillation	NOUN
fcis-25702	6	9	are	be	AUX
fcis-25702	6	10	employed	employ	VERB
fcis-25702	6	11	to	to	PART
fcis-25702	6	12	decrease	decrease	VERB
fcis-25702	6	13	the	the	DET
fcis-25702	6	14	size	size	NOUN
fcis-25702	6	15	of	of	ADP
fcis-25702	6	16	these	these	DET
fcis-25702	6	17	models	model	NOUN
fcis-25702	6	18	,	,	PUNCT
fcis-25702	6	19	but	but	CCONJ
fcis-25702	6	20	these	these	DET
fcis-25702	6	21	techniques	technique	NOUN
fcis-25702	6	22	often	often	ADV
fcis-25702	6	23	result	result	VERB
fcis-25702	6	24	in	in	ADP
fcis-25702	6	25	a	a	DET
fcis-25702	6	26	compromise	compromise	NOUN
fcis-25702	6	27	on	on	ADP
fcis-25702	6	28	performance	performance	NOUN
fcis-25702	6	29	retention	retention	NOUN
fcis-25702	6	30	.	.	PUNCT
fcis-25702	7	1	in	in	ADP
fcis-25702	7	2	this	this	DET
fcis-25702	7	3	work	work	NOUN
fcis-25702	7	4	,	,	PUNCT
fcis-25702	7	5	we	we	PRON
fcis-25702	7	6	propose	propose	VERB
fcis-25702	7	7	a	a	DET
fcis-25702	7	8	novel	novel	ADJ
fcis-25702	7	9	framework	framework	NOUN
fcis-25702	7	10	that	that	PRON
fcis-25702	7	11	uses	use	VERB
fcis-25702	7	12	dynamic	dynamic	ADJ
fcis-25702	7	13	layer	layer	NOUN
fcis-25702	7	14	skipping	skip	VERB
fcis-25702	7	15	for	for	ADP
fcis-25702	7	16	different	different	ADJ
fcis-25702	7	17	samples	sample	NOUN
fcis-25702	7	18	to	to	PART
fcis-25702	7	19	accelerate	accelerate	VERB
fcis-25702	7	20	the	the	DET
fcis-25702	7	21	inference	inference	NOUN
fcis-25702	7	22	speed	speed	NOUN
fcis-25702	7	23	of	of	ADP
fcis-25702	7	24	llms	llm	NOUN
fcis-25702	7	25	.	.	PUNCT
fcis-25702	8	1	first	first	ADV
fcis-25702	8	2	,	,	PUNCT
fcis-25702	8	3	we	we	PRON
fcis-25702	8	4	add	add	VERB
fcis-25702	8	5	an	an	DET
fcis-25702	8	6	adapter	adapter	NOUN
fcis-25702	8	7	layer	layer	NOUN
fcis-25702	8	8	at	at	ADP
fcis-25702	8	9	each	each	DET
fcis-25702	8	10	transformer	transformer	ADJ
fcis-25702	8	11	layer	layer	NOUN
fcis-25702	8	12	to	to	PART
fcis-25702	8	13	predict	predict	VERB
fcis-25702	8	14	whether	whether	SCONJ
fcis-25702	8	15	to	to	PART
fcis-25702	8	16	skip	skip	VERB
fcis-25702	8	17	the	the	DET
fcis-25702	8	18	next	next	ADJ
fcis-25702	8	19	layer	layer	NOUN
fcis-25702	8	20	or	or	CCONJ
fcis-25702	8	21	not	not	PART
fcis-25702	8	22	,	,	PUNCT
fcis-25702	8	23	and	and	CCONJ
fcis-25702	8	24	we	we	PRON
fcis-25702	8	25	propose	propose	VERB
fcis-25702	8	26	layer	layer	NOUN
fcis-25702	8	27	skip	skip	VERB
fcis-25702	8	28	pretraining	pretraine	VERB
fcis-25702	8	29	to	to	PART
fcis-25702	8	30	recover	recover	VERB
fcis-25702	8	31	the	the	DET
fcis-25702	8	32	model	model	NOUN
fcis-25702	8	33	’s	’s	PART
fcis-25702	8	34	performance	performance	NOUN
fcis-25702	8	35	.	.	PUNCT
fcis-25702	9	1	second	second	ADJ
fcis-25702	9	2	,	,	PUNCT
fcis-25702	9	3	we	we	PRON
fcis-25702	9	4	propose	propose	VERB
fcis-25702	9	5	using	use	VERB
fcis-25702	9	6	reinforcement	reinforcement	NOUN
fcis-25702	9	7	learning	learning	NOUN
fcis-25702	9	8	(	(	PUNCT
fcis-25702	9	9	rl	rl	NOUN
fcis-25702	9	10	)	)	PUNCT
fcis-25702	9	11	to	to	PART
fcis-25702	9	12	optimize	optimize	VERB
fcis-25702	9	13	the	the	DET
fcis-25702	9	14	model	model	NOUN
fcis-25702	9	15	and	and	CCONJ
fcis-25702	9	16	design	design	VERB
fcis-25702	9	17	several	several	ADJ
fcis-25702	9	18	strategies	strategy	NOUN
fcis-25702	9	19	to	to	PART
fcis-25702	9	20	stabilize	stabilize	VERB
fcis-25702	9	21	the	the	DET
fcis-25702	9	22	training	training	NOUN
fcis-25702	9	23	.	.	PUNCT
fcis-25702	10	1	extensive	extensive	ADJ
fcis-25702	10	2	experiments	experiment	NOUN
fcis-25702	10	3	on	on	ADP
fcis-25702	10	4	four	four	NUM
fcis-25702	10	5	natural	natural	ADJ
fcis-25702	10	6	language	language	NOUN
fcis-25702	10	7	understanding	understanding	NOUN
fcis-25702	10	8	(	(	PUNCT
fcis-25702	10	9	nlu	nlu	NOUN
fcis-25702	10	10	)	)	PUNCT
fcis-25702	10	11	datasets	dataset	NOUN
fcis-25702	10	12	and	and	CCONJ
fcis-25702	10	13	three	three	NUM
fcis-25702	10	14	machine	machine	NOUN
fcis-25702	10	15	translation	translation	NOUN
fcis-25702	10	16	datasets	dataset	NOUN
fcis-25702	10	17	and	and	CCONJ
fcis-25702	10	18	ablation	ablation	NOUN
fcis-25702	10	19	studies	study	NOUN
fcis-25702	10	20	show	show	VERB
fcis-25702	10	21	that	that	SCONJ
fcis-25702	10	22	our	our	PRON
fcis-25702	10	23	method	method	NOUN
fcis-25702	10	24	achieves	achieve	VERB
fcis-25702	10	25	sota	sota	ADJ
fcis-25702	10	26	performance	performance	NOUN
fcis-25702	10	27	among	among	ADP
fcis-25702	10	28	layer	layer	NOUN
fcis-25702	10	29	skipping	skip	VERB
fcis-25702	10	30	methods	method	NOUN
fcis-25702	10	31	on	on	ADP
fcis-25702	10	32	llms	llm	NOUN
fcis-25702	10	33	.	.	PUNCT
fcis-25702	11	1	keywords	keyword	NOUN
fcis-25702	11	2	:	:	PUNCT
fcis-25702	11	3	large	large	ADJ
fcis-25702	11	4	language	language	NOUN
fcis-25702	11	5	model	model	NOUN
fcis-25702	11	6	;	;	PUNCT
fcis-25702	11	7	reinforcement	reinforcement	NOUN
fcis-25702	11	8	learning	learning	NOUN
fcis-25702	11	9	;	;	PUNCT
fcis-25702	11	10	sota	sota	NOUN
fcis-25702	11	11	performance	performance	NOUN
fcis-25702	11	12	.	.	PUNCT
fcis-25702	12	1	1	1	X
fcis-25702	12	2	.	.	X
fcis-25702	12	3	introduction	introduction	NOUN
fcis-25702	12	4	since	since	SCONJ
fcis-25702	12	5	chatgpt	chatgpt	NOUN
fcis-25702	12	6	and	and	CCONJ
fcis-25702	12	7	gpt	gpt	NOUN
fcis-25702	12	8	4	4	NUM
fcis-25702	12	9	[	[	X
fcis-25702	12	10	1	1	NUM
fcis-25702	12	11	]	]	PUNCT
fcis-25702	12	12	,	,	PUNCT
fcis-25702	12	13	large	large	ADJ
fcis-25702	12	14	language	language	NOUN
fcis-25702	12	15	models	model	NOUN
fcis-25702	12	16	(	(	PUNCT
fcis-25702	12	17	llms	llm	NOUN
fcis-25702	12	18	)	)	PUNCT
fcis-25702	12	19	have	have	AUX
fcis-25702	12	20	become	become	VERB
fcis-25702	12	21	ubiquitous	ubiquitous	ADJ
fcis-25702	12	22	in	in	ADP
fcis-25702	12	23	almost	almost	ADV
fcis-25702	12	24	all	all	PRON
fcis-25702	12	25	sub	sub	NOUN
fcis-25702	12	26	-	-	NOUN
fcis-25702	12	27	fields	field	NOUN
fcis-25702	12	28	of	of	ADP
fcis-25702	12	29	natural	natural	ADJ
fcis-25702	12	30	language	language	NOUN
fcis-25702	12	31	processing	processing	NOUN
fcis-25702	12	32	.	.	PUNCT
fcis-25702	13	1	these	these	DET
fcis-25702	13	2	models	model	NOUN
fcis-25702	13	3	are	be	AUX
fcis-25702	13	4	pretrained	pretraine	VERB
fcis-25702	13	5	on	on	ADP
fcis-25702	13	6	an	an	DET
fcis-25702	13	7	enormous	enormous	ADJ
fcis-25702	13	8	number	number	NOUN
fcis-25702	13	9	of	of	ADP
fcis-25702	13	10	tokens	token	NOUN
fcis-25702	13	11	[	[	X
fcis-25702	13	12	2	2	NUM
fcis-25702	13	13	]	]	PUNCT
fcis-25702	13	14	and	and	CCONJ
fcis-25702	13	15	have	have	AUX
fcis-25702	13	16	shown	show	VERB
fcis-25702	13	17	excellent	excellent	ADJ
fcis-25702	13	18	capabilities	capability	NOUN
fcis-25702	13	19	in	in	ADP
fcis-25702	13	20	following	follow	VERB
fcis-25702	13	21	instructions	instruction	NOUN
fcis-25702	13	22	for	for	ADP
fcis-25702	13	23	sentence	sentence	NOUN
fcis-25702	13	24	generation	generation	NOUN
fcis-25702	13	25	.	.	PUNCT
fcis-25702	14	1	for	for	ADP
fcis-25702	14	2	models	model	NOUN
fcis-25702	14	3	with	with	ADP
fcis-25702	14	4	more	more	ADJ
fcis-25702	14	5	than	than	ADP
fcis-25702	14	6	10	10	NUM
fcis-25702	14	7	billion	billion	NUM
fcis-25702	14	8	parameters	parameter	NOUN
fcis-25702	14	9	,	,	PUNCT
fcis-25702	14	10	emergent	emergent	ADJ
fcis-25702	14	11	capabilities	capability	NOUN
fcis-25702	14	12	have	have	AUX
fcis-25702	14	13	occurred	occur	VERB
fcis-25702	14	14	[	[	X
fcis-25702	14	15	3	3	NUM
fcis-25702	14	16	]	]	X
fcis-25702	14	17	,	,	PUNCT
fcis-25702	14	18	such	such	ADJ
fcis-25702	14	19	as	as	ADP
fcis-25702	14	20	in	in	ADP
fcis-25702	14	21	-	-	PUNCT
fcis-25702	14	22	context	context	NOUN
fcis-25702	14	23	learning	learning	NOUN
fcis-25702	14	24	and	and	CCONJ
fcis-25702	14	25	reasoning	reason	VERB
fcis-25702	14	26	capabilities	capability	NOUN
fcis-25702	14	27	[	[	X
fcis-25702	14	28	4	4	NUM
fcis-25702	14	29	]	]	PUNCT
fcis-25702	14	30	.	.	PUNCT
fcis-25702	15	1	an	an	DET
fcis-25702	15	2	appealing	appealing	ADJ
fcis-25702	15	3	property	property	NOUN
fcis-25702	15	4	of	of	ADP
fcis-25702	15	5	llms	llm	NOUN
fcis-25702	15	6	is	be	AUX
fcis-25702	15	7	that	that	SCONJ
fcis-25702	15	8	after	after	ADP
fcis-25702	15	9	being	be	AUX
fcis-25702	15	10	trained	train	VERB
fcis-25702	15	11	with	with	ADP
fcis-25702	15	12	large	large	ADJ
fcis-25702	15	13	amounts	amount	NOUN
fcis-25702	15	14	of	of	ADP
fcis-25702	15	15	instructionresponse	instructionresponse	NOUN
fcis-25702	15	16	data	datum	NOUN
fcis-25702	15	17	,	,	PUNCT
fcis-25702	15	18	the	the	DET
fcis-25702	15	19	models	model	NOUN
fcis-25702	15	20	can	can	AUX
fcis-25702	15	21	generalize	generalize	VERB
fcis-25702	15	22	well	well	ADV
fcis-25702	15	23	across	across	ADP
fcis-25702	15	24	many	many	ADJ
fcis-25702	15	25	different	different	ADJ
fcis-25702	15	26	natural	natural	ADJ
fcis-25702	15	27	language	language	NOUN
fcis-25702	15	28	processing	processing	NOUN
fcis-25702	15	29	tasks	task	NOUN
fcis-25702	15	30	[	[	X
fcis-25702	15	31	5	5	NUM
fcis-25702	15	32	]	]	PUNCT
fcis-25702	15	33	,	,	PUNCT
fcis-25702	15	34	such	such	ADJ
fcis-25702	15	35	as	as	ADP
fcis-25702	15	36	natural	natural	ADJ
fcis-25702	15	37	language	language	NOUN
fcis-25702	15	38	inference	inference	NOUN
fcis-25702	15	39	[	[	X
fcis-25702	15	40	6	6	NUM
fcis-25702	15	41	]	]	PUNCT
fcis-25702	15	42	,	,	PUNCT
fcis-25702	15	43	information	information	NOUN
fcis-25702	15	44	extraction	extraction	NOUN
fcis-25702	15	45	[	[	X
fcis-25702	15	46	7][8][9	7][8][9	NUM
fcis-25702	15	47	]	]	PUNCT
fcis-25702	15	48	,	,	PUNCT
fcis-25702	15	49	and	and	CCONJ
fcis-25702	15	50	text	text	NOUN
fcis-25702	15	51	classification	classification	NOUN
fcis-25702	15	52	[	[	X
fcis-25702	15	53	10][11][12	10][11][12	NUM
fcis-25702	15	54	]	]	PUNCT
fcis-25702	15	55	.	.	PUNCT
fcis-25702	16	1	despite	despite	SCONJ
fcis-25702	16	2	the	the	DET
fcis-25702	16	3	impressive	impressive	ADJ
fcis-25702	16	4	capabilities	capability	NOUN
fcis-25702	16	5	of	of	ADP
fcis-25702	16	6	llms	llm	NOUN
fcis-25702	16	7	in	in	ADP
fcis-25702	16	8	completing	complete	VERB
fcis-25702	16	9	instruction	instruction	NOUN
fcis-25702	16	10	-	-	PUNCT
fcis-25702	16	11	following	follow	VERB
fcis-25702	16	12	tasks	task	NOUN
fcis-25702	16	13	[	[	X
fcis-25702	16	14	13][14][9	13][14][9	NUM
fcis-25702	16	15	]	]	PUNCT
fcis-25702	16	16	,	,	PUNCT
fcis-25702	16	17	the	the	DET
fcis-25702	16	18	computational	computational	ADJ
fcis-25702	16	19	cost	cost	NOUN
fcis-25702	16	20	of	of	ADP
fcis-25702	16	21	llms	llm	NOUN
fcis-25702	16	22	is	be	AUX
fcis-25702	16	23	a	a	DET
fcis-25702	16	24	significant	significant	ADJ
fcis-25702	16	25	issue	issue	NOUN
fcis-25702	16	26	because	because	SCONJ
fcis-25702	16	27	training	train	VERB
fcis-25702	16	28	large	large	ADJ
fcis-25702	16	29	models	model	NOUN
fcis-25702	16	30	requires	require	VERB
fcis-25702	16	31	substantial	substantial	ADJ
fcis-25702	16	32	time	time	NOUN
fcis-25702	16	33	,	,	PUNCT
fcis-25702	16	34	computational	computational	ADJ
fcis-25702	16	35	resources	resource	NOUN
fcis-25702	16	36	,	,	PUNCT
fcis-25702	16	37	and	and	CCONJ
fcis-25702	16	38	considerable	considerable	ADJ
fcis-25702	16	39	expenses	expense	NOUN
fcis-25702	16	40	.	.	PUNCT
fcis-25702	17	1	furthermore	furthermore	ADV
fcis-25702	17	2	,	,	PUNCT
fcis-25702	17	3	the	the	DET
fcis-25702	17	4	large	large	ADJ
fcis-25702	17	5	latency	latency	NOUN
fcis-25702	17	6	is	be	AUX
fcis-25702	17	7	an	an	DET
fcis-25702	17	8	obstacle	obstacle	NOUN
fcis-25702	17	9	for	for	ADP
fcis-25702	17	10	llms	llm	NOUN
fcis-25702	17	11	to	to	PART
fcis-25702	17	12	be	be	AUX
fcis-25702	17	13	applied	apply	VERB
fcis-25702	17	14	in	in	ADP
fcis-25702	17	15	many	many	ADJ
fcis-25702	17	16	industrial	industrial	ADJ
fcis-25702	17	17	scenarios	scenario	NOUN
fcis-25702	17	18	with	with	ADP
fcis-25702	17	19	low	low	ADJ
fcis-25702	17	20	latency	latency	NOUN
fcis-25702	17	21	requirements	requirement	NOUN
fcis-25702	17	22	,	,	PUNCT
fcis-25702	17	23	which	which	PRON
fcis-25702	17	24	require	require	VERB
fcis-25702	17	25	fast	fast	ADJ
fcis-25702	17	26	inference	inference	NOUN
fcis-25702	17	27	speed	speed	NOUN
fcis-25702	17	28	to	to	PART
fcis-25702	17	29	meet	meet	VERB
fcis-25702	17	30	users	user	NOUN
fcis-25702	17	31	’	'	PUNCT
fcis-25702	17	32	needs	need	NOUN
fcis-25702	17	33	.	.	PUNCT
fcis-25702	18	1	there	there	PRON
fcis-25702	18	2	is	be	VERB
fcis-25702	18	3	much	much	ADJ
fcis-25702	18	4	research	research	NOUN
fcis-25702	18	5	work	work	NOUN
fcis-25702	18	6	focusing	focus	VERB
fcis-25702	18	7	on	on	ADP
fcis-25702	18	8	speeding	speed	VERB
fcis-25702	18	9	up	up	ADP
fcis-25702	18	10	the	the	DET
fcis-25702	18	11	inference	inference	NOUN
fcis-25702	18	12	of	of	ADP
fcis-25702	18	13	bert	bert	PROPN
fcis-25702	18	14	[	[	X
fcis-25702	18	15	15	15	NUM
fcis-25702	18	16	]	]	PUNCT
fcis-25702	18	17	and	and	CCONJ
fcis-25702	18	18	llms	llm	NOUN
fcis-25702	18	19	,	,	PUNCT
fcis-25702	18	20	such	such	ADJ
fcis-25702	18	21	as	as	ADP
fcis-25702	18	22	network	network	NOUN
fcis-25702	18	23	pruning	prune	VERB
fcis-25702	18	24	[	[	X
fcis-25702	18	25	16][17	16][17	NUM
fcis-25702	18	26	]	]	PUNCT
fcis-25702	18	27	,	,	PUNCT
fcis-25702	18	28	student	student	NOUN
fcis-25702	18	29	network	network	NOUN
fcis-25702	18	30	distillation	distillation	NOUN
fcis-25702	19	1	[	[	X
fcis-25702	19	2	18][19	18][19	NUM
fcis-25702	19	3	]	]	PUNCT
fcis-25702	19	4	,	,	PUNCT
fcis-25702	19	5	quantization	quantization	NOUN
fcis-25702	19	6	[	[	X
fcis-25702	19	7	20][21	20][21	X
fcis-25702	19	8	]	]	X
fcis-25702	19	9	,	,	PUNCT
fcis-25702	19	10	and	and	CCONJ
fcis-25702	19	11	early	early	ADV
fcis-25702	19	12	exiting	exit	VERB
fcis-25702	19	13	[	[	X
fcis-25702	19	14	22][23][24][25][26	22][23][24][25][26	NUM
fcis-25702	19	15	]	]	PUNCT
fcis-25702	19	16	.	.	PUNCT
fcis-25702	20	1	network	network	NOUN
fcis-25702	20	2	pruning	pruning	NOUN
fcis-25702	20	3	,	,	PUNCT
fcis-25702	20	4	student	student	NOUN
fcis-25702	20	5	network	network	NOUN
fcis-25702	20	6	distillation	distillation	NOUN
fcis-25702	20	7	,	,	PUNCT
fcis-25702	20	8	and	and	CCONJ
fcis-25702	20	9	quantization	quantization	NOUN
fcis-25702	20	10	have	have	AUX
fcis-25702	20	11	been	be	AUX
fcis-25702	20	12	very	very	ADV
fcis-25702	20	13	effective	effective	ADJ
fcis-25702	20	14	in	in	ADP
fcis-25702	20	15	reducing	reduce	VERB
fcis-25702	20	16	the	the	DET
fcis-25702	20	17	parameters	parameter	NOUN
fcis-25702	20	18	or	or	CCONJ
fcis-25702	20	19	computation	computation	NOUN
fcis-25702	20	20	cost	cost	NOUN
fcis-25702	20	21	of	of	ADP
fcis-25702	20	22	pretrained	pretraine	VERB
fcis-25702	20	23	language	language	NOUN
fcis-25702	20	24	models	model	NOUN
fcis-25702	20	25	,	,	PUNCT
fcis-25702	20	26	but	but	CCONJ
fcis-25702	20	27	the	the	DET
fcis-25702	20	28	model	model	NOUN
fcis-25702	20	29	architecture	architecture	NOUN
fcis-25702	20	30	of	of	ADP
fcis-25702	20	31	these	these	DET
fcis-25702	20	32	methods	method	NOUN
fcis-25702	20	33	is	be	AUX
fcis-25702	20	34	usually	usually	ADV
fcis-25702	20	35	fixed	fix	VERB
fcis-25702	20	36	during	during	ADP
fcis-25702	20	37	inference	inference	NOUN
fcis-25702	20	38	.	.	PUNCT
fcis-25702	21	1	some	some	DET
fcis-25702	21	2	early	early	ADJ
fcis-25702	21	3	exiting	exiting	NOUN
fcis-25702	21	4	works	work	VERB
fcis-25702	21	5	[	[	X
fcis-25702	21	6	23][26	23][26	X
fcis-25702	21	7	]	]	X
fcis-25702	21	8	have	have	AUX
fcis-25702	21	9	proved	prove	VERB
fcis-25702	21	10	that	that	SCONJ
fcis-25702	21	11	the	the	DET
fcis-25702	21	12	network	network	NOUN
fcis-25702	21	13	structure	structure	NOUN
fcis-25702	21	14	varies	vary	VERB
fcis-25702	21	15	for	for	ADP
fcis-25702	21	16	different	different	ADJ
fcis-25702	21	17	test	test	NOUN
fcis-25702	21	18	samples	sample	NOUN
fcis-25702	21	19	during	during	ADP
fcis-25702	21	20	inference	inference	NOUN
fcis-25702	21	21	.	.	PUNCT
fcis-25702	22	1	they	they	PRON
fcis-25702	22	2	use	use	VERB
fcis-25702	22	3	dynamic	dynamic	ADJ
fcis-25702	22	4	inference	inference	NOUN
fcis-25702	22	5	methods	method	NOUN
fcis-25702	22	6	,	,	PUNCT
fcis-25702	22	7	based	base	VERB
fcis-25702	22	8	on	on	ADP
fcis-25702	22	9	certain	certain	ADJ
fcis-25702	22	10	criteria	criterion	NOUN
fcis-25702	22	11	such	such	ADJ
fcis-25702	22	12	as	as	ADP
fcis-25702	22	13	the	the	DET
fcis-25702	22	14	entropy	entropy	NOUN
fcis-25702	22	15	of	of	ADP
fcis-25702	22	16	logits	logit	NOUN
fcis-25702	22	17	,	,	PUNCT
fcis-25702	22	18	to	to	PART
fcis-25702	22	19	allocate	allocate	VERB
fcis-25702	22	20	different	different	ADJ
fcis-25702	22	21	bert	bert	NOUN
fcis-25702	22	22	layers	layer	NOUN
fcis-25702	22	23	for	for	ADP
fcis-25702	22	24	different	different	ADJ
fcis-25702	22	25	test	test	NOUN
fcis-25702	22	26	samples	sample	NOUN
fcis-25702	22	27	in	in	ADP
fcis-25702	22	28	inference	inference	NOUN
fcis-25702	22	29	stages	stage	NOUN
fcis-25702	22	30	.	.	PUNCT
fcis-25702	23	1	therefore	therefore	ADV
fcis-25702	23	2	,	,	PUNCT
fcis-25702	23	3	due	due	ADP
fcis-25702	23	4	to	to	ADP
fcis-25702	23	5	its	its	PRON
fcis-25702	23	6	potential	potential	NOUN
fcis-25702	23	7	in	in	ADP
fcis-25702	23	8	applications	application	NOUN
fcis-25702	23	9	,	,	PUNCT
fcis-25702	23	10	we	we	PRON
fcis-25702	23	11	focus	focus	VERB
fcis-25702	23	12	on	on	ADP
fcis-25702	23	13	applying	apply	VERB
fcis-25702	23	14	early	early	ADV
fcis-25702	23	15	exiting	exit	VERB
fcis-25702	23	16	to	to	ADP
fcis-25702	23	17	llms	llm	NOUN
fcis-25702	23	18	.	.	PUNCT
fcis-25702	24	1	early	early	ADJ
fcis-25702	24	2	exiting	exiting	NOUN
fcis-25702	24	3	requires	require	VERB
fcis-25702	24	4	a	a	DET
fcis-25702	24	5	multi	multi	ADJ
fcis-25702	24	6	-	-	ADJ
fcis-25702	24	7	exit	exit	NOUN
fcis-25702	24	8	pretrained	pretraine	VERB
fcis-25702	24	9	language	language	NOUN
fcis-25702	24	10	model	model	NOUN
fcis-25702	24	11	,	,	PUNCT
fcis-25702	24	12	such	such	ADJ
fcis-25702	24	13	as	as	ADP
fcis-25702	24	14	bert	bert	NOUN
fcis-25702	24	15	and	and	CCONJ
fcis-25702	24	16	gpt	gpt	NOUN
fcis-25702	25	1	[	[	X
fcis-25702	25	2	27	27	NUM
fcis-25702	25	3	]	]	PUNCT
fcis-25702	25	4	,	,	PUNCT
fcis-25702	25	5	with	with	ADP
fcis-25702	25	6	an	an	DET
fcis-25702	25	7	intermediate	intermediate	ADJ
fcis-25702	25	8	classifier	classifier	NOUN
fcis-25702	25	9	(	(	PUNCT
fcis-25702	25	10	or	or	CCONJ
fcis-25702	25	11	early	early	ADJ
fcis-25702	25	12	exit	exit	NOUN
fcis-25702	25	13	layer	layer	NOUN
fcis-25702	25	14	)	)	PUNCT
fcis-25702	25	15	installed	instal	VERB
fcis-25702	25	16	at	at	ADP
fcis-25702	25	17	each	each	DET
fcis-25702	25	18	transformer	transformer	ADJ
fcis-25702	25	19	layer	layer	NOUN
fcis-25702	25	20	.	.	PUNCT
fcis-25702	26	1	then	then	ADV
fcis-25702	26	2	,	,	PUNCT
fcis-25702	26	3	a	a	DET
fcis-25702	26	4	dynamic	dynamic	ADJ
fcis-25702	26	5	early	early	ADJ
fcis-25702	26	6	exiting	exit	VERB
fcis-25702	26	7	mechanism	mechanism	NOUN
fcis-25702	26	8	is	be	AUX
fcis-25702	26	9	applied	apply	VERB
fcis-25702	26	10	at	at	ADP
fcis-25702	26	11	each	each	DET
fcis-25702	26	12	layer	layer	NOUN
fcis-25702	26	13	during	during	ADP
fcis-25702	26	14	the	the	DET
fcis-25702	26	15	forward	forward	ADJ
fcis-25702	26	16	pass	pass	NOUN
fcis-25702	26	17	to	to	PART
fcis-25702	26	18	assess	assess	VERB
fcis-25702	26	19	whether	whether	SCONJ
fcis-25702	26	20	to	to	PART
fcis-25702	26	21	exit	exit	VERB
fcis-25702	26	22	at	at	ADP
fcis-25702	26	23	the	the	DET
fcis-25702	26	24	current	current	ADJ
fcis-25702	26	25	layer	layer	NOUN
fcis-25702	26	26	.	.	PUNCT
fcis-25702	27	1	early	early	ADJ
fcis-25702	27	2	exiting	exit	VERB
fcis-25702	27	3	can	can	AUX
fcis-25702	27	4	work	work	VERB
fcis-25702	27	5	together	together	ADV
fcis-25702	27	6	with	with	ADP
fcis-25702	27	7	static	static	ADJ
fcis-25702	27	8	model	model	NOUN
fcis-25702	27	9	compression	compression	NOUN
fcis-25702	27	10	methods	method	NOUN
fcis-25702	27	11	,	,	PUNCT
fcis-25702	27	12	such	such	ADJ
fcis-25702	27	13	as	as	ADP
fcis-25702	27	14	performing	perform	VERB
fcis-25702	27	15	early	early	ADV
fcis-25702	27	16	exiting	exit	VERB
fcis-25702	27	17	on	on	ADP
fcis-25702	27	18	a	a	DET
fcis-25702	27	19	pruned	prune	VERB
fcis-25702	27	20	model	model	NOUN
fcis-25702	27	21	[	[	X
fcis-25702	27	22	28	28	NUM
fcis-25702	27	23	]	]	PUNCT
fcis-25702	27	24	.	.	PUNCT
fcis-25702	28	1	most	most	ADJ
fcis-25702	28	2	of	of	ADP
fcis-25702	28	3	the	the	DET
fcis-25702	28	4	work	work	NOUN
fcis-25702	28	5	[	[	X
fcis-25702	28	6	22][23][25][26	22][23][25][26	X
fcis-25702	28	7	]	]	X
fcis-25702	28	8	on	on	ADP
fcis-25702	28	9	early	early	ADJ
fcis-25702	28	10	exiting	exiting	NOUN
fcis-25702	28	11	focuses	focus	VERB
fcis-25702	28	12	on	on	ADP
fcis-25702	28	13	the	the	DET
fcis-25702	28	14	classification	classification	NOUN
fcis-25702	28	15	task	task	NOUN
fcis-25702	28	16	using	use	VERB
fcis-25702	28	17	bert	bert	PROPN
fcis-25702	28	18	as	as	ADP
fcis-25702	28	19	the	the	DET
fcis-25702	28	20	backbone	backbone	NOUN
fcis-25702	28	21	model	model	NOUN
fcis-25702	28	22	,	,	PUNCT
fcis-25702	28	23	and	and	CCONJ
fcis-25702	28	24	recently	recently	ADV
fcis-25702	28	25	some	some	DET
fcis-25702	28	26	work	work	NOUN
fcis-25702	28	27	has	have	AUX
fcis-25702	28	28	started	start	VERB
fcis-25702	28	29	to	to	PART
fcis-25702	28	30	apply	apply	VERB
fcis-25702	28	31	early	early	ADV
fcis-25702	28	32	exiting	exit	VERB
fcis-25702	28	33	to	to	ADP
fcis-25702	28	34	generative	generative	ADJ
fcis-25702	28	35	tasks	task	NOUN
fcis-25702	28	36	.	.	PUNCT
fcis-25702	29	1	men	man	NOUN
fcis-25702	29	2	et	et	PROPN
fcis-25702	29	3	al	al	PROPN
fcis-25702	29	4	.	.	PUNCT
fcis-25702	30	1	[	[	X
fcis-25702	30	2	29	29	NUM
fcis-25702	30	3	]	]	PUNCT
fcis-25702	30	4	found	find	VERB
fcis-25702	30	5	that	that	SCONJ
fcis-25702	30	6	llm	llm	PROPN
fcis-25702	30	7	layers	layer	NOUN
fcis-25702	30	8	have	have	VERB
fcis-25702	30	9	high	high	ADJ
fcis-25702	30	10	redundancy	redundancy	NOUN
fcis-25702	30	11	,	,	PUNCT
fcis-25702	30	12	and	and	CCONJ
fcis-25702	30	13	removing	remove	VERB
fcis-25702	30	14	the	the	DET
fcis-25702	30	15	redundant	redundant	ADJ
fcis-25702	30	16	layers	layer	NOUN
fcis-25702	30	17	does	do	AUX
fcis-25702	30	18	not	not	PART
fcis-25702	30	19	significantly	significantly	ADV
fcis-25702	30	20	drop	drop	VERB
fcis-25702	30	21	llm	llm	NOUN
fcis-25702	30	22	performance	performance	NOUN
fcis-25702	30	23	.	.	PUNCT
fcis-25702	31	1	calm	calm	ADJ
fcis-25702	32	1	[	[	X
fcis-25702	32	2	30	30	NUM
fcis-25702	32	3	]	]	PUNCT
fcis-25702	32	4	adds	add	VERB
fcis-25702	32	5	an	an	DET
fcis-25702	32	6	early	early	ADJ
fcis-25702	32	7	exit	exit	NOUN
fcis-25702	32	8	layer	layer	NOUN
fcis-25702	32	9	at	at	ADP
fcis-25702	32	10	each	each	DET
fcis-25702	32	11	decoder	decoder	NOUN
fcis-25702	32	12	layer	layer	NOUN
fcis-25702	32	13	of	of	ADP
fcis-25702	32	14	a	a	DET
fcis-25702	32	15	small	small	ADJ
fcis-25702	32	16	t5	t5	PROPN
fcis-25702	32	17	model	model	NOUN
fcis-25702	32	18	(	(	PUNCT
fcis-25702	32	19	8	8	NUM
fcis-25702	32	20	layers	layer	NOUN
fcis-25702	32	21	)	)	PUNCT
fcis-25702	33	1	[	[	X
fcis-25702	33	2	31	31	NUM
fcis-25702	33	3	]	]	PUNCT
fcis-25702	33	4	and	and	CCONJ
fcis-25702	33	5	uses	use	VERB
fcis-25702	33	6	entropy	entropy	NOUN
fcis-25702	33	7	as	as	ADP
fcis-25702	33	8	the	the	DET
fcis-25702	33	9	criterion	criterion	NOUN
fcis-25702	33	10	to	to	ADP
fcis-25702	33	11	exit	exit	NOUN
fcis-25702	33	12	.	.	PUNCT
fcis-25702	34	1	however	however	ADV
fcis-25702	34	2	,	,	PUNCT
fcis-25702	34	3	for	for	ADP
fcis-25702	34	4	llms	llm	NOUN
fcis-25702	34	5	such	such	ADJ
fcis-25702	34	6	as	as	ADP
fcis-25702	34	7	llama	llama	NOUN
fcis-25702	34	8	[	[	X
fcis-25702	34	9	32	32	NUM
fcis-25702	34	10	]	]	PUNCT
fcis-25702	34	11	with	with	ADP
fcis-25702	34	12	a	a	DET
fcis-25702	34	13	large	large	ADJ
fcis-25702	34	14	language	language	NOUN
fcis-25702	34	15	modeling	modeling	NOUN
fcis-25702	34	16	head	head	NOUN
fcis-25702	34	17	size	size	NOUN
fcis-25702	34	18	and	and	CCONJ
fcis-25702	34	19	deep	deep	ADJ
fcis-25702	34	20	layers	layer	NOUN
fcis-25702	34	21	,	,	PUNCT
fcis-25702	34	22	it	it	PRON
fcis-25702	34	23	would	would	AUX
fcis-25702	34	24	incur	incur	VERB
fcis-25702	34	25	significant	significant	ADJ
fcis-25702	34	26	computational	computational	ADJ
fcis-25702	34	27	costs	cost	NOUN
fcis-25702	34	28	.	.	PUNCT
fcis-25702	35	1	skipdecode	skipdecode	NOUN
fcis-25702	36	1	[	[	X
fcis-25702	36	2	33	33	NUM
fcis-25702	36	3	]	]	PUNCT
fcis-25702	36	4	proposes	propose	VERB
fcis-25702	36	5	using	use	VERB
fcis-25702	36	6	layer	layer	NOUN
fcis-25702	36	7	skipping	skipping	NOUN
fcis-25702	36	8	,	,	PUNCT
fcis-25702	36	9	employing	employ	VERB
fcis-25702	36	10	rule	rule	NOUN
fcis-25702	36	11	-	-	PUNCT
fcis-25702	36	12	based	base	VERB
fcis-25702	36	13	methods	method	NOUN
fcis-25702	36	14	,	,	PUNCT
fcis-25702	36	15	to	to	PART
fcis-25702	36	16	gradually	gradually	ADV
fcis-25702	36	17	skip	skip	VERB
fcis-25702	36	18	more	more	ADJ
fcis-25702	36	19	intermediate	intermediate	ADJ
fcis-25702	36	20	gpt	gpt	NOUN
fcis-25702	36	21	layers	layer	NOUN
fcis-25702	36	22	as	as	SCONJ
fcis-25702	36	23	more	more	ADJ
fcis-25702	36	24	tokens	token	NOUN
fcis-25702	36	25	are	be	AUX
fcis-25702	36	26	generated	generate	VERB
fcis-25702	36	27	.	.	PUNCT
fcis-25702	37	1	this	this	DET
fcis-25702	37	2	approach	approach	NOUN
fcis-25702	37	3	avoids	avoid	VERB
fcis-25702	37	4	installing	instal	VERB
fcis-25702	37	5	an	an	DET
fcis-25702	37	6	exit	exit	NOUN
fcis-25702	37	7	layer	layer	NOUN
fcis-25702	37	8	at	at	ADP
fcis-25702	37	9	each	each	DET
fcis-25702	37	10	gpt	gpt	NOUN
fcis-25702	37	11	layer	layer	NOUN
fcis-25702	37	12	.	.	PUNCT
fcis-25702	38	1	however	however	ADV
fcis-25702	38	2	,	,	PUNCT
fcis-25702	38	3	it	it	PRON
fcis-25702	38	4	applies	apply	VERB
fcis-25702	38	5	the	the	DET
fcis-25702	38	6	same	same	ADJ
fcis-25702	38	7	skipping	skip	VERB
fcis-25702	38	8	layers	layer	NOUN
fcis-25702	38	9	for	for	ADP
fcis-25702	38	10	all	all	DET
fcis-25702	38	11	samples	sample	NOUN
fcis-25702	38	12	,	,	PUNCT
fcis-25702	38	13	which	which	PRON
fcis-25702	38	14	does	do	AUX
fcis-25702	38	15	not	not	PART
fcis-25702	38	16	constitute	constitute	VERB
fcis-25702	38	17	dynamic	dynamic	ADJ
fcis-25702	38	18	inference	inference	NOUN
fcis-25702	38	19	.	.	PUNCT
fcis-25702	39	1	yuan	yuan	NOUN
fcis-25702	39	2	et	et	PROPN
fcis-25702	39	3	al	al	PROPN
fcis-25702	39	4	.	.	PUNCT
fcis-25702	40	1	[	[	X
fcis-25702	40	2	34	34	NUM
fcis-25702	40	3	]	]	PUNCT
fcis-25702	40	4	directly	directly	ADV
fcis-25702	40	5	prunes	prune	VERB
fcis-25702	40	6	the	the	DET
fcis-25702	40	7	layers	layer	NOUN
fcis-25702	40	8	of	of	ADP
fcis-25702	40	9	llms	llm	NOUN
fcis-25702	40	10	,	,	PUNCT
fcis-25702	40	11	retains	retain	VERB
fcis-25702	40	12	the	the	DET
fcis-25702	40	13	last	last	ADJ
fcis-25702	40	14	layer	layer	NOUN
fcis-25702	40	15	,	,	PUNCT
fcis-25702	40	16	and	and	CCONJ
fcis-25702	40	17	finetunes	finetune	VERB
fcis-25702	40	18	the	the	DET
fcis-25702	40	19	model	model	NOUN
fcis-25702	40	20	on	on	ADP
fcis-25702	40	21	nlu	nlu	NOUN
fcis-25702	40	22	datasets	dataset	NOUN
fcis-25702	40	23	.	.	PUNCT
fcis-25702	41	1	it	it	PRON
fcis-25702	41	2	shows	show	VERB
fcis-25702	41	3	that	that	SCONJ
fcis-25702	41	4	skipping	skip	VERB
fcis-25702	41	5	more	more	ADJ
fcis-25702	41	6	than	than	ADP
fcis-25702	41	7	half	half	NOUN
fcis-25702	41	8	of	of	ADP
fcis-25702	41	9	the	the	DET
fcis-25702	41	10	llm	llm	PROPN
fcis-25702	41	11	’s	’s	PART
fcis-25702	41	12	layers	layer	NOUN
fcis-25702	41	13	can	can	AUX
fcis-25702	41	14	still	still	ADV
fcis-25702	41	15	achieve	achieve	VERB
fcis-25702	41	16	comparable	comparable	ADJ
fcis-25702	41	17	or	or	CCONJ
fcis-25702	41	18	slightly	slightly	ADV
fcis-25702	41	19	better	well	ADJ
fcis-25702	41	20	results	result	NOUN
fcis-25702	41	21	.	.	PUNCT
fcis-25702	42	1	however	however	ADV
fcis-25702	42	2	,	,	PUNCT
fcis-25702	42	3	these	these	DET
fcis-25702	42	4	pruning	prune	VERB
fcis-25702	42	5	methods	method	NOUN
fcis-25702	42	6	do	do	AUX
fcis-25702	42	7	not	not	PART
fcis-25702	42	8	consider	consider	VERB
fcis-25702	42	9	dynamic	dynamic	ADJ
fcis-25702	42	10	inference	inference	NOUN
fcis-25702	42	11	for	for	ADP
fcis-25702	42	12	different	different	ADJ
fcis-25702	42	13	samples	sample	NOUN
fcis-25702	42	14	.	.	PUNCT
fcis-25702	43	1	some	some	DET
fcis-25702	43	2	hard	hard	ADJ
fcis-25702	43	3	samples	sample	NOUN
fcis-25702	43	4	may	may	AUX
fcis-25702	43	5	need	need	VERB
fcis-25702	43	6	more	more	ADJ
fcis-25702	43	7	layers	layer	NOUN
fcis-25702	43	8	to	to	PART
fcis-25702	43	9	compute	compute	VERB
fcis-25702	43	10	,	,	PUNCT
fcis-25702	43	11	while	while	SCONJ
fcis-25702	43	12	easy	easy	ADJ
fcis-25702	43	13	samples	sample	NOUN
fcis-25702	43	14	may	may	AUX
fcis-25702	43	15	need	need	VERB
fcis-25702	43	16	fewer	few	ADJ
fcis-25702	43	17	layers	layer	NOUN
fcis-25702	43	18	.	.	PUNCT
fcis-25702	44	1	also	also	ADV
fcis-25702	44	2	,	,	PUNCT
fcis-25702	44	3	certain	certain	ADJ
fcis-25702	44	4	layers	layer	NOUN
fcis-25702	44	5	may	may	AUX
fcis-25702	44	6	be	be	AUX
fcis-25702	44	7	important	important	ADJ
fcis-25702	44	8	for	for	ADP
fcis-25702	44	9	some	some	DET
fcis-25702	44	10	samples	sample	NOUN
fcis-25702	44	11	while	while	SCONJ
fcis-25702	44	12	2	2	NUM
fcis-25702	44	13	unimportant	unimportant	ADJ
fcis-25702	44	14	for	for	ADP
fcis-25702	44	15	others	other	NOUN
fcis-25702	44	16	.	.	PUNCT
fcis-25702	45	1	since	since	SCONJ
fcis-25702	45	2	previous	previous	ADJ
fcis-25702	45	3	work	work	NOUN
fcis-25702	45	4	has	have	AUX
fcis-25702	45	5	not	not	PART
fcis-25702	45	6	applied	apply	VERB
fcis-25702	45	7	dynamic	dynamic	ADJ
fcis-25702	45	8	inference	inference	NOUN
fcis-25702	45	9	to	to	ADP
fcis-25702	45	10	llms	llm	NOUN
fcis-25702	45	11	,	,	PUNCT
fcis-25702	45	12	we	we	PRON
fcis-25702	45	13	propose	propose	VERB
fcis-25702	45	14	a	a	DET
fcis-25702	45	15	novel	novel	ADJ
fcis-25702	45	16	framework	framework	NOUN
fcis-25702	45	17	for	for	ADP
fcis-25702	45	18	speeding	speed	VERB
fcis-25702	45	19	up	up	ADP
fcis-25702	45	20	llms	llm	NOUN
fcis-25702	45	21	by	by	ADP
fcis-25702	45	22	using	use	VERB
fcis-25702	45	23	controllers	controller	NOUN
fcis-25702	45	24	to	to	ADP
fcis-25702	45	25	dy	dy	X
fcis-25702	45	26	namically	namically	ADV
fcis-25702	45	27	skip	skip	VERB
fcis-25702	45	28	layers	layer	NOUN
fcis-25702	45	29	for	for	ADP
fcis-25702	45	30	different	different	ADJ
fcis-25702	45	31	samples	sample	NOUN
fcis-25702	45	32	.	.	PUNCT
fcis-25702	46	1	we	we	PRON
fcis-25702	46	2	use	use	VERB
fcis-25702	46	3	reinforcement	reinforcement	NOUN
fcis-25702	46	4	learning	learning	NOUN
fcis-25702	46	5	(	(	PUNCT
fcis-25702	46	6	rl	rl	NOUN
fcis-25702	46	7	)	)	PUNCT
fcis-25702	46	8	to	to	PART
fcis-25702	46	9	train	train	VERB
fcis-25702	46	10	the	the	DET
fcis-25702	46	11	controller	controller	NOUN
fcis-25702	46	12	,	,	PUNCT
fcis-25702	46	13	which	which	PRON
fcis-25702	46	14	decides	decide	VERB
fcis-25702	46	15	whether	whether	SCONJ
fcis-25702	46	16	to	to	PART
fcis-25702	46	17	dynamically	dynamically	ADV
fcis-25702	46	18	skip	skip	VERB
fcis-25702	46	19	the	the	DET
fcis-25702	46	20	next	next	ADJ
fcis-25702	46	21	layer	layer	NOUN
fcis-25702	46	22	.	.	PUNCT
fcis-25702	47	1	since	since	SCONJ
fcis-25702	47	2	the	the	DET
fcis-25702	47	3	training	training	NOUN
fcis-25702	47	4	of	of	ADP
fcis-25702	47	5	rl	rl	PROPN
fcis-25702	47	6	is	be	AUX
fcis-25702	47	7	unstableand	unstableand	NOUN
fcis-25702	47	8	hard	hard	ADJ
fcis-25702	47	9	to	to	PART
fcis-25702	47	10	converge	converge	VERB
fcis-25702	47	11	,	,	PUNCT
fcis-25702	47	12	we	we	PRON
fcis-25702	47	13	design	design	VERB
fcis-25702	47	14	several	several	ADJ
fcis-25702	47	15	strategies	strategy	NOUN
fcis-25702	47	16	such	such	ADJ
fcis-25702	47	17	as	as	ADP
fcis-25702	47	18	introducing	introduce	VERB
fcis-25702	47	19	adapters	adapter	NOUN
fcis-25702	47	20	and	and	CCONJ
fcis-25702	47	21	layer	layer	NOUN
fcis-25702	47	22	skipping	skip	VERB
fcis-25702	47	23	pre	pre	ADJ
fcis-25702	47	24	-	-	NOUN
fcis-25702	47	25	training	training	NOUN
fcis-25702	47	26	.	.	PUNCT
fcis-25702	48	1	we	we	PRON
fcis-25702	48	2	also	also	ADV
fcis-25702	48	3	refine	refine	VERB
fcis-25702	48	4	the	the	DET
fcis-25702	48	5	sampling	sample	VERB
fcis-25702	48	6	strategy	strategy	NOUN
fcis-25702	48	7	of	of	ADP
fcis-25702	48	8	rl	rl	PRON
fcis-25702	48	9	by	by	ADP
fcis-25702	48	10	not	not	PART
fcis-25702	48	11	skipping	skip	VERB
fcis-25702	48	12	layers	layer	NOUN
fcis-25702	48	13	that	that	PRON
fcis-25702	48	14	are	be	AUX
fcis-25702	48	15	dissimilar	dissimilar	ADJ
fcis-25702	48	16	to	to	ADP
fcis-25702	48	17	previous	previous	ADJ
fcis-25702	48	18	layers	layer	NOUN
fcis-25702	48	19	.	.	PUNCT
fcis-25702	49	1	we	we	PRON
fcis-25702	49	2	conduct	conduct	VERB
fcis-25702	49	3	experiments	experiment	NOUN
fcis-25702	49	4	on	on	ADP
fcis-25702	49	5	natural	natural	ADJ
fcis-25702	49	6	language	language	NOUN
fcis-25702	49	7	understanding	understanding	NOUN
fcis-25702	49	8	(	(	PUNCT
fcis-25702	49	9	nlu	nlu	NOUN
fcis-25702	49	10	)	)	PUNCT
fcis-25702	49	11	datasets	dataset	NOUN
fcis-25702	49	12	and	and	CCONJ
fcis-25702	49	13	machine	machine	NOUN
fcis-25702	49	14	translation	translation	NOUN
fcis-25702	49	15	(	(	PUNCT
fcis-25702	49	16	mt	mt	PROPN
fcis-25702	49	17	)	)	PUNCT
fcis-25702	49	18	task	task	NOUN
fcis-25702	49	19	,	,	PUNCT
fcis-25702	49	20	formulating	formulate	VERB
fcis-25702	49	21	this	this	DET
fcis-25702	49	22	task	task	NOUN
fcis-25702	49	23	as	as	ADP
fcis-25702	49	24	a	a	DET
fcis-25702	49	25	prompt	prompt	NOUN
fcis-25702	49	26	-	-	PUNCT
fcis-25702	49	27	based	base	VERB
fcis-25702	49	28	language	language	NOUN
fcis-25702	49	29	modeling	modeling	NOUN
fcis-25702	49	30	task	task	NOUN
fcis-25702	49	31	using	use	VERB
fcis-25702	49	32	instructions	instruction	NOUN
fcis-25702	49	33	.	.	PUNCT
fcis-25702	50	1	the	the	DET
fcis-25702	50	2	experiments	experiment	NOUN
fcis-25702	50	3	on	on	ADP
fcis-25702	50	4	four	four	NUM
fcis-25702	50	5	benchmark	benchmark	NOUN
fcis-25702	50	6	nlu	nlu	NOUN
fcis-25702	50	7	datasets	dataset	NOUN
fcis-25702	50	8	and	and	CCONJ
fcis-25702	50	9	three	three	NUM
fcis-25702	50	10	mt	mt	PROPN
fcis-25702	50	11	datasets	dataset	NOUN
fcis-25702	50	12	show	show	VERB
fcis-25702	50	13	that	that	SCONJ
fcis-25702	50	14	our	our	PRON
fcis-25702	50	15	work	work	NOUN
fcis-25702	50	16	achieves	achieve	VERB
fcis-25702	50	17	state	state	NOUN
fcis-25702	50	18	-	-	PUNCT
fcis-25702	50	19	of	of	ADP
fcis-25702	50	20	-	-	PUNCT
fcis-25702	50	21	the	the	DET
fcis-25702	50	22	-	-	PUNCT
fcis-25702	50	23	art	art	NOUN
fcis-25702	50	24	performance	performance	NOUN
fcis-25702	50	25	compared	compare	VERB
fcis-25702	50	26	to	to	ADP
fcis-25702	50	27	previous	previous	ADJ
fcis-25702	50	28	work	work	NOUN
fcis-25702	50	29	.	.	PUNCT
fcis-25702	51	1	our	our	PRON
fcis-25702	51	2	contributions	contribution	NOUN
fcis-25702	51	3	are	be	AUX
fcis-25702	51	4	as	as	SCONJ
fcis-25702	51	5	follows	follow	VERB
fcis-25702	51	6	:	:	PUNCT
fcis-25702	51	7	(	(	PUNCT
fcis-25702	51	8	1	1	X
fcis-25702	51	9	)	)	PUNCT
fcis-25702	51	10	to	to	ADP
fcis-25702	51	11	the	the	DET
fcis-25702	51	12	best	good	ADJ
fcis-25702	51	13	of	of	ADP
fcis-25702	51	14	our	our	PRON
fcis-25702	51	15	knowledge	knowledge	NOUN
fcis-25702	51	16	,	,	PUNCT
fcis-25702	51	17	we	we	PRON
fcis-25702	51	18	are	be	AUX
fcis-25702	51	19	the	the	DET
fcis-25702	51	20	first	first	ADJ
fcis-25702	51	21	to	to	PART
fcis-25702	51	22	apply	apply	VERB
fcis-25702	51	23	a	a	DET
fcis-25702	51	24	dynamic	dynamic	ADJ
fcis-25702	51	25	inference	inference	NOUN
fcis-25702	51	26	method	method	NOUN
fcis-25702	51	27	to	to	ADP
fcis-25702	51	28	large	large	ADJ
fcis-25702	51	29	language	language	NOUN
fcis-25702	51	30	models	model	NOUN
fcis-25702	51	31	to	to	PART
fcis-25702	51	32	achieve	achieve	VERB
fcis-25702	51	33	llm	llm	PROPN
fcis-25702	51	34	inference	inference	NOUN
fcis-25702	51	35	speed	speed	NOUN
fcis-25702	51	36	-	-	PUNCT
fcis-25702	51	37	up	up	NOUN
fcis-25702	51	38	.	.	PUNCT
fcis-25702	52	1	(	(	PUNCT
fcis-25702	52	2	2	2	X
fcis-25702	52	3	)	)	PUNCT
fcis-25702	52	4	we	we	PRON
fcis-25702	52	5	propose	propose	VERB
fcis-25702	52	6	a	a	DET
fcis-25702	52	7	novel	novel	ADJ
fcis-25702	52	8	framework	framework	NOUN
fcis-25702	52	9	that	that	PRON
fcis-25702	52	10	uses	use	VERB
fcis-25702	52	11	an	an	DET
fcis-25702	52	12	adapter	adapter	NOUN
fcis-25702	52	13	layer	layer	NOUN
fcis-25702	52	14	to	to	PART
fcis-25702	52	15	dynamically	dynamically	ADV
fcis-25702	52	16	skip	skip	VERB
fcis-25702	52	17	llm	llm	ADJ
fcis-25702	52	18	layers	layer	NOUN
fcis-25702	52	19	without	without	ADP
fcis-25702	52	20	introducing	introduce	VERB
fcis-25702	52	21	high	high	ADJ
fcis-25702	52	22	computational	computational	ADJ
fcis-25702	52	23	costs	cost	NOUN
fcis-25702	52	24	.	.	PUNCT
fcis-25702	53	1	(	(	PUNCT
fcis-25702	53	2	3	3	X
fcis-25702	53	3	)	)	PUNCT
fcis-25702	53	4	we	we	PRON
fcis-25702	53	5	propose	propose	VERB
fcis-25702	53	6	using	use	VERB
fcis-25702	53	7	reinforcement	reinforcement	NOUN
fcis-25702	53	8	learning	learning	NOUN
fcis-25702	53	9	to	to	PART
fcis-25702	53	10	optimize	optimize	VERB
fcis-25702	53	11	our	our	PRON
fcis-25702	53	12	framework	framework	NOUN
fcis-25702	53	13	,	,	PUNCT
fcis-25702	53	14	and	and	CCONJ
fcis-25702	53	15	we	we	PRON
fcis-25702	53	16	introduce	introduce	VERB
fcis-25702	53	17	some	some	DET
fcis-25702	53	18	sampling	sample	VERB
fcis-25702	53	19	strategies	strategy	NOUN
fcis-25702	53	20	to	to	PART
fcis-25702	53	21	improve	improve	VERB
fcis-25702	53	22	the	the	DET
fcis-25702	53	23	training	training	NOUN
fcis-25702	53	24	of	of	ADP
fcis-25702	53	25	rl	rl	PROPN
fcis-25702	53	26	.	.	PUNCT
fcis-25702	54	1	(	(	PUNCT
fcis-25702	54	2	4	4	X
fcis-25702	54	3	)	)	PUNCT
fcis-25702	54	4	we	we	PRON
fcis-25702	54	5	conduct	conduct	VERB
fcis-25702	54	6	experiments	experiment	NOUN
fcis-25702	54	7	to	to	PART
fcis-25702	54	8	show	show	VERB
fcis-25702	54	9	that	that	SCONJ
fcis-25702	54	10	our	our	PRON
fcis-25702	54	11	method	method	NOUN
fcis-25702	54	12	achieves	achieve	VERB
fcis-25702	54	13	sota	sota	ADJ
fcis-25702	54	14	performance	performance	NOUN
fcis-25702	54	15	for	for	ADP
fcis-25702	54	16	llm	llm	NOUN
fcis-25702	54	17	inference	inference	NOUN
fcis-25702	54	18	speed	speed	NOUN
fcis-25702	54	19	-	-	PUNCT
fcis-25702	54	20	up	up	NOUN
fcis-25702	54	21	.	.	PUNCT
fcis-25702	55	1	in	in	ADP
fcis-25702	55	2	addition	addition	NOUN
fcis-25702	55	3	,	,	PUNCT
fcis-25702	55	4	ablations	ablation	NOUN
fcis-25702	55	5	and	and	CCONJ
fcis-25702	55	6	intrinsic	intrinsic	ADJ
fcis-25702	55	7	evaluations	evaluation	NOUN
fcis-25702	55	8	demonstrate	demonstrate	VERB
fcis-25702	55	9	the	the	DET
fcis-25702	55	10	effectiveness	effectiveness	NOUN
fcis-25702	55	11	of	of	ADP
fcis-25702	55	12	our	our	PRON
fcis-25702	55	13	proposed	propose	VERB
fcis-25702	55	14	method	method	NOUN
fcis-25702	55	15	.	.	PUNCT
fcis-25702	56	1	2	2	X
fcis-25702	56	2	.	.	NUM
fcis-25702	56	3	preliminaries	preliminary	NOUN
fcis-25702	56	4	due	due	ADP
fcis-25702	56	5	to	to	ADP
fcis-25702	56	6	the	the	DET
fcis-25702	56	7	length	length	NOUN
fcis-25702	56	8	limit	limit	NOUN
fcis-25702	56	9	,	,	PUNCT
fcis-25702	56	10	readers	reader	NOUN
fcis-25702	56	11	are	be	AUX
fcis-25702	56	12	referred	refer	VERB
fcis-25702	56	13	to	to	AUX
fcis-25702	56	14	appendix	appendix	VERB
fcis-25702	56	15	a.1	a.1	PROPN
fcis-25702	56	16	for	for	ADP
fcis-25702	56	17	llm	llm	PROPN
fcis-25702	56	18	applications	application	NOUN
fcis-25702	56	19	on	on	ADP
fcis-25702	56	20	natural	natural	ADJ
fcis-25702	56	21	language	language	NOUN
fcis-25702	56	22	understanding	understanding	NOUN
fcis-25702	56	23	(	(	PUNCT
fcis-25702	56	24	nlu	nlu	NOUN
fcis-25702	56	25	)	)	PUNCT
fcis-25702	56	26	datasets	dataset	NOUN
fcis-25702	56	27	.	.	PUNCT
fcis-25702	57	1	2.1	2.1	NUM
fcis-25702	57	2	.	.	PUNCT
fcis-25702	57	3	layer	layer	NOUN
fcis-25702	57	4	skipping	skip	VERB
fcis-25702	57	5	skip	skip	NOUN
fcis-25702	57	6	-	-	PUNCT
fcis-25702	57	7	decode	decode	NOUN
fcis-25702	57	8	[	[	X
fcis-25702	57	9	33	33	NUM
fcis-25702	57	10	]	]	PUNCT
fcis-25702	57	11	is	be	AUX
fcis-25702	57	12	the	the	DET
fcis-25702	57	13	first	first	ADJ
fcis-25702	57	14	to	to	PART
fcis-25702	57	15	propose	propose	VERB
fcis-25702	57	16	a	a	DET
fcis-25702	57	17	layer	layer	NOUN
fcis-25702	57	18	skipping	skip	VERB
fcis-25702	57	19	method	method	NOUN
fcis-25702	57	20	,	,	PUNCT
fcis-25702	57	21	which	which	PRON
fcis-25702	57	22	can	can	AUX
fcis-25702	57	23	be	be	AUX
fcis-25702	57	24	viewed	view	VERB
fcis-25702	57	25	as	as	ADP
fcis-25702	57	26	another	another	DET
fcis-25702	57	27	form	form	NOUN
fcis-25702	57	28	of	of	ADP
fcis-25702	57	29	early	early	ADJ
fcis-25702	57	30	exiting	exiting	NOUN
fcis-25702	57	31	.	.	PUNCT
fcis-25702	58	1	the	the	DET
fcis-25702	58	2	main	main	ADJ
fcis-25702	58	3	difference	difference	NOUN
fcis-25702	58	4	is	be	AUX
fcis-25702	58	5	that	that	SCONJ
fcis-25702	58	6	layer	layer	NOUN
fcis-25702	58	7	skipping	skip	VERB
fcis-25702	58	8	can	can	AUX
fcis-25702	58	9	bypass	bypass	VERB
fcis-25702	58	10	any	any	DET
fcis-25702	58	11	layer	layer	NOUN
fcis-25702	58	12	,	,	PUNCT
fcis-25702	58	13	while	while	SCONJ
fcis-25702	58	14	early	early	ADJ
fcis-25702	58	15	exiting	exiting	NOUN
fcis-25702	58	16	exits	exit	VERB
fcis-25702	58	17	at	at	ADP
fcis-25702	58	18	a	a	DET
fcis-25702	58	19	specific	specific	ADJ
fcis-25702	58	20	layer	layer	NOUN
fcis-25702	58	21	and	and	CCONJ
fcis-25702	58	22	stops	stop	VERB
fcis-25702	58	23	computing	compute	VERB
fcis-25702	58	24	the	the	DET
fcis-25702	58	25	subsequent	subsequent	ADJ
fcis-25702	58	26	layers	layer	NOUN
fcis-25702	58	27	.	.	PUNCT
fcis-25702	59	1	however	however	ADV
fcis-25702	59	2	,	,	PUNCT
fcis-25702	59	3	it	it	PRON
fcis-25702	59	4	is	be	AUX
fcis-25702	59	5	very	very	ADV
fcis-25702	59	6	challenging	challenging	ADJ
fcis-25702	59	7	to	to	PART
fcis-25702	59	8	apply	apply	VERB
fcis-25702	59	9	early	early	ADV
fcis-25702	59	10	exiting	exit	VERB
fcis-25702	59	11	to	to	ADP
fcis-25702	59	12	llms	llm	NOUN
fcis-25702	59	13	,	,	PUNCT
fcis-25702	59	14	as	as	SCONJ
fcis-25702	59	15	early	early	ADJ
fcis-25702	59	16	exiting	exit	VERB
fcis-25702	59	17	methods	method	NOUN
fcis-25702	59	18	install	install	VERB
fcis-25702	59	19	an	an	DET
fcis-25702	59	20	early	early	ADJ
fcis-25702	59	21	exit	exit	NOUN
fcis-25702	59	22	layer	layer	NOUN
fcis-25702	59	23	,	,	PUNCT
fcis-25702	59	24	which	which	PRON
fcis-25702	59	25	is	be	AUX
fcis-25702	59	26	a	a	DET
fcis-25702	59	27	classifier	classifier	NOUN
fcis-25702	59	28	layer	layer	NOUN
fcis-25702	59	29	,	,	PUNCT
fcis-25702	59	30	at	at	ADP
fcis-25702	59	31	each	each	DET
fcis-25702	59	32	transformer	transformer	ADJ
fcis-25702	59	33	layer	layer	NOUN
fcis-25702	59	34	to	to	PART
fcis-25702	59	35	compute	compute	VERB
fcis-25702	59	36	the	the	DET
fcis-25702	59	37	entropy	entropy	NOUN
fcis-25702	59	38	of	of	ADP
fcis-25702	59	39	the	the	DET
fcis-25702	59	40	logits	logit	NOUN
fcis-25702	59	41	and	and	CCONJ
fcis-25702	59	42	predict	predict	VERB
fcis-25702	59	43	whether	whether	SCONJ
fcis-25702	59	44	to	to	PART
fcis-25702	59	45	exit	exit	VERB
fcis-25702	59	46	.	.	PUNCT
fcis-25702	60	1	however	however	ADV
fcis-25702	60	2	,	,	PUNCT
fcis-25702	60	3	for	for	ADP
fcis-25702	60	4	ll	ll	PROPN
fcis-25702	60	5	ms	ms	PROPN
fcis-25702	60	6	,	,	PUNCT
fcis-25702	60	7	the	the	DET
fcis-25702	60	8	classifier	classifier	NOUN
fcis-25702	60	9	layer	layer	NOUN
fcis-25702	60	10	would	would	AUX
fcis-25702	60	11	be	be	AUX
fcis-25702	60	12	a	a	DET
fcis-25702	60	13	language	language	NOUN
fcis-25702	60	14	modeling	modeling	NOUN
fcis-25702	60	15	(	(	PUNCT
fcis-25702	60	16	lm	lm	ADJ
fcis-25702	60	17	)	)	PUNCT
fcis-25702	60	18	head	head	NOUN
fcis-25702	60	19	classifier	classifier	NOUN
fcis-25702	60	20	,	,	PUNCT
fcis-25702	60	21	typically	typically	ADV
fcis-25702	60	22	larger	large	ADJ
fcis-25702	60	23	than	than	ADP
fcis-25702	60	24	100,000	100,000	NUM
fcis-25702	60	25	units	unit	NOUN
fcis-25702	60	26	.	.	PUNCT
fcis-25702	61	1	additionally	additionally	ADV
fcis-25702	61	2	,	,	PUNCT
fcis-25702	61	3	since	since	SCONJ
fcis-25702	61	4	llms	llm	NOUN
fcis-25702	61	5	often	often	ADV
fcis-25702	61	6	have	have	VERB
fcis-25702	61	7	many	many	ADJ
fcis-25702	61	8	deep	deep	ADJ
fcis-25702	61	9	layers	layer	NOUN
fcis-25702	61	10	,	,	PUNCT
fcis-25702	61	11	such	such	ADJ
fcis-25702	61	12	as	as	ADP
fcis-25702	61	13	36	36	NUM
fcis-25702	61	14	or	or	CCONJ
fcis-25702	61	15	48	48	NUM
fcis-25702	61	16	,	,	PUNCT
fcis-25702	61	17	installing	instal	VERB
fcis-25702	61	18	an	an	DET
fcis-25702	61	19	lm	lm	ADJ
fcis-25702	61	20	head	head	NOUN
fcis-25702	61	21	at	at	ADP
fcis-25702	61	22	each	each	DET
fcis-25702	61	23	layer	layer	NOUN
fcis-25702	61	24	would	would	AUX
fcis-25702	61	25	incur	incur	VERB
fcis-25702	61	26	unbearably	unbearably	ADV
fcis-25702	61	27	high	high	ADJ
fcis-25702	61	28	computational	computational	ADJ
fcis-25702	61	29	costs	cost	NOUN
fcis-25702	61	30	.	.	PUNCT
fcis-25702	62	1	therefore	therefore	ADV
fcis-25702	62	2	,	,	PUNCT
fcis-25702	62	3	skip	skip	VERB
fcis-25702	62	4	-	-	PUNCT
fcis-25702	62	5	decode	decode	NOUN
fcis-25702	62	6	directly	directly	ADV
fcis-25702	62	7	assigns	assign	VERB
fcis-25702	62	8	the	the	DET
fcis-25702	62	9	number	number	NOUN
fcis-25702	62	10	of	of	ADP
fcis-25702	62	11	layers	layer	NOUN
fcis-25702	62	12	to	to	PART
fcis-25702	62	13	skip	skip	VERB
fcis-25702	62	14	instead	instead	ADV
fcis-25702	62	15	of	of	ADP
fcis-25702	62	16	assessing	assess	VERB
fcis-25702	62	17	each	each	DET
fcis-25702	62	18	transformer	transformer	NOUN
fcis-25702	62	19	layer	layer	NOUN
fcis-25702	62	20	.	.	PUNCT
fcis-25702	63	1	the	the	DET
fcis-25702	63	2	model	model	NOUN
fcis-25702	63	3	gradually	gradually	ADV
fcis-25702	63	4	skips	skip	VERB
fcis-25702	63	5	more	more	ADV
fcis-25702	63	6	intermediate	intermediate	ADJ
fcis-25702	63	7	layers	layer	NOUN
fcis-25702	63	8	as	as	SCONJ
fcis-25702	63	9	it	it	PRON
fcis-25702	63	10	generates	generate	VERB
fcis-25702	63	11	more	more	ADJ
fcis-25702	63	12	tokens	token	NOUN
fcis-25702	63	13	.	.	PUNCT
fcis-25702	64	1	it	it	PRON
fcis-25702	64	2	applies	apply	VERB
fcis-25702	64	3	the	the	DET
fcis-25702	64	4	same	same	ADJ
fcis-25702	64	5	skipping	skip	VERB
fcis-25702	64	6	layers	layer	NOUN
fcis-25702	64	7	for	for	ADP
fcis-25702	64	8	all	all	DET
fcis-25702	64	9	samples	sample	NOUN
fcis-25702	64	10	,	,	PUNCT
fcis-25702	64	11	which	which	PRON
fcis-25702	64	12	is	be	AUX
fcis-25702	64	13	not	not	PART
fcis-25702	64	14	dynamic	dynamic	ADJ
fcis-25702	64	15	.	.	PUNCT
fcis-25702	65	1	additionally	additionally	ADV
fcis-25702	65	2	,	,	PUNCT
fcis-25702	65	3	this	this	DET
fcis-25702	65	4	approach	approach	NOUN
fcis-25702	65	5	assumes	assume	VERB
fcis-25702	65	6	that	that	SCONJ
fcis-25702	65	7	the	the	DET
fcis-25702	65	8	tokens	token	NOUN
fcis-25702	65	9	at	at	ADP
fcis-25702	65	10	the	the	DET
fcis-25702	65	11	end	end	NOUN
fcis-25702	65	12	of	of	ADP
fcis-25702	65	13	the	the	DET
fcis-25702	65	14	responses	response	NOUN
fcis-25702	65	15	are	be	AUX
fcis-25702	65	16	simpler	simple	ADJ
fcis-25702	65	17	than	than	ADP
fcis-25702	65	18	those	those	PRON
fcis-25702	65	19	at	at	ADP
fcis-25702	65	20	the	the	DET
fcis-25702	65	21	beginning	beginning	NOUN
fcis-25702	65	22	.	.	PUNCT
fcis-25702	66	1	a	a	DET
fcis-25702	66	2	counterexample	counterexample	NOUN
fcis-25702	66	3	is	be	AUX
fcis-25702	66	4	using	use	VERB
fcis-25702	66	5	llms	llm	NOUN
fcis-25702	66	6	for	for	ADP
fcis-25702	66	7	the	the	DET
fcis-25702	66	8	named	name	VERB
fcis-25702	66	9	entity	entity	NOUN
fcis-25702	66	10	recognition	recognition	NOUN
fcis-25702	66	11	task	task	NOUN
fcis-25702	66	12	,	,	PUNCT
fcis-25702	66	13	where	where	SCONJ
fcis-25702	66	14	the	the	DET
fcis-25702	66	15	llm	llm	NOUN
fcis-25702	66	16	generates	generate	VERB
fcis-25702	66	17	many	many	ADJ
fcis-25702	66	18	entities	entity	NOUN
fcis-25702	66	19	mentions	mention	NOUN
fcis-25702	66	20	as	as	ADP
fcis-25702	66	21	results	result	NOUN
fcis-25702	66	22	.	.	PUNCT
fcis-25702	67	1	however	however	ADV
fcis-25702	67	2	,	,	PUNCT
fcis-25702	67	3	the	the	DET
fcis-25702	67	4	entity	entity	NOUN
fcis-25702	67	5	mentions	mention	NOUN
fcis-25702	67	6	generated	generate	VERB
fcis-25702	67	7	at	at	ADP
fcis-25702	67	8	later	later	ADJ
fcis-25702	67	9	steps	step	NOUN
fcis-25702	67	10	are	be	AUX
fcis-25702	67	11	just	just	ADV
fcis-25702	67	12	as	as	ADV
fcis-25702	67	13	important	important	ADJ
fcis-25702	67	14	as	as	ADP
fcis-25702	67	15	those	those	PRON
fcis-25702	67	16	at	at	ADP
fcis-25702	67	17	earlier	early	ADJ
fcis-25702	67	18	steps	step	NOUN
fcis-25702	67	19	,	,	PUNCT
fcis-25702	67	20	making	make	VERB
fcis-25702	67	21	it	it	PRON
fcis-25702	67	22	unfair	unfair	ADJ
fcis-25702	67	23	to	to	PART
fcis-25702	67	24	assign	assign	VERB
fcis-25702	67	25	fewer	few	ADJ
fcis-25702	67	26	layers	layer	NOUN
fcis-25702	67	27	to	to	ADP
fcis-25702	67	28	these	these	DET
fcis-25702	67	29	tokens	token	NOUN
fcis-25702	67	30	at	at	ADP
fcis-25702	67	31	later	later	ADJ
fcis-25702	67	32	generation	generation	NOUN
fcis-25702	67	33	steps	step	NOUN
fcis-25702	67	34	.	.	PUNCT
fcis-25702	68	1	also	also	ADV
fcis-25702	68	2	,	,	PUNCT
fcis-25702	68	3	skip	skip	ADJ
fcis-25702	68	4	-	-	PUNCT
fcis-25702	68	5	decode	decode	NOUN
fcis-25702	68	6	skips	skip	VERB
fcis-25702	68	7	the	the	DET
fcis-25702	68	8	shallow	shallow	ADJ
fcis-25702	68	9	layers	layer	NOUN
fcis-25702	68	10	first	first	ADV
fcis-25702	68	11	and	and	CCONJ
fcis-25702	68	12	then	then	ADV
fcis-25702	68	13	the	the	DET
fcis-25702	68	14	deep	deep	ADJ
fcis-25702	68	15	layers	layer	NOUN
fcis-25702	68	16	.	.	PUNCT
fcis-25702	69	1	much	much	ADJ
fcis-25702	69	2	work	work	NOUN
fcis-25702	70	1	[	[	X
fcis-25702	70	2	29][24][35	29][24][35	NOUN
fcis-25702	70	3	]	]	PUNCT
fcis-25702	70	4	has	have	AUX
fcis-25702	70	5	demonstrated	demonstrate	VERB
fcis-25702	70	6	that	that	SCONJ
fcis-25702	70	7	the	the	DET
fcis-25702	70	8	later	later	ADJ
fcis-25702	70	9	layers	layer	NOUN
fcis-25702	70	10	of	of	ADP
fcis-25702	70	11	bert	bert	PROPN
fcis-25702	70	12	or	or	CCONJ
fcis-25702	70	13	llms	llm	NOUN
fcis-25702	70	14	have	have	VERB
fcis-25702	70	15	more	more	ADJ
fcis-25702	70	16	redundancy	redundancy	NOUN
fcis-25702	70	17	compared	compare	VERB
fcis-25702	70	18	to	to	ADP
fcis-25702	70	19	earlier	early	ADJ
fcis-25702	70	20	layers	layer	NOUN
fcis-25702	70	21	.	.	PUNCT
fcis-25702	71	1	thus	thus	ADV
fcis-25702	71	2	,	,	PUNCT
fcis-25702	71	3	we	we	PRON
fcis-25702	71	4	aim	aim	VERB
fcis-25702	71	5	to	to	PART
fcis-25702	71	6	develop	develop	VERB
fcis-25702	71	7	an	an	DET
fcis-25702	71	8	adaptive	adaptive	ADJ
fcis-25702	71	9	method	method	NOUN
fcis-25702	71	10	to	to	PART
fcis-25702	71	11	automatically	automatically	ADV
fcis-25702	71	12	determine	determine	VERB
fcis-25702	71	13	which	which	DET
fcis-25702	71	14	layers	layer	NOUN
fcis-25702	71	15	to	to	PART
fcis-25702	71	16	skip	skip	VERB
fcis-25702	71	17	for	for	ADP
fcis-25702	71	18	predicting	predict	VERB
fcis-25702	71	19	each	each	DET
fcis-25702	71	20	token	token	ADJ
fcis-25702	71	21	.	.	PUNCT
fcis-25702	72	1	2.2	2.2	NUM
fcis-25702	72	2	.	.	PUNCT
fcis-25702	73	1	problem	problem	NOUN
fcis-25702	73	2	statement	statement	NOUN
fcis-25702	73	3	we	we	PRON
fcis-25702	73	4	apply	apply	VERB
fcis-25702	73	5	prompts	prompt	NOUN
fcis-25702	73	6	to	to	PART
fcis-25702	73	7	use	use	VERB
fcis-25702	73	8	llms	llm	NOUN
fcis-25702	73	9	on	on	ADP
fcis-25702	73	10	nlu	nlu	NOUN
fcis-25702	73	11	and	and	CCONJ
fcis-25702	73	12	mt	mt	PROPN
fcis-25702	73	13	datasets	dataset	NOUN
fcis-25702	73	14	.	.	PUNCT
fcis-25702	74	1	for	for	ADP
fcis-25702	74	2	each	each	DET
fcis-25702	74	3	dataset	dataset	NOUN
fcis-25702	74	4	,	,	PUNCT
fcis-25702	74	5	we	we	PRON
fcis-25702	74	6	design	design	VERB
fcis-25702	74	7	a	a	DET
fcis-25702	74	8	specific	specific	ADJ
fcis-25702	74	9	prompt	prompt	NOUN
fcis-25702	74	10	,	,	PUNCT
fcis-25702	74	11	and	and	CCONJ
fcis-25702	74	12	during	during	ADP
fcis-25702	74	13	training	training	NOUN
fcis-25702	74	14	,	,	PUNCT
fcis-25702	74	15	we	we	PRON
fcis-25702	74	16	combine	combine	VERB
fcis-25702	74	17	all	all	DET
fcis-25702	74	18	datasets	dataset	NOUN
fcis-25702	74	19	to	to	PART
fcis-25702	74	20	train	train	VERB
fcis-25702	74	21	one	one	NUM
fcis-25702	74	22	model	model	NOUN
fcis-25702	74	23	.	.	PUNCT
fcis-25702	75	1	we	we	PRON
fcis-25702	75	2	show	show	VERB
fcis-25702	75	3	our	our	PRON
fcis-25702	75	4	prompt	prompt	ADJ
fcis-25702	75	5	design	design	NOUN
fcis-25702	75	6	in	in	ADP
fcis-25702	75	7	appendix	appendix	NOUN
fcis-25702	75	8	a.2	a.2	PUNCT
fcis-25702	75	9	.	.	PUNCT
fcis-25702	76	1	the	the	DET
fcis-25702	76	2	prompt	prompt	NOUN
fcis-25702	76	3	is	be	AUX
fcis-25702	76	4	the	the	DET
fcis-25702	76	5	input	input	NOUN
fcis-25702	76	6	to	to	ADP
fcis-25702	76	7	llms	llm	NOUN
fcis-25702	76	8	that	that	PRON
fcis-25702	76	9	contains	contain	VERB
fcis-25702	76	10	the	the	DET
fcis-25702	76	11	instructions	instruction	NOUN
fcis-25702	76	12	and	and	CCONJ
fcis-25702	76	13	the	the	DET
fcis-25702	76	14	input	input	NOUN
fcis-25702	76	15	sentence	sentence	NOUN
fcis-25702	76	16	to	to	PART
fcis-25702	76	17	be	be	AUX
fcis-25702	76	18	classified	classify	VERB
fcis-25702	76	19	.	.	PUNCT
fcis-25702	77	1	denote	denote	VERB
fcis-25702	77	2	the	the	DET
fcis-25702	77	3	prompt	prompt	NOUN
fcis-25702	77	4	as	as	ADP
fcis-25702	77	5	w	w	PROPN
fcis-25702	77	6	,	,	PUNCT
fcis-25702	77	7	w	w	PROPN
fcis-25702	77	8	,	,	PUNCT
fcis-25702	77	9	w	w	PROPN
fcis-25702	77	10	,	,	PUNCT
fcis-25702	77	11	…	…	PUNCT
fcis-25702	77	12	,	,	PUNCT
fcis-25702	77	13	w	w	NOUN
fcis-25702	77	14	,	,	PUNCT
fcis-25702	77	15	and	and	CCONJ
fcis-25702	77	16	the	the	DET
fcis-25702	77	17	llm	llm	NOUN
fcis-25702	77	18	we	we	PRON
fcis-25702	77	19	use	use	VERB
fcis-25702	77	20	has	have	VERB
fcis-25702	77	21	l	l	NOUN
fcis-25702	77	22	layers	layer	NOUN
fcis-25702	77	23	.	.	PUNCT
fcis-25702	78	1	for	for	ADP
fcis-25702	78	2	generating	generate	VERB
fcis-25702	78	3	responses	response	NOUN
fcis-25702	78	4	,	,	PUNCT
fcis-25702	78	5	the	the	DET
fcis-25702	78	6	prompt	prompt	ADJ
fcis-25702	78	7	sequence	sequence	NOUN
fcis-25702	78	8	will	will	AUX
fcis-25702	78	9	be	be	AUX
fcis-25702	78	10	embedded	embed	VERB
fcis-25702	78	11	in	in	ADP
fcis-25702	78	12	the	the	DET
fcis-25702	78	13	embedding	embed	VERB
fcis-25702	78	14	layer	layer	NOUN
fcis-25702	78	15	:	:	PUNCT
fcis-25702	78	16	h	h	NOUN
fcis-25702	78	17	h	h	NOUN
fcis-25702	78	18	,	,	PUNCT
fcis-25702	78	19	h	h	NOUN
fcis-25702	78	20	,	,	PUNCT
fcis-25702	78	21	…	…	PUNCT
fcis-25702	78	22	,	,	PUNCT
fcis-25702	79	1	h	h	NOUN
fcis-25702	79	2	embedding	embed	VERB
fcis-25702	79	3	w	w	ADP
fcis-25702	79	4	,	,	PUNCT
fcis-25702	79	5	w	w	PROPN
fcis-25702	79	6	,	,	PUNCT
fcis-25702	79	7	…	…	PUNCT
fcis-25702	79	8	,	,	PUNCT
fcis-25702	79	9	w	w	NOUN
fcis-25702	79	10	(	(	PUNCT
fcis-25702	79	11	1	1	NUM
fcis-25702	79	12	)	)	PUNCT
fcis-25702	79	13	and	and	CCONJ
fcis-25702	79	14	go	go	VERB
fcis-25702	79	15	through	through	ADP
fcis-25702	79	16	the	the	DET
fcis-25702	79	17	transformer	transformer	ADJ
fcis-25702	79	18	layers	layer	NOUN
fcis-25702	79	19	:	:	PUNCT
fcis-25702	79	20	h	h	NOUN
fcis-25702	79	21	h	h	NOUN
fcis-25702	79	22	,	,	PUNCT
fcis-25702	79	23	h	h	NOUN
fcis-25702	79	24	,	,	PUNCT
fcis-25702	79	25	…	…	PUNCT
fcis-25702	79	26	,	,	PUNCT
fcis-25702	79	27	h	h	PROPN
fcis-25702	79	28	trans	trans	PROPN
fcis-25702	79	29	h	h	PROPN
fcis-25702	79	30	,	,	PUNCT
fcis-25702	79	31	h	h	NOUN
fcis-25702	79	32	,	,	PUNCT
fcis-25702	79	33	…	…	PUNCT
fcis-25702	79	34	,	,	PUNCT
fcis-25702	79	35	h	h	NOUN
fcis-25702	79	36	(	(	PUNCT
fcis-25702	79	37	2	2	NUM
fcis-25702	79	38	)	)	PUNCT
fcis-25702	79	39	for	for	ADP
fcis-25702	79	40	i	i	PROPN
fcis-25702	79	41	0,1	0,1	NUM
fcis-25702	79	42	,	,	PUNCT
fcis-25702	79	43	…	…	PUNCT
fcis-25702	79	44	,	,	PUNCT
fcis-25702	79	45	l	l	NOUN
fcis-25702	79	46	1	1	X
fcis-25702	79	47	.	.	PUNCT
fcis-25702	80	1	at	at	ADP
fcis-25702	80	2	the	the	DET
fcis-25702	80	3	final	final	ADJ
fcis-25702	80	4	layer	layer	NOUN
fcis-25702	80	5	,	,	PUNCT
fcis-25702	80	6	the	the	DET
fcis-25702	80	7	hidden	hide	VERB
fcis-25702	80	8	representation	representation	NOUN
fcis-25702	80	9	h	h	NOUN
fcis-25702	80	10	will	will	AUX
fcis-25702	80	11	be	be	AUX
fcis-25702	80	12	fed	feed	VERB
fcis-25702	80	13	into	into	ADP
fcis-25702	80	14	the	the	DET
fcis-25702	80	15	language	language	NOUN
fcis-25702	80	16	modeling	model	VERB
fcis-25702	80	17	head	head	NOUN
fcis-25702	80	18	to	to	PART
fcis-25702	80	19	predict	predict	VERB
fcis-25702	80	20	the	the	DET
fcis-25702	80	21	next	next	ADJ
fcis-25702	80	22	token	token	VERB
fcis-25702	80	23	by	by	ADP
fcis-25702	80	24	calculating	calculate	VERB
fcis-25702	80	25	its	its	PRON
fcis-25702	80	26	probability	probability	NOUN
fcis-25702	80	27	distribution	distribution	NOUN
fcis-25702	80	28	with	with	ADP
fcis-25702	80	29	the	the	DET
fcis-25702	80	30	language	language	NOUN
fcis-25702	80	31	modeling	model	VERB
fcis-25702	80	32	head	head	NOUN
fcis-25702	80	33	:	:	PUNCT
fcis-25702	80	34	p	p	X
fcis-25702	80	35	w	w	PROPN
fcis-25702	80	36	∣	∣	PROPN
fcis-25702	80	37	w	w	PROPN
fcis-25702	80	38	,	,	PUNCT
fcis-25702	80	39	w	w	PROPN
fcis-25702	80	40	,	,	PUNCT
fcis-25702	80	41	w	w	PROPN
fcis-25702	80	42	,	,	PUNCT
fcis-25702	80	43	…	…	PUNCT
fcis-25702	80	44	,	,	PUNCT
fcis-25702	80	45	w	w	PROPN
fcis-25702	80	46	lmh	lmh	PROPN
fcis-25702	80	47	h	h	PROPN
fcis-25702	80	48	(	(	PUNCT
fcis-25702	80	49	3	3	X
fcis-25702	80	50	)	)	PUNCT
fcis-25702	80	51	we	we	PRON
fcis-25702	80	52	use	use	VERB
fcis-25702	80	53	the	the	DET
fcis-25702	80	54	cross	cross	NOUN
fcis-25702	80	55	entropy	entropy	NOUN
fcis-25702	80	56	loss	loss	NOUN
fcis-25702	80	57	to	to	ADP
fcis-25702	80	58	fine	fine	ADJ
fcis-25702	80	59	-	-	PUNCT
fcis-25702	80	60	tune	tune	NOUN
fcis-25702	80	61	the	the	DET
fcis-25702	80	62	llms	llm	NOUN
fcis-25702	80	63	:	:	PUNCT
fcis-25702	80	64	∑	∑	ADP
fcis-25702	80	65	   	   	SPACE
fcis-25702	80	66	y	y	PROPN
fcis-25702	80	67	log	log	VERB
fcis-25702	80	68	p	p	X
fcis-25702	80	69	(	(	PUNCT
fcis-25702	80	70	4	4	NUM
fcis-25702	80	71	)	)	PUNCT
fcis-25702	80	72	note	note	VERB
fcis-25702	80	73	that	that	SCONJ
fcis-25702	80	74	with	with	ADP
fcis-25702	80	75	the	the	DET
fcis-25702	80	76	attention	attention	NOUN
fcis-25702	80	77	mechanism	mechanism	NOUN
fcis-25702	80	78	of	of	ADP
fcis-25702	80	79	transformerbased	transformerbase	VERB
fcis-25702	80	80	models	model	NOUN
fcis-25702	80	81	,	,	PUNCT
fcis-25702	80	82	h	h	NOUN
fcis-25702	80	83	contains	contain	VERB
fcis-25702	80	84	information	information	NOUN
fcis-25702	80	85	about	about	ADP
fcis-25702	80	86	the	the	DET
fcis-25702	80	87	tokens	token	NOUN
fcis-25702	80	88	in	in	ADP
fcis-25702	80	89	the	the	DET
fcis-25702	80	90	previous	previous	ADJ
fcis-25702	80	91	steps	step	NOUN
fcis-25702	80	92	.	.	PUNCT
fcis-25702	81	1	p	p	NOUN
fcis-25702	81	2	)	)	PUNCT
fcis-25702	81	3	denotes	denote	VERB
fcis-25702	81	4	the	the	DET
fcis-25702	81	5	calculated	calculated	ADJ
fcis-25702	81	6	distribution	distribution	NOUN
fcis-25702	81	7	over	over	ADP
fcis-25702	81	8	the	the	DET
fcis-25702	81	9	vocabulary	vocabulary	NOUN
fcis-25702	81	10	,	,	PUNCT
fcis-25702	81	11	y	y	PROPN
fcis-25702	81	12	is	be	AUX
fcis-25702	81	13	the	the	DET
fcis-25702	81	14	ground	ground	NOUN
fcis-25702	81	15	truth	truth	NOUN
fcis-25702	81	16	.	.	PUNCT
fcis-25702	82	1	with	with	ADP
fcis-25702	82	2	a	a	DET
fcis-25702	82	3	given	give	VERB
fcis-25702	82	4	decoding	decode	VERB
fcis-25702	82	5	algorithm	algorithm	NOUN
fcis-25702	82	6	like	like	ADP
fcis-25702	82	7	beam	beam	NOUN
fcis-25702	82	8	search	search	NOUN
fcis-25702	82	9	or	or	CCONJ
fcis-25702	82	10	nucleus	nucleus	NOUN
fcis-25702	82	11	sampling	sample	VERB
fcis-25702	82	12	[	[	X
fcis-25702	82	13	36	36	NUM
fcis-25702	82	14	]	]	PUNCT
fcis-25702	82	15	,	,	PUNCT
fcis-25702	82	16	one	one	PRON
fcis-25702	82	17	can	can	AUX
fcis-25702	82	18	select	select	VERB
fcis-25702	82	19	the	the	DET
fcis-25702	82	20	predicted	predict	VERB
fcis-25702	82	21	next	next	ADJ
fcis-25702	82	22	token	token	PROPN
fcis-25702	82	23	w	w	PROPN
fcis-25702	82	24	,	,	PUNCT
fcis-25702	82	25	and	and	CCONJ
fcis-25702	82	26	the	the	DET
fcis-25702	82	27	input	input	NOUN
fcis-25702	82	28	sequence	sequence	NOUN
fcis-25702	82	29	will	will	AUX
fcis-25702	82	30	be	be	AUX
fcis-25702	82	31	augmented	augment	VERB
fcis-25702	82	32	to	to	ADP
fcis-25702	82	33	x	x	PROPN
fcis-25702	82	34	w	w	PROPN
fcis-25702	82	35	,	,	PUNCT
fcis-25702	82	36	w	w	PROPN
fcis-25702	82	37	,	,	PUNCT
fcis-25702	82	38	w	w	PROPN
fcis-25702	82	39	,	,	PUNCT
fcis-25702	82	40	…	…	PUNCT
fcis-25702	82	41	,	,	PUNCT
fcis-25702	82	42	w	w	NOUN
fcis-25702	82	43	,	,	PUNCT
fcis-25702	82	44	w	w	PROPN
fcis-25702	82	45	.	.	PUNCT
fcis-25702	83	1	the	the	DET
fcis-25702	83	2	above	above	ADJ
fcis-25702	83	3	process	process	NOUN
fcis-25702	83	4	will	will	AUX
fcis-25702	83	5	continue	continue	VERB
fcis-25702	83	6	until	until	SCONJ
fcis-25702	83	7	the	the	DET
fcis-25702	83	8	end	end	NOUN
fcis-25702	83	9	-	-	PUNCT
fcis-25702	83	10	of	of	ADP
fcis-25702	83	11	-	-	PUNCT
fcis-25702	83	12	sentence	sentence	NOUN
fcis-25702	83	13	token	token	NOUN
fcis-25702	83	14	is	be	AUX
fcis-25702	83	15	met	meet	VERB
fcis-25702	83	16	.	.	PUNCT
fcis-25702	84	1	note	note	VERB
fcis-25702	84	2	that	that	SCONJ
fcis-25702	84	3	in	in	ADP
fcis-25702	84	4	practice	practice	NOUN
fcis-25702	84	5	,	,	PUNCT
fcis-25702	84	6	one	one	PRON
fcis-25702	84	7	will	will	AUX
fcis-25702	84	8	not	not	PART
fcis-25702	84	9	let	let	VERB
fcis-25702	84	10	the	the	DET
fcis-25702	84	11	new	new	ADJ
fcis-25702	84	12	sequence	sequence	NOUN
fcis-25702	84	13	w	w	PROPN
fcis-25702	84	14	,	,	PUNCT
fcis-25702	84	15	w	w	PROPN
fcis-25702	84	16	,	,	PUNCT
fcis-25702	84	17	w	w	PROPN
fcis-25702	84	18	,	,	PUNCT
fcis-25702	84	19	…	…	PUNCT
fcis-25702	84	20	,	,	PUNCT
fcis-25702	84	21	w	w	X
fcis-25702	84	22	,	,	PUNCT
fcis-25702	84	23	w	w	PROPN
fcis-25702	84	24	go	go	VERB
fcis-25702	84	25	through	through	ADP
fcis-25702	84	26	all	all	DET
fcis-25702	84	27	the	the	DET
fcis-25702	84	28	transformers	transformer	NOUN
fcis-25702	84	29	repeatedly	repeatedly	ADV
fcis-25702	84	30	since	since	SCONJ
fcis-25702	84	31	this	this	PRON
fcis-25702	84	32	requires	require	VERB
fcis-25702	84	33	too	too	ADV
fcis-25702	84	34	much	much	ADJ
fcis-25702	84	35	computation	computation	NOUN
fcis-25702	84	36	.	.	PUNCT
fcis-25702	85	1	for	for	ADP
fcis-25702	85	2	efficient	efficient	ADJ
fcis-25702	85	3	token	token	VERB
fcis-25702	85	4	generation	generation	NOUN
fcis-25702	85	5	,	,	PUNCT
fcis-25702	85	6	one	one	PRON
fcis-25702	85	7	will	will	AUX
fcis-25702	85	8	adopt	adopt	VERB
fcis-25702	85	9	the	the	DET
fcis-25702	85	10	key	key	ADJ
fcis-25702	85	11	-	-	PUNCT
fcis-25702	85	12	value	value	NOUN
fcis-25702	85	13	(	(	PUNCT
fcis-25702	85	14	kv	kv	NOUN
fcis-25702	85	15	)	)	PUNCT
fcis-25702	85	16	caching	cache	VERB
fcis-25702	85	17	mechanism	mechanism	NOUN
fcis-25702	85	18	,	,	PUNCT
fcis-25702	85	19	which	which	PRON
fcis-25702	85	20	is	be	AUX
fcis-25702	85	21	an	an	DET
fcis-25702	85	22	efficient	efficient	ADJ
fcis-25702	85	23	implementation	implementation	NOUN
fcis-25702	85	24	for	for	ADP
fcis-25702	85	25	token	token	ADJ
fcis-25702	85	26	generation	generation	NOUN
fcis-25702	85	27	in	in	ADP
fcis-25702	85	28	transformer	transformer	NOUN
fcis-25702	85	29	models	model	NOUN
fcis-25702	85	30	.	.	PUNCT
fcis-25702	86	1	with	with	ADP
fcis-25702	86	2	the	the	DET
fcis-25702	86	3	causal	causal	PROPN
fcis-25702	86	4	mask	mask	NOUN
fcis-25702	86	5	,	,	PUNCT
fcis-25702	86	6	the	the	DET
fcis-25702	86	7	newly	newly	ADV
fcis-25702	86	8	generated	generate	VERB
fcis-25702	86	9	token	token	VERB
fcis-25702	86	10	will	will	AUX
fcis-25702	86	11	not	not	PART
fcis-25702	86	12	affect	affect	VERB
fcis-25702	86	13	the	the	DET
fcis-25702	86	14	hidden	hide	VERB
fcis-25702	86	15	states	state	NOUN
fcis-25702	86	16	of	of	ADP
fcis-25702	86	17	the	the	DET
fcis-25702	86	18	previous	previous	ADJ
fcis-25702	86	19	shorter	short	ADJ
fcis-25702	86	20	sequences	sequence	NOUN
fcis-25702	86	21	.	.	PUNCT
fcis-25702	87	1	thus	thus	ADV
fcis-25702	87	2	,	,	PUNCT
fcis-25702	87	3	we	we	PRON
fcis-25702	87	4	only	only	ADV
fcis-25702	87	5	need	need	VERB
fcis-25702	87	6	to	to	PART
fcis-25702	87	7	calculate	calculate	VERB
fcis-25702	87	8	the	the	DET
fcis-25702	87	9	key	key	NOUN
fcis-25702	87	10	and	and	CCONJ
fcis-25702	87	11	value	value	NOUN
fcis-25702	87	12	of	of	ADP
fcis-25702	87	13	the	the	DET
fcis-25702	87	14	new	new	ADJ
fcis-25702	87	15	token	token	NOUN
fcis-25702	87	16	.	.	PUNCT
fcis-25702	88	1	the	the	DET
fcis-25702	88	2	model	model	NOUN
fcis-25702	88	3	can	can	AUX
fcis-25702	88	4	significantly	significantly	ADV
fcis-25702	88	5	reduce	reduce	VERB
fcis-25702	88	6	redundant	redundant	ADJ
fcis-25702	88	7	computations	computation	NOUN
fcis-25702	88	8	during	during	ADP
fcis-25702	88	9	subsequent	subsequent	ADJ
fcis-25702	88	10	steps	step	NOUN
fcis-25702	88	11	by	by	ADP
fcis-25702	88	12	storing	store	VERB
fcis-25702	88	13	the	the	DET
fcis-25702	88	14	computed	computed	ADJ
fcis-25702	88	15	keys	key	NOUN
fcis-25702	88	16	and	and	CCONJ
fcis-25702	88	17	values	value	NOUN
fcis-25702	88	18	for	for	ADP
fcis-25702	88	19	previously	previously	ADV
fcis-25702	88	20	processed	process	VERB
fcis-25702	88	21	tokens	token	NOUN
fcis-25702	88	22	.	.	PUNCT
fcis-25702	89	1	however	however	ADV
fcis-25702	89	2	,	,	PUNCT
fcis-25702	89	3	in	in	ADP
fcis-25702	89	4	scenarios	scenario	NOUN
fcis-25702	89	5	of	of	ADP
fcis-25702	89	6	early	early	ADJ
fcis-25702	89	7	exiting	exiting	NOUN
fcis-25702	89	8	,	,	PUNCT
fcis-25702	89	9	for	for	ADP
fcis-25702	89	10	example	example	NOUN
fcis-25702	89	11	,	,	PUNCT
fcis-25702	89	12	if	if	SCONJ
fcis-25702	89	13	the	the	DET
fcis-25702	89	14	next	next	ADJ
fcis-25702	89	15	token	token	NOUN
fcis-25702	89	16	exits	exit	VERB
fcis-25702	89	17	later	later	ADV
fcis-25702	89	18	than	than	ADP
fcis-25702	89	19	the	the	DET
fcis-25702	89	20	previous	previous	ADJ
fcis-25702	89	21	one	one	NUM
fcis-25702	89	22	,	,	PUNCT
fcis-25702	89	23	we	we	PRON
fcis-25702	89	24	need	need	VERB
fcis-25702	89	25	to	to	PART
fcis-25702	89	26	recompute	recompute	VERB
fcis-25702	89	27	the	the	DET
fcis-25702	89	28	kv	kv	PROPN
fcis-25702	89	29	values	value	NOUN
fcis-25702	89	30	for	for	ADP
fcis-25702	89	31	the	the	DET
fcis-25702	89	32	previous	previous	ADJ
fcis-25702	89	33	tokens	token	NOUN
fcis-25702	89	34	.	.	PUNCT
fcis-25702	90	1	skipdecode	skipdecode	PROPN
fcis-25702	91	1	[	[	X
fcis-25702	91	2	33	33	NUM
fcis-25702	91	3	]	]	PUNCT
fcis-25702	91	4	then	then	ADV
fcis-25702	91	5	proposes	propose	VERB
fcis-25702	91	6	to	to	PART
fcis-25702	91	7	monotonically	monotonically	ADV
fcis-25702	91	8	decrease	decrease	VERB
fcis-25702	91	9	exit	exit	NOUN
fcis-25702	91	10	points	point	NOUN
fcis-25702	91	11	to	to	PART
fcis-25702	91	12	eliminate	eliminate	VERB
fcis-25702	91	13	the	the	DET
fcis-25702	91	14	necessity	necessity	NOUN
fcis-25702	91	15	to	to	PART
fcis-25702	91	16	recalculate	recalculate	VERB
fcis-25702	91	17	kv	kv	PROPN
fcis-25702	91	18	caches	cache	NOUN
fcis-25702	91	19	for	for	ADP
fcis-25702	91	20	preceding	precede	VERB
fcis-25702	91	21	tokens	token	NOUN
fcis-25702	91	22	.	.	PUNCT
fcis-25702	92	1	3	3	X
fcis-25702	92	2	.	.	X
fcis-25702	92	3	methodology	methodology	NOUN
fcis-25702	92	4	3.1	3.1	NUM
fcis-25702	92	5	.	.	PUNCT
fcis-25702	92	6	dynamic	dynamic	ADJ
fcis-25702	92	7	layer	layer	NOUN
fcis-25702	92	8	skipping	skip	VERB
fcis-25702	92	9	architecture	architecture	NOUN
fcis-25702	92	10	and	and	CCONJ
fcis-25702	92	11	pretraining	pretraine	VERB
fcis-25702	92	12	skip	skip	NOUN
fcis-25702	92	13	-	-	PUNCT
fcis-25702	92	14	decode	decode	NOUN
fcis-25702	92	15	[	[	X
fcis-25702	92	16	33	33	NUM
fcis-25702	92	17	]	]	PUNCT
fcis-25702	92	18	directly	directly	ADV
fcis-25702	92	19	skips	skip	VERB
fcis-25702	92	20	the	the	DET
fcis-25702	92	21	layers	layer	NOUN
fcis-25702	92	22	of	of	ADP
fcis-25702	92	23	llms	llm	NOUN
fcis-25702	92	24	without	without	ADP
fcis-25702	92	25	training	training	NOUN
fcis-25702	92	26	,	,	PUNCT
fcis-25702	92	27	lacking	lack	VERB
fcis-25702	92	28	investigation	investigation	NOUN
fcis-25702	92	29	into	into	ADP
fcis-25702	92	30	training	training	NOUN
fcis-25702	92	31	methods	method	NOUN
fcis-25702	92	32	that	that	PRON
fcis-25702	92	33	enhance	enhance	VERB
fcis-25702	92	34	the	the	DET
fcis-25702	92	35	performance	performance	NOUN
fcis-25702	92	36	of	of	ADP
fcis-25702	92	37	layer	layer	NOUN
fcis-25702	92	38	skipping	skip	VERB
fcis-25702	92	39	.	.	PUNCT
fcis-25702	93	1	yuan	yuan	NOUN
fcis-25702	93	2	et	et	PROPN
fcis-25702	93	3	al	al	PROPN
fcis-25702	93	4	.	.	PUNCT
fcis-25702	94	1	[	[	X
fcis-25702	94	2	34	34	NUM
fcis-25702	94	3	]	]	SYM
fcis-25702	94	4	3	3	NUM
fcis-25702	94	5	directly	directly	ADV
fcis-25702	94	6	cuts	cut	VERB
fcis-25702	94	7	off	off	ADP
fcis-25702	94	8	some	some	DET
fcis-25702	94	9	layers	layer	NOUN
fcis-25702	94	10	and	and	CCONJ
fcis-25702	94	11	then	then	ADV
fcis-25702	94	12	fine	fine	ADJ
fcis-25702	94	13	-	-	PUNCT
fcis-25702	94	14	tunes	tune	NOUN
fcis-25702	94	15	the	the	DET
fcis-25702	94	16	model	model	NOUN
fcis-25702	94	17	.	.	PUNCT
fcis-25702	95	1	both	both	PRON
fcis-25702	95	2	of	of	ADP
fcis-25702	95	3	these	these	DET
fcis-25702	95	4	methods	method	NOUN
fcis-25702	95	5	are	be	AUX
fcis-25702	95	6	static	static	ADJ
fcis-25702	95	7	layer	layer	NOUN
fcis-25702	95	8	skipping	skip	VERB
fcis-25702	95	9	.	.	PUNCT
fcis-25702	96	1	in	in	ADP
fcis-25702	96	2	the	the	DET
fcis-25702	96	3	following	following	ADJ
fcis-25702	96	4	part	part	NOUN
fcis-25702	96	5	,	,	PUNCT
fcis-25702	96	6	we	we	PRON
fcis-25702	96	7	will	will	AUX
fcis-25702	96	8	introduce	introduce	VERB
fcis-25702	96	9	the	the	DET
fcis-25702	96	10	architecture	architecture	NOUN
fcis-25702	96	11	we	we	PRON
fcis-25702	96	12	use	use	VERB
fcis-25702	96	13	for	for	ADP
fcis-25702	96	14	dynamic	dynamic	ADJ
fcis-25702	96	15	layer	layer	NOUN
fcis-25702	96	16	skipping	skip	VERB
fcis-25702	96	17	and	and	CCONJ
fcis-25702	96	18	the	the	DET
fcis-25702	96	19	method	method	NOUN
fcis-25702	96	20	we	we	PRON
fcis-25702	96	21	used	use	VERB
fcis-25702	96	22	to	to	PART
fcis-25702	96	23	pretrain	pretrain	VERB
fcis-25702	96	24	the	the	DET
fcis-25702	96	25	model	model	NOUN
fcis-25702	96	26	.	.	PUNCT
fcis-25702	97	1	the	the	DET
fcis-25702	97	2	pretraining	pretraine	VERB
fcis-25702	97	3	is	be	AUX
fcis-25702	97	4	designed	design	VERB
fcis-25702	97	5	to	to	PART
fcis-25702	97	6	enhance	enhance	VERB
fcis-25702	97	7	llms	llm	NOUN
fcis-25702	97	8	’	'	PUNCT
fcis-25702	97	9	ability	ability	NOUN
fcis-25702	97	10	for	for	ADP
fcis-25702	97	11	dynamic	dynamic	ADJ
fcis-25702	97	12	layer	layer	NOUN
fcis-25702	97	13	skipping	skip	VERB
fcis-25702	97	14	and	and	CCONJ
fcis-25702	97	15	to	to	PART
fcis-25702	97	16	initialize	initialize	VERB
fcis-25702	97	17	the	the	DET
fcis-25702	97	18	parameters	parameter	NOUN
fcis-25702	97	19	of	of	ADP
fcis-25702	97	20	the	the	DET
fcis-25702	97	21	controllers	controller	NOUN
fcis-25702	97	22	.	.	PUNCT
fcis-25702	98	1	figure	figure	VERB
fcis-25702	98	2	1	1	NUM
fcis-25702	98	3	.	.	PUNCT
fcis-25702	99	1	the	the	DET
fcis-25702	99	2	adapter	adapter	NOUN
fcis-25702	99	3	enhanced	enhance	VERB
fcis-25702	99	4	transformer	transformer	NOUN
fcis-25702	99	5	layers	layer	NOUN
fcis-25702	99	6	.	.	PUNCT
fcis-25702	100	1	we	we	PRON
fcis-25702	100	2	add	add	VERB
fcis-25702	100	3	an	an	DET
fcis-25702	100	4	adapter	adapter	NOUN
fcis-25702	100	5	layer	layer	NOUN
fcis-25702	100	6	at	at	ADP
fcis-25702	100	7	each	each	DET
fcis-25702	100	8	transformer	transformer	NOUN
fcis-25702	100	9	layer	layer	NOUN
fcis-25702	100	10	,	,	PUNCT
fcis-25702	100	11	if	if	SCONJ
fcis-25702	100	12	current	current	ADJ
fcis-25702	100	13	transformer	transformer	NOUN
fcis-25702	100	14	layer	layer	NOUN
fcis-25702	100	15	is	be	AUX
fcis-25702	100	16	skipped	skip	VERB
fcis-25702	100	17	,	,	PUNCT
fcis-25702	100	18	then	then	ADV
fcis-25702	100	19	the	the	DET
fcis-25702	100	20	hidden	hidden	ADJ
fcis-25702	100	21	state	state	NOUN
fcis-25702	100	22	goes	go	VERB
fcis-25702	100	23	through	through	ADP
fcis-25702	100	24	the	the	DET
fcis-25702	100	25	adapter	adapter	NOUN
fcis-25702	100	26	layer	layer	NOUN
fcis-25702	100	27	dynamic	dynamic	ADJ
fcis-25702	100	28	layer	layer	NOUN
fcis-25702	100	29	skipping	skip	VERB
fcis-25702	100	30	architecture	architecture	NOUN
fcis-25702	100	31	:	:	PUNCT
fcis-25702	100	32	we	we	PRON
fcis-25702	100	33	insert	insert	VERB
fcis-25702	100	34	an	an	DET
fcis-25702	100	35	adapter	adapter	NOUN
fcis-25702	100	36	layer	layer	NOUN
fcis-25702	100	37	in	in	ADP
fcis-25702	100	38	parallel	parallel	NOUN
fcis-25702	100	39	with	with	ADP
fcis-25702	100	40	each	each	DET
fcis-25702	100	41	transformer	transformer	NOUN
fcis-25702	100	42	layer	layer	NOUN
fcis-25702	100	43	,	,	PUNCT
fcis-25702	100	44	just	just	ADV
fcis-25702	100	45	like	like	ADP
fcis-25702	100	46	lora	lora	PROPN
fcis-25702	101	1	[	[	X
fcis-25702	101	2	37	37	NUM
fcis-25702	101	3	]	]	PUNCT
fcis-25702	101	4	,	,	PUNCT
fcis-25702	101	5	as	as	SCONJ
fcis-25702	101	6	shown	show	VERB
fcis-25702	101	7	in	in	ADP
fcis-25702	101	8	figure	figure	NOUN
fcis-25702	101	9	1	1	NUM
fcis-25702	101	10	.	.	PUNCT
fcis-25702	102	1	this	this	DET
fcis-25702	102	2	adapter	adapter	NOUN
fcis-25702	102	3	consists	consist	VERB
fcis-25702	102	4	of	of	ADP
fcis-25702	102	5	a	a	DET
fcis-25702	102	6	twolayer	twolayer	NOUN
fcis-25702	102	7	mlp	mlp	NOUN
fcis-25702	102	8	with	with	ADP
fcis-25702	102	9	a	a	DET
fcis-25702	102	10	gelu	gelu	ADJ
fcis-25702	102	11	activation	activation	NOUN
fcis-25702	102	12	and	and	CCONJ
fcis-25702	102	13	takes	take	VERB
fcis-25702	102	14	the	the	DET
fcis-25702	102	15	hidden	hide	VERB
fcis-25702	102	16	states	state	NOUN
fcis-25702	102	17	h	h	NOUN
fcis-25702	102	18	of	of	ADP
fcis-25702	102	19	w	w	PROPN
fcis-25702	102	20	as	as	ADP
fcis-25702	102	21	the	the	DET
fcis-25702	102	22	input	input	NOUN
fcis-25702	102	23	feature	feature	NOUN
fcis-25702	102	24	and	and	CCONJ
fcis-25702	102	25	outputs	output	VERB
fcis-25702	102	26	the	the	DET
fcis-25702	102	27	hidden	hide	VERB
fcis-25702	102	28	states	state	NOUN
fcis-25702	102	29	h	h	NOUN
fcis-25702	102	30	.	.	PUNCT
fcis-25702	103	1	since	since	SCONJ
fcis-25702	103	2	the	the	DET
fcis-25702	103	3	adapters	adapter	NOUN
fcis-25702	103	4	use	use	VERB
fcis-25702	103	5	a	a	DET
fcis-25702	103	6	bottleneck	bottleneck	NOUN
fcis-25702	103	7	architecture	architecture	NOUN
fcis-25702	103	8	[	[	X
fcis-25702	103	9	38	38	NUM
fcis-25702	103	10	]	]	PUNCT
fcis-25702	103	11	,	,	PUNCT
fcis-25702	103	12	they	they	PRON
fcis-25702	103	13	are	be	AUX
fcis-25702	103	14	computationally	computationally	ADV
fcis-25702	103	15	efficient	efficient	ADJ
fcis-25702	103	16	.	.	PUNCT
fcis-25702	104	1	for	for	ADP
fcis-25702	104	2	a	a	DET
fcis-25702	104	3	transformer	transformer	NOUN
fcis-25702	104	4	layer	layer	NOUN
fcis-25702	104	5	i	i	PRON
fcis-25702	104	6	,	,	PUNCT
fcis-25702	104	7	we	we	PRON
fcis-25702	104	8	input	input	VERB
fcis-25702	104	9	the	the	DET
fcis-25702	104	10	hidden	hide	VERB
fcis-25702	104	11	states	state	NOUN
fcis-25702	104	12	of	of	ADP
fcis-25702	104	13	the	the	DET
fcis-25702	104	14	last	last	ADJ
fcis-25702	104	15	layer	layer	NOUN
fcis-25702	104	16	h	h	NOUN
fcis-25702	104	17	into	into	ADP
fcis-25702	104	18	the	the	DET
fcis-25702	104	19	transformer	transformer	NOUN
fcis-25702	104	20	layer	layer	NOUN
fcis-25702	104	21	and	and	CCONJ
fcis-25702	104	22	the	the	DET
fcis-25702	104	23	adapter	adapter	NOUN
fcis-25702	104	24	layer	layer	NOUN
fcis-25702	104	25	,	,	PUNCT
fcis-25702	104	26	and	and	CCONJ
fcis-25702	104	27	we	we	PRON
fcis-25702	104	28	sum	sum	VERB
fcis-25702	104	29	them	they	PRON
fcis-25702	104	30	to	to	PART
fcis-25702	104	31	get	get	VERB
fcis-25702	104	32	the	the	DET
fcis-25702	104	33	output	output	NOUN
fcis-25702	104	34	hidden	hide	VERB
fcis-25702	104	35	states	state	NOUN
fcis-25702	104	36	of	of	ADP
fcis-25702	104	37	layer	layer	NOUN
fcis-25702	105	1	i	i	PRON
fcis-25702	105	2	:	:	PUNCT
fcis-25702	105	3	h	h	PROPN
fcis-25702	105	4	adapter	adapter	PROPN
fcis-25702	105	5	h	h	PROPN
fcis-25702	105	6	trans	trans	PROPN
fcis-25702	105	7	h	h	PROPN
fcis-25702	105	8	(	(	PUNCT
fcis-25702	105	9	5	5	X
fcis-25702	105	10	)	)	PUNCT
fcis-25702	105	11	we	we	PRON
fcis-25702	105	12	use	use	VERB
fcis-25702	105	13	the	the	DET
fcis-25702	105	14	adapter	adapter	NOUN
fcis-25702	105	15	layer	layer	NOUN
fcis-25702	105	16	adapter	adapter	NOUN
fcis-25702	105	17	as	as	SCONJ
fcis-25702	105	18	the	the	DET
fcis-25702	105	19	controller	controller	NOUN
fcis-25702	105	20	to	to	PART
fcis-25702	105	21	decide	decide	VERB
fcis-25702	105	22	whether	whether	SCONJ
fcis-25702	105	23	to	to	PART
fcis-25702	105	24	skip	skip	VERB
fcis-25702	105	25	transformer	transformer	NOUN
fcis-25702	105	26	layer	layer	NOUN
fcis-25702	105	27	i	i	PRON
fcis-25702	105	28	.	.	PUNCT
fcis-25702	106	1	by	by	ADP
fcis-25702	106	2	taking	take	VERB
fcis-25702	106	3	the	the	DET
fcis-25702	106	4	output	output	NOUN
fcis-25702	106	5	of	of	ADP
fcis-25702	106	6	the	the	DET
fcis-25702	106	7	adapter	adapter	NOUN
fcis-25702	106	8	layers	layer	NOUN
fcis-25702	106	9	as	as	ADP
fcis-25702	106	10	input	input	NOUN
fcis-25702	106	11	for	for	ADP
fcis-25702	106	12	the	the	DET
fcis-25702	106	13	next	next	ADJ
fcis-25702	106	14	layer	layer	NOUN
fcis-25702	106	15	,	,	PUNCT
fcis-25702	106	16	the	the	DET
fcis-25702	106	17	adapters	adapter	NOUN
fcis-25702	106	18	learn	learn	VERB
fcis-25702	106	19	the	the	DET
fcis-25702	106	20	essential	essential	ADJ
fcis-25702	106	21	information	information	NOUN
fcis-25702	106	22	needed	need	VERB
fcis-25702	106	23	to	to	PART
fcis-25702	106	24	pass	pass	VERB
fcis-25702	106	25	to	to	ADP
fcis-25702	106	26	the	the	DET
fcis-25702	106	27	next	next	ADJ
fcis-25702	106	28	layer	layer	NOUN
fcis-25702	106	29	.	.	PUNCT
fcis-25702	107	1	this	this	DET
fcis-25702	107	2	information	information	NOUN
fcis-25702	107	3	can	can	AUX
fcis-25702	107	4	greatly	greatly	ADV
fcis-25702	107	5	help	help	VERB
fcis-25702	107	6	the	the	DET
fcis-25702	107	7	adapter	adapter	NOUN
fcis-25702	107	8	assess	assess	VERB
fcis-25702	107	9	whether	whether	SCONJ
fcis-25702	107	10	to	to	PART
fcis-25702	107	11	skip	skip	VERB
fcis-25702	107	12	the	the	DET
fcis-25702	107	13	next	next	ADJ
fcis-25702	107	14	layer	layer	NOUN
fcis-25702	107	15	.	.	PUNCT
fcis-25702	108	1	by	by	ADP
fcis-25702	108	2	pretraining	pretraine	VERB
fcis-25702	108	3	the	the	DET
fcis-25702	108	4	adapters	adapter	NOUN
fcis-25702	108	5	,	,	PUNCT
fcis-25702	108	6	it	it	PRON
fcis-25702	108	7	provides	provide	VERB
fcis-25702	108	8	a	a	DET
fcis-25702	108	9	good	good	ADJ
fcis-25702	108	10	initialization	initialization	NOUN
fcis-25702	108	11	of	of	ADP
fcis-25702	108	12	the	the	DET
fcis-25702	108	13	adapter	adapter	NOUN
fcis-25702	108	14	parameters	parameter	NOUN
fcis-25702	108	15	compared	compare	VERB
fcis-25702	108	16	to	to	ADP
fcis-25702	108	17	basic	basic	ADJ
fcis-25702	108	18	initialization	initialization	NOUN
fcis-25702	108	19	methods	method	NOUN
fcis-25702	108	20	such	such	ADJ
fcis-25702	108	21	as	as	ADP
fcis-25702	108	22	xavier	xavier	PROPN
fcis-25702	108	23	initialization	initialization	PROPN
fcis-25702	108	24	.	.	PUNCT
fcis-25702	109	1	a	a	DET
fcis-25702	109	2	good	good	ADJ
fcis-25702	109	3	initialization	initialization	NOUN
fcis-25702	109	4	offers	offer	VERB
fcis-25702	109	5	a	a	DET
fcis-25702	109	6	smaller	small	ADJ
fcis-25702	109	7	search	search	NOUN
fcis-25702	109	8	space	space	NOUN
fcis-25702	109	9	for	for	ADP
fcis-25702	109	10	reinforcement	reinforcement	NOUN
fcis-25702	109	11	learning	learning	NOUN
fcis-25702	109	12	,	,	PUNCT
fcis-25702	109	13	and	and	CCONJ
fcis-25702	109	14	thus	thus	ADV
fcis-25702	109	15	helps	help	VERB
fcis-25702	109	16	rl	rl	PART
fcis-25702	109	17	converge	converge	VERB
fcis-25702	109	18	to	to	ADP
fcis-25702	109	19	the	the	DET
fcis-25702	109	20	optimal	optimal	ADJ
fcis-25702	109	21	policy	policy	NOUN
fcis-25702	109	22	.	.	PUNCT
fcis-25702	110	1	layer	layer	NOUN
fcis-25702	110	2	skipping	skip	VERB
fcis-25702	110	3	pretraining	pretraine	VERB
fcis-25702	110	4	:	:	PUNCT
fcis-25702	110	5	to	to	PART
fcis-25702	110	6	train	train	VERB
fcis-25702	110	7	a	a	DET
fcis-25702	110	8	model	model	NOUN
fcis-25702	110	9	with	with	ADP
fcis-25702	110	10	layer	layer	NOUN
fcis-25702	110	11	skipping	skip	VERB
fcis-25702	110	12	capabilities	capability	NOUN
fcis-25702	110	13	,	,	PUNCT
fcis-25702	110	14	we	we	PRON
fcis-25702	110	15	assign	assign	VERB
fcis-25702	110	16	a	a	DET
fcis-25702	110	17	random	random	ADJ
fcis-25702	110	18	0,1	0,1	NUM
fcis-25702	110	19	mask	mask	NOUN
fcis-25702	110	20	m	m	VERB
fcis-25702	110	21	(	(	PUNCT
fcis-25702	110	22	i.e.	i.e.	X
fcis-25702	110	23	,	,	PUNCT
fcis-25702	110	24	a	a	DET
fcis-25702	110	25	bernoulli	bernoulli	NOUN
fcis-25702	110	26	random	random	ADJ
fcis-25702	110	27	variable	variable	NOUN
fcis-25702	110	28	)	)	PUNCT
fcis-25702	110	29	with	with	ADP
fcis-25702	110	30	a	a	DET
fcis-25702	110	31	probability	probability	NOUN
fcis-25702	110	32	of	of	ADP
fcis-25702	110	33	being	be	AUX
fcis-25702	110	34	0	0	NUM
fcis-25702	110	35	set	set	VERB
fcis-25702	110	36	to	to	ADP
fcis-25702	110	37	p	p	NOUN
fcis-25702	110	38	0	0	PUNCT
fcis-25702	111	1	p	p	NOUN
fcis-25702	111	2	1	1	NUM
fcis-25702	111	3	)	)	PUNCT
fcis-25702	111	4	for	for	ADP
fcis-25702	111	5	each	each	DET
fcis-25702	111	6	layer	layer	NOUN
fcis-25702	111	7	i.	i.	NOUN
fcis-25702	111	8	thus	thus	ADV
fcis-25702	111	9	,	,	PUNCT
fcis-25702	111	10	the	the	DET
fcis-25702	111	11	above	above	ADJ
fcis-25702	111	12	equation	equation	NOUN
fcis-25702	111	13	becomes	become	VERB
fcis-25702	111	14	:	:	PUNCT
fcis-25702	111	15	h	h	PROPN
fcis-25702	111	16	adapter	adapter	PROPN
fcis-25702	111	17	h	h	PROPN
fcis-25702	111	18	m	m	PROPN
fcis-25702	111	19	∗	∗	PROPN
fcis-25702	111	20	trans	trans	PROPN
fcis-25702	111	21	h	h	PROPN
fcis-25702	111	22	(	(	PUNCT
fcis-25702	111	23	6	6	NUM
fcis-25702	111	24	)	)	PUNCT
fcis-25702	111	25	that	that	PRON
fcis-25702	111	26	is	be	AUX
fcis-25702	111	27	,	,	PUNCT
fcis-25702	111	28	if	if	SCONJ
fcis-25702	111	29	layer	layer	NOUN
fcis-25702	111	30	i	i	PRON
fcis-25702	111	31	is	be	AUX
fcis-25702	111	32	skipped	skip	VERB
fcis-25702	111	33	,	,	PUNCT
fcis-25702	111	34	the	the	DET
fcis-25702	111	35	hidden	hide	VERB
fcis-25702	111	36	states	state	NOUN
fcis-25702	111	37	will	will	AUX
fcis-25702	111	38	only	only	ADV
fcis-25702	111	39	go	go	VERB
fcis-25702	111	40	through	through	ADP
fcis-25702	111	41	adapter	adapter	NOUN
fcis-25702	111	42	.	.	PUNCT
fcis-25702	112	1	during	during	ADP
fcis-25702	112	2	training	training	NOUN
fcis-25702	112	3	,	,	PUNCT
fcis-25702	112	4	the	the	DET
fcis-25702	112	5	parameters	parameter	NOUN
fcis-25702	112	6	of	of	ADP
fcis-25702	112	7	transformer	transformer	NOUN
fcis-25702	112	8	layer	layer	NOUN
fcis-25702	112	9	trans	tran	NOUN
fcis-25702	112	10	can	can	AUX
fcis-25702	112	11	be	be	AUX
fcis-25702	112	12	either	either	ADV
fcis-25702	112	13	fully	fully	ADV
fcis-25702	112	14	tuned	tune	VERB
fcis-25702	112	15	or	or	CCONJ
fcis-25702	112	16	tuned	tune	VERB
fcis-25702	112	17	with	with	ADP
fcis-25702	112	18	lora	lora	PROPN
fcis-25702	113	1	[	[	X
fcis-25702	113	2	37	37	NUM
fcis-25702	113	3	]	]	PUNCT
fcis-25702	113	4	to	to	PART
fcis-25702	113	5	save	save	VERB
fcis-25702	113	6	computational	computational	ADJ
fcis-25702	113	7	resources	resource	NOUN
fcis-25702	113	8	.	.	PUNCT
fcis-25702	114	1	note	note	VERB
fcis-25702	114	2	that	that	SCONJ
fcis-25702	114	3	separate	separate	ADJ
fcis-25702	114	4	forward	forward	ADJ
fcis-25702	114	5	steps	step	NOUN
fcis-25702	114	6	may	may	AUX
fcis-25702	114	7	result	result	VERB
fcis-25702	114	8	in	in	ADP
fcis-25702	114	9	different	different	ADJ
fcis-25702	114	10	representations	representation	NOUN
fcis-25702	114	11	with	with	ADP
fcis-25702	114	12	random	random	ADJ
fcis-25702	114	13	layer	layer	NOUN
fcis-25702	114	14	masks	mask	NOUN
fcis-25702	114	15	.	.	PUNCT
fcis-25702	115	1	we	we	PRON
fcis-25702	115	2	can	can	AUX
fcis-25702	115	3	add	add	VERB
fcis-25702	115	4	a	a	DET
fcis-25702	115	5	consistency	consistency	NOUN
fcis-25702	115	6	regularization	regularization	NOUN
fcis-25702	115	7	objective	objective	NOUN
fcis-25702	115	8	to	to	ADP
fcis-25702	115	9	the	the	DET
fcis-25702	115	10	ce	ce	PROPN
fcis-25702	115	11	loss	loss	NOUN
fcis-25702	115	12	term	term	NOUN
fcis-25702	115	13	to	to	PART
fcis-25702	115	14	ensure	ensure	VERB
fcis-25702	115	15	model	model	NOUN
fcis-25702	115	16	consistency	consistency	NOUN
fcis-25702	115	17	after	after	ADP
fcis-25702	115	18	random	random	ADJ
fcis-25702	115	19	layer	layer	NOUN
fcis-25702	115	20	skipping	skip	VERB
fcis-25702	115	21	.	.	PUNCT
fcis-25702	116	1	for	for	ADP
fcis-25702	116	2	the	the	DET
fcis-25702	116	3	separate	separate	ADJ
fcis-25702	116	4	forward	forward	ADJ
fcis-25702	116	5	passes	pass	NOUN
fcis-25702	116	6	of	of	ADP
fcis-25702	116	7	the	the	DET
fcis-25702	116	8	same	same	ADJ
fcis-25702	116	9	input	input	NOUN
fcis-25702	116	10	w	w	PROPN
fcis-25702	116	11	,	,	PUNCT
fcis-25702	116	12	w	w	PROPN
fcis-25702	116	13	,	,	PUNCT
fcis-25702	116	14	w	w	PROPN
fcis-25702	116	15	,	,	PUNCT
fcis-25702	116	16	…	…	PUNCT
fcis-25702	116	17	,	,	PUNCT
fcis-25702	116	18	w	w	X
fcis-25702	116	19	,	,	PUNCT
fcis-25702	116	20	one	one	PRON
fcis-25702	116	21	can	can	AUX
fcis-25702	116	22	obtain	obtain	VERB
fcis-25702	116	23	three	three	NUM
fcis-25702	116	24	different	different	ADJ
fcis-25702	116	25	distributions	distribution	NOUN
fcis-25702	116	26	p	p	NOUN
fcis-25702	116	27	,	,	PUNCT
fcis-25702	116	28	p	p	X
fcis-25702	116	29	,	,	PUNCT
fcis-25702	116	30	and	and	CCONJ
fcis-25702	116	31	p	p	X
fcis-25702	116	32	for	for	ADP
fcis-25702	116	33	w	w	PROPN
fcis-25702	116	34	.	.	PUNCT
fcis-25702	117	1	p	p	NOUN
fcis-25702	117	2	denotes	denote	VERB
fcis-25702	117	3	the	the	DET
fcis-25702	117	4	teacher	teacher	NOUN
fcis-25702	117	5	model	model	NOUN
fcis-25702	117	6	that	that	PRON
fcis-25702	117	7	does	do	AUX
fcis-25702	117	8	not	not	PART
fcis-25702	117	9	skip	skip	VERB
fcis-25702	117	10	,	,	PUNCT
fcis-25702	117	11	while	while	SCONJ
fcis-25702	117	12	p	p	NOUN
fcis-25702	117	13	and	and	CCONJ
fcis-25702	117	14	p	p	NOUN
fcis-25702	117	15	are	be	AUX
fcis-25702	117	16	student	student	NOUN
fcis-25702	117	17	models	model	NOUN
fcis-25702	117	18	that	that	PRON
fcis-25702	117	19	randomly	randomly	ADV
fcis-25702	117	20	skip	skip	VERB
fcis-25702	117	21	different	different	ADJ
fcis-25702	117	22	layers	layer	NOUN
fcis-25702	117	23	.	.	PUNCT
fcis-25702	118	1	we	we	PRON
fcis-25702	118	2	use	use	VERB
fcis-25702	118	3	kldivergence	kldivergence	NOUN
fcis-25702	118	4	to	to	PART
fcis-25702	118	5	provide	provide	VERB
fcis-25702	118	6	consistency	consistency	NOUN
fcis-25702	118	7	regularization	regularization	NOUN
fcis-25702	118	8	:	:	PUNCT
fcis-25702	118	9	reg	reg	NUM
fcis-25702	118	10	kl	kl	NOUN
fcis-25702	118	11	p	p	NOUN
fcis-25702	118	12	,	,	PUNCT
fcis-25702	118	13	p	p	X
fcis-25702	118	14	(	(	PUNCT
fcis-25702	118	15	7	7	NUM
fcis-25702	118	16	)	)	PUNCT
fcis-25702	118	17	,	,	PUNCT
fcis-25702	118	18	and	and	CCONJ
fcis-25702	118	19	distill	distill	VERB
fcis-25702	118	20	knowledge	knowledge	NOUN
fcis-25702	118	21	from	from	ADP
fcis-25702	118	22	teacher	teacher	NOUN
fcis-25702	118	23	model	model	NOUN
fcis-25702	118	24	:	:	PUNCT
fcis-25702	118	25	distill	distill	VERB
fcis-25702	118	26	kl	kl	PROPN
fcis-25702	118	27	p	p	NOUN
fcis-25702	118	28	,	,	PUNCT
fcis-25702	118	29	p	p	X
fcis-25702	118	30	kl	kl	NOUN
fcis-25702	118	31	p	p	NOUN
fcis-25702	118	32	,	,	PUNCT
fcis-25702	118	33	p	p	X
fcis-25702	118	34	(	(	PUNCT
fcis-25702	118	35	8)	8)	NUM
fcis-25702	118	36	the	the	DET
fcis-25702	118	37	final	final	ADJ
fcis-25702	118	38	loss	loss	NOUN
fcis-25702	118	39	for	for	ADP
fcis-25702	118	40	pretraining	pretraine	VERB
fcis-25702	118	41	is	be	AUX
fcis-25702	118	42	:	:	PUNCT
fcis-25702	118	43	pretrain	pretrain	NOUN
fcis-25702	118	44	α	α	PROPN
fcis-25702	118	45	∗	∗	NOUN
fcis-25702	118	46	distill	distill	VERB
fcis-25702	118	47	reg	reg	PRON
fcis-25702	118	48	cls	cls	NOUN
fcis-25702	118	49	(	(	PUNCT
fcis-25702	118	50	9	9	NUM
fcis-25702	118	51	)	)	PUNCT
fcis-25702	118	52	,	,	PUNCT
fcis-25702	118	53	where	where	SCONJ
fcis-25702	118	54	α	α	NOUN
fcis-25702	118	55	is	be	AUX
fcis-25702	118	56	the	the	DET
fcis-25702	118	57	hyper	hyper	NOUN
fcis-25702	118	58	-	-	NOUN
fcis-25702	118	59	parameter	parameter	NOUN
fcis-25702	118	60	controlling	control	VERB
fcis-25702	118	61	the	the	DET
fcis-25702	118	62	weight	weight	NOUN
fcis-25702	118	63	of	of	ADP
fcis-25702	118	64	loss	loss	NOUN
fcis-25702	118	65	.	.	PUNCT
fcis-25702	119	1	3.2	3.2	NUM
fcis-25702	119	2	.	.	PUNCT
fcis-25702	119	3	reinforcement	reinforcement	NOUN
fcis-25702	119	4	learning	learning	NOUN
fcis-25702	119	5	framework	framework	NOUN
fcis-25702	119	6	we	we	PRON
fcis-25702	119	7	now	now	ADV
fcis-25702	119	8	propose	propose	VERB
fcis-25702	119	9	a	a	DET
fcis-25702	119	10	reinforcement	reinforcement	NOUN
fcis-25702	119	11	learning	learning	NOUN
fcis-25702	119	12	(	(	PUNCT
fcis-25702	119	13	rl	rl	NOUN
fcis-25702	119	14	)	)	PUNCT
fcis-25702	119	15	based	base	VERB
fcis-25702	119	16	approach	approach	NOUN
fcis-25702	119	17	for	for	ADP
fcis-25702	119	18	dynamically	dynamically	ADV
fcis-25702	119	19	determining	determine	VERB
fcis-25702	119	20	which	which	DET
fcis-25702	119	21	layers	layer	NOUN
fcis-25702	119	22	to	to	PART
fcis-25702	119	23	skip	skip	VERB
fcis-25702	119	24	when	when	SCONJ
fcis-25702	119	25	predicting	predict	VERB
fcis-25702	119	26	the	the	DET
fcis-25702	119	27	next	next	ADJ
fcis-25702	119	28	token	token	NOUN
fcis-25702	119	29	w	w	NOUN
fcis-25702	119	30	given	give	VERB
fcis-25702	119	31	w	w	PROPN
fcis-25702	119	32	,	,	PUNCT
fcis-25702	119	33	…	…	PUNCT
fcis-25702	119	34	,	,	PUNCT
fcis-25702	119	35	w	w	X
fcis-25702	119	36	.	.	PUNCT
fcis-25702	120	1	at	at	ADP
fcis-25702	120	2	each	each	DET
fcis-25702	120	3	transformer	transformer	NOUN
fcis-25702	120	4	layer	layer	NOUN
fcis-25702	121	1	i	i	PRON
fcis-25702	121	2	,	,	PUNCT
fcis-25702	121	3	we	we	PRON
fcis-25702	121	4	use	use	VERB
fcis-25702	121	5	the	the	DET
fcis-25702	121	6	adapter	adapter	NOUN
fcis-25702	121	7	layer	layer	NOUN
fcis-25702	121	8	to	to	PART
fcis-25702	121	9	get	get	VERB
fcis-25702	121	10	the	the	DET
fcis-25702	121	11	hidden	hide	VERB
fcis-25702	121	12	states	state	NOUN
fcis-25702	121	13	h	h	NOUN
fcis-25702	121	14	,	,	PUNCT
fcis-25702	121	15	then	then	ADV
fcis-25702	121	16	we	we	PRON
fcis-25702	121	17	add	add	VERB
fcis-25702	121	18	an	an	DET
fcis-25702	121	19	rl	rl	NOUN
fcis-25702	121	20	-	-	PUNCT
fcis-25702	121	21	specific	specific	ADJ
fcis-25702	121	22	head	head	NOUN
fcis-25702	121	23	,	,	PUNCT
fcis-25702	121	24	which	which	PRON
fcis-25702	121	25	is	be	AUX
fcis-25702	121	26	a	a	DET
fcis-25702	121	27	one	one	NUM
fcis-25702	121	28	-	-	PUNCT
fcis-25702	121	29	layer	layer	NOUN
fcis-25702	121	30	mlp	mlp	NOUN
fcis-25702	121	31	with	with	ADP
fcis-25702	121	32	sigmoid	sigmoid	NOUN
fcis-25702	121	33	activation	activation	NOUN
fcis-25702	121	34	,	,	PUNCT
fcis-25702	121	35	to	to	ADP
fcis-25702	121	36	the	the	DET
fcis-25702	121	37	adapter	adapter	NOUN
fcis-25702	121	38	.	.	PUNCT
fcis-25702	122	1	thus	thus	ADV
fcis-25702	122	2	,	,	PUNCT
fcis-25702	122	3	the	the	DET
fcis-25702	122	4	controller	controller	NOUN
fcis-25702	122	5	consists	consist	VERB
fcis-25702	122	6	of	of	ADP
fcis-25702	122	7	an	an	DET
fcis-25702	122	8	adapter	adapter	NOUN
fcis-25702	122	9	layer	layer	NOUN
fcis-25702	122	10	and	and	CCONJ
fcis-25702	122	11	an	an	DET
fcis-25702	122	12	rlspecific	rlspecific	ADJ
fcis-25702	122	13	head	head	NOUN
fcis-25702	122	14	.	.	PUNCT
fcis-25702	123	1	we	we	PRON
fcis-25702	123	2	use	use	VERB
fcis-25702	123	3	the	the	DET
fcis-25702	123	4	output	output	NOUN
fcis-25702	123	5	of	of	ADP
fcis-25702	123	6	the	the	DET
fcis-25702	123	7	controller	controller	NOUN
fcis-25702	123	8	as	as	ADP
fcis-25702	123	9	the	the	DET
fcis-25702	123	10	predicted	predict	VERB
fcis-25702	123	11	probability	probability	NOUN
fcis-25702	123	12	of	of	ADP
fcis-25702	123	13	skipping	skip	VERB
fcis-25702	123	14	,	,	PUNCT
fcis-25702	123	15	pskip	pskip	NOUN
fcis-25702	123	16	.	.	PUNCT
fcis-25702	124	1	if	if	SCONJ
fcis-25702	124	2	pskip	pskip	NOUN
fcis-25702	124	3	exceeds	exceed	VERB
fcis-25702	124	4	a	a	DET
fcis-25702	124	5	threshold	threshold	NOUN
fcis-25702	124	6	,	,	PUNCT
fcis-25702	124	7	we	we	PRON
fcis-25702	124	8	will	will	AUX
fcis-25702	124	9	skip	skip	VERB
fcis-25702	124	10	the	the	DET
fcis-25702	124	11	next	next	ADJ
fcis-25702	124	12	layer	layer	NOUN
fcis-25702	124	13	.	.	PUNCT
fcis-25702	125	1	note	note	VERB
fcis-25702	125	2	that	that	SCONJ
fcis-25702	125	3	we	we	PRON
fcis-25702	125	4	do	do	AUX
fcis-25702	125	5	not	not	PART
fcis-25702	125	6	have	have	VERB
fcis-25702	125	7	any	any	DET
fcis-25702	125	8	ground	ground	NOUN
fcis-25702	125	9	truth	truth	NOUN
fcis-25702	125	10	labels	label	NOUN
fcis-25702	125	11	for	for	ADP
fcis-25702	125	12	each	each	DET
fcis-25702	125	13	layer	layer	NOUN
fcis-25702	125	14	’s	’s	PART
fcis-25702	125	15	controller	controller	NOUN
fcis-25702	125	16	.	.	PUNCT
fcis-25702	126	1	thus	thus	ADV
fcis-25702	126	2	,	,	PUNCT
fcis-25702	126	3	the	the	DET
fcis-25702	126	4	controllers	controller	NOUN
fcis-25702	126	5	are	be	AUX
fcis-25702	126	6	optimized	optimize	VERB
fcis-25702	126	7	via	via	ADP
fcis-25702	126	8	reinforcement	reinforcement	NOUN
fcis-25702	126	9	learning	learning	NOUN
fcis-25702	126	10	algorithms	algorithm	NOUN
fcis-25702	126	11	like	like	INTJ
fcis-25702	126	12	reinforce	reinforce	VERB
fcis-25702	126	13	[	[	X
fcis-25702	126	14	39	39	NUM
fcis-25702	126	15	]	]	PUNCT
fcis-25702	126	16	.	.	PUNCT
fcis-25702	127	1	to	to	PART
fcis-25702	127	2	be	be	AUX
fcis-25702	127	3	more	more	ADV
fcis-25702	127	4	specific	specific	ADJ
fcis-25702	127	5	,	,	PUNCT
fcis-25702	127	6	we	we	PRON
fcis-25702	127	7	define	define	VERB
fcis-25702	127	8	the	the	DET
fcis-25702	127	9	state	state	NOUN
fcis-25702	127	10	,	,	PUNCT
fcis-25702	127	11	action	action	NOUN
fcis-25702	127	12	,	,	PUNCT
fcis-25702	127	13	reward	reward	NOUN
fcis-25702	127	14	,	,	PUNCT
fcis-25702	127	15	advantage	advantage	NOUN
fcis-25702	127	16	function	function	NOUN
fcis-25702	127	17	,	,	PUNCT
fcis-25702	127	18	and	and	CCONJ
fcis-25702	127	19	objective	objective	ADJ
fcis-25702	127	20	function	function	NOUN
fcis-25702	127	21	below	below	ADV
fcis-25702	127	22	.	.	PUNCT
fcis-25702	128	1	figure	figure	NOUN
fcis-25702	128	2	2	2	NUM
fcis-25702	128	3	.	.	PUNCT
fcis-25702	128	4	dynamic	dynamic	ADJ
fcis-25702	128	5	layer	layer	NOUN
fcis-25702	128	6	skipping	skip	VERB
fcis-25702	128	7	framework	framework	NOUN
fcis-25702	128	8	:	:	PUNCT
fcis-25702	128	9	at	at	ADP
fcis-25702	128	10	each	each	DET
fcis-25702	128	11	transformer	transformer	NOUN
fcis-25702	128	12	layer	layer	NOUN
fcis-25702	128	13	we	we	PRON
fcis-25702	128	14	have	have	VERB
fcis-25702	128	15	a	a	DET
fcis-25702	128	16	controller	controller	NOUN
fcis-25702	128	17	,	,	PUNCT
fcis-25702	128	18	which	which	PRON
fcis-25702	128	19	consists	consist	VERB
fcis-25702	128	20	of	of	ADP
fcis-25702	128	21	an	an	DET
fcis-25702	128	22	adapter	adapter	NOUN
fcis-25702	128	23	layer	layer	NOUN
fcis-25702	128	24	and	and	CCONJ
fcis-25702	128	25	a	a	DET
fcis-25702	128	26	rl	rl	NOUN
fcis-25702	128	27	head	head	NOUN
fcis-25702	128	28	.	.	PUNCT
fcis-25702	129	1	we	we	PRON
fcis-25702	129	2	use	use	VERB
fcis-25702	129	3	the	the	DET
fcis-25702	129	4	output	output	NOUN
fcis-25702	129	5	of	of	ADP
fcis-25702	129	6	the	the	DET
fcis-25702	129	7	controller	controller	NOUN
fcis-25702	129	8	to	to	PART
fcis-25702	129	9	determine	determine	VERB
fcis-25702	129	10	whether	whether	SCONJ
fcis-25702	129	11	to	to	PART
fcis-25702	129	12	skip	skip	VERB
fcis-25702	129	13	computing	compute	VERB
fcis-25702	129	14	the	the	DET
fcis-25702	129	15	next	next	ADJ
fcis-25702	129	16	transformer	transformer	NOUN
fcis-25702	129	17	layer	layer	NOUN
fcis-25702	129	18	4	4	NUM
fcis-25702	129	19	state	state	NOUN
fcis-25702	129	20	state	state	NOUN
fcis-25702	129	21	s	s	PART
fcis-25702	129	22	consists	consist	NOUN
fcis-25702	129	23	of	of	ADP
fcis-25702	129	24	the	the	DET
fcis-25702	129	25	sequence	sequence	NOUN
fcis-25702	129	26	representations	representation	VERB
fcis-25702	129	27	from	from	ADP
fcis-25702	129	28	transformer	transformer	ADJ
fcis-25702	129	29	layer	layer	NOUN
fcis-25702	129	30	l.	l.	NOUN
fcis-25702	129	31	action	action	NOUN
fcis-25702	129	32	the	the	DET
fcis-25702	129	33	actions	action	NOUN
fcis-25702	129	34	at	at	ADP
fcis-25702	129	35	each	each	DET
fcis-25702	129	36	layer	layer	NOUN
fcis-25702	129	37	are	be	AUX
fcis-25702	129	38	skip	skip	ADJ
fcis-25702	129	39	,	,	PUNCT
fcis-25702	129	40	select	select	ADJ
fcis-25702	129	41	.	.	PUNCT
fcis-25702	130	1	skip	skip	ADJ
fcis-25702	130	2	means	mean	NOUN
fcis-25702	130	3	to	to	PART
fcis-25702	130	4	skip	skip	VERB
fcis-25702	130	5	computing	compute	VERB
fcis-25702	130	6	the	the	DET
fcis-25702	130	7	hidden	hide	VERB
fcis-25702	130	8	states	state	NOUN
fcis-25702	130	9	of	of	ADP
fcis-25702	130	10	the	the	DET
fcis-25702	130	11	next	next	ADJ
fcis-25702	130	12	layer	layer	NOUN
fcis-25702	130	13	,	,	PUNCT
fcis-25702	130	14	while	while	SCONJ
fcis-25702	130	15	select	select	ADJ
fcis-25702	130	16	means	mean	VERB
fcis-25702	130	17	the	the	DET
fcis-25702	130	18	hidden	hide	VERB
fcis-25702	130	19	states	state	NOUN
fcis-25702	130	20	go	go	VERB
fcis-25702	130	21	through	through	ADP
fcis-25702	130	22	the	the	DET
fcis-25702	130	23	next	next	ADJ
fcis-25702	130	24	transformer	transformer	NOUN
fcis-25702	130	25	layer	layer	NOUN
fcis-25702	130	26	to	to	PART
fcis-25702	130	27	compute	compute	VERB
fcis-25702	130	28	.	.	PUNCT
fcis-25702	131	1	the	the	DET
fcis-25702	131	2	controllers	controller	NOUN
fcis-25702	131	3	we	we	PRON
fcis-25702	131	4	use	use	VERB
fcis-25702	131	5	,	,	PUNCT
fcis-25702	131	6	which	which	PRON
fcis-25702	131	7	are	be	AUX
fcis-25702	131	8	called	call	VERB
fcis-25702	131	9	policy	policy	NOUN
fcis-25702	131	10	networks	network	NOUN
fcis-25702	131	11	in	in	ADP
fcis-25702	131	12	rl	rl	NOUN
fcis-25702	131	13	,	,	PUNCT
fcis-25702	131	14	generate	generate	VERB
fcis-25702	131	15	probabilities	probability	NOUN
fcis-25702	131	16	for	for	ADP
fcis-25702	131	17	actions	action	NOUN
fcis-25702	131	18	and	and	CCONJ
fcis-25702	131	19	contain	contain	VERB
fcis-25702	131	20	an	an	DET
fcis-25702	131	21	adapter	adapter	NOUN
fcis-25702	131	22	layer	layer	NOUN
fcis-25702	131	23	and	and	CCONJ
fcis-25702	131	24	an	an	DET
fcis-25702	131	25	rl	rl	NOUN
fcis-25702	131	26	-	-	PUNCT
fcis-25702	131	27	specific	specific	ADJ
fcis-25702	131	28	head	head	NOUN
fcis-25702	131	29	.	.	PUNCT
fcis-25702	132	1	more	more	ADV
fcis-25702	132	2	formally	formally	ADV
fcis-25702	132	3	,	,	PUNCT
fcis-25702	132	4	the	the	DET
fcis-25702	132	5	policy	policy	NOUN
fcis-25702	132	6	network	network	NOUN
fcis-25702	132	7	is	be	AUX
fcis-25702	132	8	defined	define	VERB
fcis-25702	132	9	as	as	ADP
fcis-25702	132	10	:	:	PUNCT
fcis-25702	132	11	∣	∣	PROPN
fcis-25702	132	12	σ	σ	PROPN
fcis-25702	132	13	w	w	PROPN
fcis-25702	132	14	gelu	gelu	PROPN
fcis-25702	132	15	w	w	PROPN
fcis-25702	132	16	gelu	gelu	PROPN
fcis-25702	132	17	w	w	PROPN
fcis-25702	132	18	hlast	hlast	PROPN
fcis-25702	132	19	b	b	PROPN
fcis-25702	132	20	b	b	PROPN
fcis-25702	132	21	(	(	PUNCT
fcis-25702	132	22	10	10	NUM
fcis-25702	132	23	)	)	PUNCT
fcis-25702	132	24	where	where	SCONJ
fcis-25702	132	25	σ	σ	PROPN
fcis-25702	132	26	is	be	AUX
fcis-25702	132	27	the	the	DET
fcis-25702	132	28	sigmoid	sigmoid	NOUN
fcis-25702	132	29	function	function	NOUN
fcis-25702	132	30	,	,	PUNCT
fcis-25702	132	31	a	a	DET
fcis-25702	132	32	denotes	denote	NOUN
fcis-25702	132	33	the	the	DET
fcis-25702	132	34	action	action	NOUN
fcis-25702	132	35	at	at	ADP
fcis-25702	132	36	state	state	NOUN
fcis-25702	132	37	s	s	PART
fcis-25702	132	38	,	,	PUNCT
fcis-25702	132	39	l	l	NOUN
fcis-25702	132	40	is	be	AUX
fcis-25702	132	41	the	the	DET
fcis-25702	132	42	current	current	ADJ
fcis-25702	132	43	layer	layer	NOUN
fcis-25702	132	44	,	,	PUNCT
fcis-25702	132	45	h	h	PROPN
fcis-25702	132	46	last	last	ADJ
fcis-25702	132	47	denotes	denote	VERB
fcis-25702	132	48	the	the	DET
fcis-25702	132	49	last	last	ADJ
fcis-25702	132	50	token	token	VERB
fcis-25702	132	51	representation	representation	NOUN
fcis-25702	132	52	at	at	ADP
fcis-25702	132	53	layer	layer	NOUN
fcis-25702	132	54	l	l	NOUN
fcis-25702	132	55	,	,	PUNCT
fcis-25702	132	56	and	and	CCONJ
fcis-25702	132	57	θ	θ	PROPN
fcis-25702	132	58	are	be	AUX
fcis-25702	132	59	the	the	DET
fcis-25702	132	60	parameters	parameter	NOUN
fcis-25702	132	61	of	of	ADP
fcis-25702	132	62	the	the	DET
fcis-25702	132	63	policy	policy	NOUN
fcis-25702	132	64	network	network	NOUN
fcis-25702	132	65	.	.	PUNCT
fcis-25702	133	1	we	we	PRON
fcis-25702	133	2	will	will	AUX
fcis-25702	133	3	use	use	VERB
fcis-25702	133	4	a	a	DET
fcis-25702	133	5	a	a	NOUN
fcis-25702	133	6	,	,	PUNCT
fcis-25702	133	7	a	a	PRON
fcis-25702	133	8	,	,	PUNCT
fcis-25702	133	9	…	…	PUNCT
fcis-25702	133	10	,	,	PUNCT
fcis-25702	133	11	a	a	DET
fcis-25702	133	12	to	to	PART
fcis-25702	133	13	denote	denote	VERB
fcis-25702	133	14	the	the	DET
fcis-25702	133	15	action	action	NOUN
fcis-25702	133	16	trajectory	trajectory	NOUN
fcis-25702	133	17	the	the	DET
fcis-25702	133	18	policy	policy	NOUN
fcis-25702	133	19	takes	take	VERB
fcis-25702	133	20	up	up	ADP
fcis-25702	133	21	to	to	ADP
fcis-25702	133	22	layer	layer	NOUN
fcis-25702	133	23	l.	l.	PROPN
fcis-25702	133	24	reward	reward	PROPN
fcis-25702	133	25	denote	denote	VERB
fcis-25702	133	26	the	the	DET
fcis-25702	133	27	ground	ground	NOUN
fcis-25702	133	28	truth	truth	NOUN
fcis-25702	133	29	of	of	ADP
fcis-25702	133	30	token	token	ADJ
fcis-25702	133	31	w	w	PROPN
fcis-25702	133	32	as	as	ADP
fcis-25702	133	33	w∗	w∗	NOUN
fcis-25702	133	34	(	(	PUNCT
fcis-25702	133	35	provided	provide	VERB
fcis-25702	133	36	in	in	ADP
fcis-25702	133	37	the	the	DET
fcis-25702	133	38	training	training	NOUN
fcis-25702	133	39	data	datum	NOUN
fcis-25702	133	40	)	)	PUNCT
fcis-25702	133	41	.	.	PUNCT
fcis-25702	134	1	at	at	ADP
fcis-25702	134	2	an	an	DET
fcis-25702	134	3	intermediate	intermediate	ADJ
fcis-25702	134	4	layer	layer	NOUN
fcis-25702	134	5	l	l	NOUN
fcis-25702	134	6	,	,	PUNCT
fcis-25702	134	7	if	if	SCONJ
fcis-25702	134	8	the	the	DET
fcis-25702	134	9	layer	layer	NOUN
fcis-25702	134	10	is	be	AUX
fcis-25702	134	11	not	not	PART
fcis-25702	134	12	skipped	skip	VERB
fcis-25702	134	13	,	,	PUNCT
fcis-25702	134	14	then	then	ADV
fcis-25702	134	15	the	the	DET
fcis-25702	134	16	reward	reward	NOUN
fcis-25702	134	17	is	be	AUX
fcis-25702	134	18	r	r	NOUN
fcis-25702	134	19	a	a	DET
fcis-25702	134	20	∣	∣	NOUN
fcis-25702	134	21	s	s	NOUN
fcis-25702	134	22	1	1	NUM
fcis-25702	134	23	.	.	PUNCT
fcis-25702	135	1	if	if	SCONJ
fcis-25702	135	2	an	an	DET
fcis-25702	135	3	intermediate	intermediate	ADJ
fcis-25702	135	4	layer	layer	NOUN
fcis-25702	135	5	l	l	NOUN
fcis-25702	135	6	is	be	AUX
fcis-25702	135	7	skipped	skip	VERB
fcis-25702	135	8	,	,	PUNCT
fcis-25702	135	9	the	the	DET
fcis-25702	135	10	reward	reward	NOUN
fcis-25702	135	11	is	be	AUX
fcis-25702	135	12	r	r	NOUN
fcis-25702	135	13	a	a	DET
fcis-25702	135	14	∣	∣	NOUN
fcis-25702	135	15	s	s	NOUN
fcis-25702	135	16	1	1	NUM
fcis-25702	135	17	.	.	PUNCT
fcis-25702	136	1	at	at	ADP
fcis-25702	136	2	the	the	DET
fcis-25702	136	3	final	final	ADJ
fcis-25702	136	4	layer	layer	NOUN
fcis-25702	136	5	,	,	PUNCT
fcis-25702	136	6	after	after	ADP
fcis-25702	136	7	obtaining	obtain	VERB
fcis-25702	136	8	the	the	DET
fcis-25702	136	9	probability	probability	NOUN
fcis-25702	136	10	distribution	distribution	NOUN
fcis-25702	136	11	from	from	ADP
fcis-25702	136	12	the	the	DET
fcis-25702	136	13	lm	lm	ADJ
fcis-25702	136	14	head	head	NOUN
fcis-25702	136	15	as	as	ADP
fcis-25702	136	16	p	p	PROPN
fcis-25702	136	17	w	w	PROPN
fcis-25702	136	18	∣	∣	PROPN
fcis-25702	136	19	h	h	NOUN
fcis-25702	136	20	,	,	PUNCT
fcis-25702	136	21	we	we	PRON
fcis-25702	136	22	will	will	AUX
fcis-25702	136	23	receive	receive	VERB
fcis-25702	136	24	the	the	DET
fcis-25702	136	25	final	final	ADJ
fcis-25702	136	26	reward	reward	NOUN
fcis-25702	136	27	ce	ce	PROPN
fcis-25702	136	28	w∗	w∗	PROPN
fcis-25702	136	29	,	,	PUNCT
fcis-25702	136	30	p	p	PROPN
fcis-25702	136	31	w	w	PROPN
fcis-25702	136	32	∣	∣	PROPN
fcis-25702	136	33	h	h	NOUN
fcis-25702	136	34	.	.	PUNCT
fcis-25702	137	1	the	the	DET
fcis-25702	137	2	cumulative	cumulative	ADJ
fcis-25702	137	3	intermediate	intermediate	ADJ
fcis-25702	137	4	reward	reward	NOUN
fcis-25702	137	5	is	be	AUX
fcis-25702	137	6	l	l	NOUN
fcis-25702	137	7	∑	∑	NOUN
fcis-25702	137	8	 	 	SPACE
fcis-25702	137	9	r	r	NOUN
fcis-25702	137	10	a	a	DET
fcis-25702	137	11	∣	∣	NOUN
fcis-25702	137	12	s	s	NOUN
fcis-25702	137	13	.	.	PUNCT
fcis-25702	138	1	thus	thus	ADV
fcis-25702	138	2	,	,	PUNCT
fcis-25702	138	3	the	the	DET
fcis-25702	138	4	total	total	ADJ
fcis-25702	138	5	reward	reward	NOUN
fcis-25702	138	6	for	for	ADP
fcis-25702	138	7	the	the	DET
fcis-25702	138	8	controller	controller	NOUN
fcis-25702	138	9	is	be	AUX
fcis-25702	138	10	:	:	PUNCT
fcis-25702	138	11	r	r	VERB
fcis-25702	138	12	a	a	DET
fcis-25702	138	13	∣	∣	ADJ
fcis-25702	138	14	θ	θ	PROPN
fcis-25702	138	15	ce	ce	NOUN
fcis-25702	138	16	w∗	w∗	PROPN
fcis-25702	138	17	,	,	PUNCT
fcis-25702	138	18	p	p	PROPN
fcis-25702	138	19	w	w	PROPN
fcis-25702	138	20	∣	∣	PROPN
fcis-25702	138	21	h	h	NOUN
fcis-25702	138	22	λ	λ	PROPN
fcis-25702	138	23	∗	∗	NOUN
fcis-25702	138	24	l	l	NOUN
fcis-25702	138	25	(	(	PUNCT
fcis-25702	138	26	11	11	NUM
fcis-25702	138	27	)	)	PUNCT
fcis-25702	138	28	where	where	SCONJ
fcis-25702	138	29	l	l	NOUN
fcis-25702	138	30	is	be	AUX
fcis-25702	138	31	the	the	DET
fcis-25702	138	32	number	number	NOUN
fcis-25702	138	33	of	of	ADP
fcis-25702	138	34	transformer	transformer	NOUN
fcis-25702	138	35	layers	layer	NOUN
fcis-25702	138	36	actually	actually	ADV
fcis-25702	138	37	processed	process	VERB
fcis-25702	138	38	,	,	PUNCT
fcis-25702	138	39	we	we	PRON
fcis-25702	138	40	encourage	encourage	VERB
fcis-25702	138	41	the	the	DET
fcis-25702	138	42	model	model	NOUN
fcis-25702	138	43	to	to	PART
fcis-25702	138	44	skip	skip	VERB
fcis-25702	138	45	as	as	ADP
fcis-25702	138	46	many	many	ADJ
fcis-25702	138	47	layers	layer	NOUN
fcis-25702	138	48	as	as	ADP
fcis-25702	138	49	possible	possible	ADJ
fcis-25702	138	50	.	.	PUNCT
fcis-25702	139	1	λ	λ	NOUN
fcis-25702	139	2	is	be	AUX
fcis-25702	139	3	a	a	DET
fcis-25702	139	4	hyperparameter	hyperparameter	NOUN
fcis-25702	139	5	for	for	ADP
fcis-25702	139	6	controlling	control	VERB
fcis-25702	139	7	the	the	DET
fcis-25702	139	8	penalty	penalty	NOUN
fcis-25702	139	9	of	of	ADP
fcis-25702	139	10	computational	computational	ADJ
fcis-25702	139	11	costs	cost	NOUN
fcis-25702	139	12	.	.	PUNCT
fcis-25702	140	1	ce	ce	PROPN
fcis-25702	140	2	)	)	PUNCT
fcis-25702	140	3	is	be	AUX
fcis-25702	140	4	the	the	DET
fcis-25702	140	5	cross	cross	ADJ
fcis-25702	140	6	-	-	ADJ
fcis-25702	140	7	entropy	entropy	ADJ
fcis-25702	140	8	loss	loss	NOUN
fcis-25702	140	9	,	,	PUNCT
fcis-25702	140	10	reflecting	reflect	VERB
fcis-25702	140	11	the	the	DET
fcis-25702	140	12	performance	performance	NOUN
fcis-25702	140	13	of	of	ADP
fcis-25702	140	14	layer	layer	NOUN
fcis-25702	140	15	skipping	skip	VERB
fcis-25702	140	16	.	.	PUNCT
fcis-25702	141	1	by	by	ADP
fcis-25702	141	2	maximizing	maximize	VERB
fcis-25702	141	3	the	the	DET
fcis-25702	141	4	reward	reward	NOUN
fcis-25702	141	5	,	,	PUNCT
fcis-25702	141	6	the	the	DET
fcis-25702	141	7	controllers	controller	NOUN
fcis-25702	141	8	learn	learn	VERB
fcis-25702	141	9	to	to	PART
fcis-25702	141	10	skip	skip	VERB
fcis-25702	141	11	layers	layer	NOUN
fcis-25702	141	12	while	while	SCONJ
fcis-25702	141	13	maintaining	maintain	VERB
fcis-25702	141	14	accurate	accurate	ADJ
fcis-25702	141	15	token	token	ADJ
fcis-25702	141	16	predictions	prediction	NOUN
fcis-25702	141	17	.	.	PUNCT
fcis-25702	142	1	it	it	PRON
fcis-25702	142	2	should	should	AUX
fcis-25702	142	3	be	be	AUX
fcis-25702	142	4	noted	note	VERB
fcis-25702	142	5	that	that	SCONJ
fcis-25702	142	6	the	the	DET
fcis-25702	142	7	computation	computation	NOUN
fcis-25702	142	8	of	of	ADP
fcis-25702	142	9	the	the	DET
fcis-25702	142	10	reward	reward	NOUN
fcis-25702	142	11	is	be	AUX
fcis-25702	142	12	based	base	VERB
fcis-25702	142	13	only	only	ADV
fcis-25702	142	14	on	on	ADP
fcis-25702	142	15	the	the	DET
fcis-25702	142	16	lm	lm	NOUN
fcis-25702	142	17	-	-	PUNCT
fcis-25702	142	18	logits	logit	NOUN
fcis-25702	142	19	of	of	ADP
fcis-25702	142	20	the	the	DET
fcis-25702	142	21	last	last	ADJ
fcis-25702	142	22	layer	layer	NOUN
fcis-25702	142	23	because	because	SCONJ
fcis-25702	142	24	if	if	SCONJ
fcis-25702	142	25	we	we	PRON
fcis-25702	142	26	were	be	AUX
fcis-25702	142	27	to	to	PART
fcis-25702	142	28	compute	compute	VERB
fcis-25702	142	29	the	the	DET
fcis-25702	142	30	reward	reward	NOUN
fcis-25702	142	31	for	for	ADP
fcis-25702	142	32	each	each	DET
fcis-25702	142	33	layer	layer	NOUN
fcis-25702	142	34	,	,	PUNCT
fcis-25702	142	35	we	we	PRON
fcis-25702	142	36	would	would	AUX
fcis-25702	142	37	need	need	VERB
fcis-25702	142	38	to	to	PART
fcis-25702	142	39	assign	assign	VERB
fcis-25702	142	40	an	an	DET
fcis-25702	142	41	individual	individual	ADJ
fcis-25702	142	42	lm	lm	INTJ
fcis-25702	142	43	head	head	NOUN
fcis-25702	142	44	at	at	ADP
fcis-25702	142	45	each	each	DET
fcis-25702	142	46	layer	layer	NOUN
fcis-25702	142	47	,	,	PUNCT
fcis-25702	142	48	which	which	PRON
fcis-25702	142	49	would	would	AUX
fcis-25702	142	50	require	require	VERB
fcis-25702	142	51	a	a	DET
fcis-25702	142	52	significant	significant	ADJ
fcis-25702	142	53	number	number	NOUN
fcis-25702	142	54	of	of	ADP
fcis-25702	142	55	computational	computational	ADJ
fcis-25702	142	56	resources	resource	NOUN
fcis-25702	142	57	.	.	PUNCT
fcis-25702	143	1	sampling	sample	VERB
fcis-25702	143	2	strategy	strategy	NOUN
fcis-25702	143	3	since	since	SCONJ
fcis-25702	143	4	we	we	PRON
fcis-25702	143	5	can	can	AUX
fcis-25702	143	6	not	not	PART
fcis-25702	143	7	directly	directly	ADV
fcis-25702	143	8	optimize	optimize	VERB
fcis-25702	143	9	the	the	DET
fcis-25702	143	10	reward	reward	NOUN
fcis-25702	143	11	of	of	ADP
fcis-25702	143	12	each	each	DET
fcis-25702	143	13	layer	layer	NOUN
fcis-25702	143	14	,	,	PUNCT
fcis-25702	143	15	in	in	ADP
fcis-25702	143	16	the	the	DET
fcis-25702	143	17	early	early	ADJ
fcis-25702	143	18	training	training	NOUN
fcis-25702	143	19	stages	stage	NOUN
fcis-25702	143	20	of	of	ADP
fcis-25702	143	21	rl	rl	NOUN
fcis-25702	143	22	,	,	PUNCT
fcis-25702	143	23	the	the	DET
fcis-25702	143	24	search	search	NOUN
fcis-25702	143	25	space	space	NOUN
fcis-25702	143	26	would	would	AUX
fcis-25702	143	27	be	be	AUX
fcis-25702	143	28	very	very	ADV
fcis-25702	143	29	large	large	ADJ
fcis-25702	143	30	,	,	PUNCT
fcis-25702	143	31	making	make	VERB
fcis-25702	143	32	it	it	PRON
fcis-25702	143	33	difficult	difficult	ADJ
fcis-25702	143	34	for	for	SCONJ
fcis-25702	143	35	the	the	DET
fcis-25702	143	36	model	model	NOUN
fcis-25702	143	37	to	to	PART
fcis-25702	143	38	converge	converge	VERB
fcis-25702	143	39	to	to	ADP
fcis-25702	143	40	the	the	DET
fcis-25702	143	41	optimal	optimal	ADJ
fcis-25702	143	42	policy	policy	NOUN
fcis-25702	143	43	.	.	PUNCT
fcis-25702	144	1	to	to	PART
fcis-25702	144	2	address	address	VERB
fcis-25702	144	3	this	this	DET
fcis-25702	144	4	problem	problem	NOUN
fcis-25702	144	5	,	,	PUNCT
fcis-25702	144	6	we	we	PRON
fcis-25702	144	7	use	use	VERB
fcis-25702	144	8	the	the	DET
fcis-25702	144	9	following	follow	VERB
fcis-25702	144	10	strategies	strategy	NOUN
fcis-25702	144	11	to	to	PART
fcis-25702	144	12	stabilize	stabilize	VERB
fcis-25702	144	13	the	the	DET
fcis-25702	144	14	reinforcement	reinforcement	NOUN
fcis-25702	144	15	learning	learning	NOUN
fcis-25702	144	16	:	:	PUNCT
fcis-25702	144	17	(	(	PUNCT
fcis-25702	144	18	1	1	X
fcis-25702	144	19	)	)	PUNCT
fcis-25702	144	20	we	we	PRON
fcis-25702	144	21	do	do	AUX
fcis-25702	144	22	not	not	PART
fcis-25702	144	23	drop	drop	VERB
fcis-25702	144	24	the	the	DET
fcis-25702	144	25	first	first	ADJ
fcis-25702	144	26	layer	layer	NOUN
fcis-25702	144	27	and	and	CCONJ
fcis-25702	144	28	the	the	DET
fcis-25702	144	29	last	last	ADJ
fcis-25702	144	30	layer	layer	NOUN
fcis-25702	144	31	,	,	PUNCT
fcis-25702	144	32	because	because	SCONJ
fcis-25702	144	33	it	it	PRON
fcis-25702	144	34	would	would	AUX
fcis-25702	144	35	cause	cause	VERB
fcis-25702	144	36	a	a	DET
fcis-25702	144	37	large	large	ADJ
fcis-25702	144	38	performance	performance	NOUN
fcis-25702	144	39	drop	drop	NOUN
fcis-25702	144	40	.	.	PUNCT
fcis-25702	145	1	(	(	PUNCT
fcis-25702	145	2	2	2	X
fcis-25702	145	3	)	)	PUNCT
fcis-25702	145	4	we	we	PRON
fcis-25702	145	5	calculate	calculate	VERB
fcis-25702	145	6	the	the	DET
fcis-25702	145	7	cosine	cosine	NOUN
fcis-25702	145	8	similarity	similarity	NOUN
fcis-25702	145	9	between	between	ADP
fcis-25702	145	10	two	two	NUM
fcis-25702	145	11	consecutive	consecutive	ADJ
fcis-25702	145	12	layers	layer	NOUN
fcis-25702	145	13	.	.	PUNCT
fcis-25702	146	1	zhang	zhang	PROPN
fcis-25702	146	2	et	et	PROPN
fcis-25702	146	3	al	al	PROPN
fcis-25702	146	4	.	.	PUNCT
fcis-25702	147	1	[	[	X
fcis-25702	147	2	26	26	NUM
fcis-25702	147	3	]	]	PUNCT
fcis-25702	147	4	found	find	VERB
fcis-25702	147	5	that	that	SCONJ
fcis-25702	147	6	in	in	ADP
fcis-25702	147	7	the	the	DET
fcis-25702	147	8	bert	bert	PROPN
fcis-25702	147	9	model	model	NOUN
fcis-25702	147	10	,	,	PUNCT
fcis-25702	147	11	if	if	SCONJ
fcis-25702	147	12	the	the	DET
fcis-25702	147	13	predictions	prediction	NOUN
fcis-25702	147	14	of	of	ADP
fcis-25702	147	15	several	several	ADJ
fcis-25702	147	16	consecutive	consecutive	ADJ
fcis-25702	147	17	layers	layer	NOUN
fcis-25702	147	18	have	have	VERB
fcis-25702	147	19	high	high	ADJ
fcis-25702	147	20	similarity	similarity	NOUN
fcis-25702	147	21	,	,	PUNCT
fcis-25702	147	22	then	then	ADV
fcis-25702	147	23	redundancy	redundancy	NOUN
fcis-25702	147	24	may	may	AUX
fcis-25702	147	25	exist	exist	VERB
fcis-25702	147	26	in	in	ADP
fcis-25702	147	27	those	those	DET
fcis-25702	147	28	layers	layer	NOUN
fcis-25702	147	29	,	,	PUNCT
fcis-25702	147	30	allowing	allow	VERB
fcis-25702	147	31	for	for	ADP
fcis-25702	147	32	early	early	ADJ
fcis-25702	147	33	exiting	exiting	NOUN
fcis-25702	147	34	.	.	PUNCT
fcis-25702	148	1	based	base	VERB
fcis-25702	148	2	on	on	ADP
fcis-25702	148	3	this	this	PRON
fcis-25702	148	4	,	,	PUNCT
fcis-25702	148	5	we	we	PRON
fcis-25702	148	6	skip	skip	VERB
fcis-25702	148	7	those	those	DET
fcis-25702	148	8	layers	layer	NOUN
fcis-25702	148	9	if	if	SCONJ
fcis-25702	148	10	their	their	PRON
fcis-25702	148	11	previous	previous	ADJ
fcis-25702	148	12	layer	layer	NOUN
fcis-25702	148	13	is	be	AUX
fcis-25702	148	14	similar	similar	ADJ
fcis-25702	148	15	to	to	ADP
fcis-25702	148	16	the	the	DET
fcis-25702	148	17	earlier	early	ADJ
fcis-25702	148	18	layers	layer	NOUN
fcis-25702	148	19	.	.	PUNCT
fcis-25702	149	1	in	in	ADP
fcis-25702	149	2	our	our	PRON
fcis-25702	149	3	setting	setting	NOUN
fcis-25702	149	4	,	,	PUNCT
fcis-25702	149	5	the	the	DET
fcis-25702	149	6	similarity	similarity	NOUN
fcis-25702	149	7	between	between	ADP
fcis-25702	149	8	layer	layer	NOUN
fcis-25702	149	9	i	i	PRON
fcis-25702	149	10	and	and	CCONJ
fcis-25702	149	11	layer	layer	NOUN
fcis-25702	149	12	i	i	PRON
fcis-25702	149	13	1	1	NUM
fcis-25702	149	14	is	be	AUX
fcis-25702	149	15	calculated	calculate	VERB
fcis-25702	149	16	by	by	ADP
fcis-25702	149	17	the	the	DET
fcis-25702	149	18	cosine	cosine	NOUN
fcis-25702	149	19	similarity	similarity	NOUN
fcis-25702	149	20	of	of	ADP
fcis-25702	149	21	the	the	DET
fcis-25702	149	22	last	last	ADJ
fcis-25702	149	23	token	token	ADJ
fcis-25702	149	24	hidden	hide	VERB
fcis-25702	149	25	states	state	NOUN
fcis-25702	149	26	:	:	PUNCT
fcis-25702	149	27	cos	cos	PROPN
fcis-25702	149	28	cos	cos	PROPN
fcis-25702	149	29	hlast	hlast	PROPN
fcis-25702	149	30	,	,	PUNCT
fcis-25702	149	31	hlast	hlast	NOUN
fcis-25702	149	32	.	.	PUNCT
fcis-25702	150	1	if	if	SCONJ
fcis-25702	150	2	cos	cos	PROPN
fcis-25702	150	3	is	be	AUX
fcis-25702	150	4	higher	high	ADJ
fcis-25702	150	5	than	than	ADP
fcis-25702	150	6	a	a	DET
fcis-25702	150	7	threshold	threshold	NOUN
fcis-25702	150	8	τ	τ	NOUN
fcis-25702	150	9	,	,	PUNCT
fcis-25702	150	10	we	we	PRON
fcis-25702	150	11	say	say	VERB
fcis-25702	150	12	layer	layer	NOUN
fcis-25702	150	13	i	i	PRON
fcis-25702	150	14	is	be	AUX
fcis-25702	150	15	s	s	PROPN
fcis-25702	150	16	p	p	NOUN
fcis-25702	150	17	,	,	PUNCT
fcis-25702	150	18	meaning	mean	VERB
fcis-25702	150	19	it	it	PRON
fcis-25702	150	20	’s	’	VERB
fcis-25702	150	21	similar	similar	ADJ
fcis-25702	150	22	to	to	ADP
fcis-25702	150	23	its	its	PRON
fcis-25702	150	24	previous	previous	ADJ
fcis-25702	150	25	layer	layer	NOUN
fcis-25702	150	26	.	.	PUNCT
fcis-25702	151	1	for	for	ADP
fcis-25702	151	2	a	a	DET
fcis-25702	151	3	layer	layer	NOUN
fcis-25702	152	1	i	i	PRON
fcis-25702	152	2	,	,	PUNCT
fcis-25702	152	3	if	if	SCONJ
fcis-25702	152	4	its	its	PRON
fcis-25702	152	5	m	m	VERB
fcis-25702	152	6	previous	previous	ADJ
fcis-25702	152	7	layers	layer	NOUN
fcis-25702	152	8	are	be	AUX
fcis-25702	152	9	all	all	PRON
fcis-25702	152	10	s	s	VERB
fcis-25702	152	11	p	p	X
fcis-25702	152	12	,	,	PUNCT
fcis-25702	152	13	it	it	PRON
fcis-25702	152	14	indicates	indicate	VERB
fcis-25702	152	15	that	that	SCONJ
fcis-25702	152	16	the	the	DET
fcis-25702	152	17	model	model	NOUN
fcis-25702	152	18	’s	’s	PART
fcis-25702	152	19	hidden	hide	VERB
fcis-25702	152	20	states	state	NOUN
fcis-25702	152	21	have	have	AUX
fcis-25702	152	22	not	not	PART
fcis-25702	152	23	changed	change	VERB
fcis-25702	152	24	much	much	ADV
fcis-25702	152	25	for	for	ADP
fcis-25702	152	26	m	m	PROPN
fcis-25702	152	27	consecutive	consecutive	ADJ
fcis-25702	152	28	layers	layer	NOUN
fcis-25702	152	29	.	.	PUNCT
fcis-25702	153	1	then	then	ADV
fcis-25702	153	2	,	,	PUNCT
fcis-25702	153	3	for	for	ADP
fcis-25702	153	4	the	the	DET
fcis-25702	153	5	skipping	skip	VERB
fcis-25702	153	6	prediction	prediction	NOUN
fcis-25702	153	7	probability	probability	NOUN
fcis-25702	153	8	p	p	NOUN
fcis-25702	153	9	skip	skip	NOUN
fcis-25702	153	10	for	for	ADP
fcis-25702	153	11	the	the	DET
fcis-25702	153	12	next	next	ADJ
fcis-25702	153	13	layer	layer	NOUN
fcis-25702	153	14	i	i	PRON
fcis-25702	153	15	1	1	NUM
fcis-25702	153	16	,	,	PUNCT
fcis-25702	153	17	we	we	PRON
fcis-25702	153	18	add	add	VERB
fcis-25702	153	19	a	a	DET
fcis-25702	153	20	small	small	ADJ
fcis-25702	153	21	value	value	NOUN
fcis-25702	153	22	t	t	NOUN
fcis-25702	153	23	to	to	PART
fcis-25702	153	24	increase	increase	VERB
fcis-25702	153	25	its	its	PRON
fcis-25702	153	26	probability	probability	NOUN
fcis-25702	153	27	of	of	ADP
fcis-25702	153	28	exiting	exit	VERB
fcis-25702	153	29	at	at	ADP
fcis-25702	153	30	the	the	DET
fcis-25702	153	31	next	next	ADJ
fcis-25702	153	32	layer	layer	NOUN
fcis-25702	153	33	,	,	PUNCT
fcis-25702	153	34	such	such	ADJ
fcis-25702	153	35	that	that	SCONJ
fcis-25702	153	36	p	p	NOUN
fcis-25702	153	37	skip	skip	NOUN
fcis-25702	153	38	t	t	NOUN
fcis-25702	153	39	pskip	pskip	NOUN
fcis-25702	153	40	.	.	PUNCT
fcis-25702	154	1	during	during	ADP
fcis-25702	154	2	training	training	NOUN
fcis-25702	154	3	,	,	PUNCT
fcis-25702	154	4	we	we	PRON
fcis-25702	154	5	gradually	gradually	ADV
fcis-25702	154	6	reduce	reduce	VERB
fcis-25702	154	7	the	the	DET
fcis-25702	154	8	number	number	NOUN
fcis-25702	154	9	m	m	VERB
fcis-25702	154	10	to	to	ADP
fcis-25702	154	11	0	0	NUM
fcis-25702	154	12	,	,	PUNCT
fcis-25702	154	13	where	where	SCONJ
fcis-25702	154	14	0	0	NUM
fcis-25702	154	15	means	mean	VERB
fcis-25702	154	16	that	that	SCONJ
fcis-25702	154	17	we	we	PRON
fcis-25702	154	18	wo	will	AUX
fcis-25702	154	19	n’t	not	PART
fcis-25702	154	20	increase	increase	VERB
fcis-25702	154	21	the	the	DET
fcis-25702	154	22	probability	probability	NOUN
fcis-25702	154	23	for	for	ADP
fcis-25702	154	24	any	any	DET
fcis-25702	154	25	layer	layer	NOUN
fcis-25702	154	26	.	.	PUNCT
fcis-25702	155	1	initially	initially	ADV
fcis-25702	155	2	,	,	PUNCT
fcis-25702	155	3	m	m	VERB
fcis-25702	155	4	is	be	AUX
fcis-25702	155	5	set	set	VERB
fcis-25702	155	6	at	at	ADP
fcis-25702	155	7	3	3	NUM
fcis-25702	155	8	.	.	PUNCT
fcis-25702	156	1	therefore	therefore	ADV
fcis-25702	156	2	,	,	PUNCT
fcis-25702	156	3	at	at	ADP
fcis-25702	156	4	the	the	DET
fcis-25702	156	5	initial	initial	ADJ
fcis-25702	156	6	stage	stage	NOUN
fcis-25702	156	7	of	of	ADP
fcis-25702	156	8	rl	rl	ADP
fcis-25702	156	9	training	training	NOUN
fcis-25702	156	10	,	,	PUNCT
fcis-25702	156	11	we	we	PRON
fcis-25702	156	12	encourage	encourage	VERB
fcis-25702	156	13	the	the	DET
fcis-25702	156	14	model	model	NOUN
fcis-25702	156	15	to	to	PART
fcis-25702	156	16	skip	skip	VERB
fcis-25702	156	17	more	more	ADV
fcis-25702	156	18	confident	confident	ADJ
fcis-25702	156	19	layers	layer	NOUN
fcis-25702	156	20	by	by	ADP
fcis-25702	156	21	measuring	measure	VERB
fcis-25702	156	22	their	their	PRON
fcis-25702	156	23	similarities	similarity	NOUN
fcis-25702	156	24	with	with	ADP
fcis-25702	156	25	previous	previous	ADJ
fcis-25702	156	26	layers	layer	NOUN
fcis-25702	156	27	,	,	PUNCT
fcis-25702	156	28	thereby	thereby	ADV
fcis-25702	156	29	eliminating	eliminate	VERB
fcis-25702	156	30	the	the	DET
fcis-25702	156	31	sampling	sampling	NOUN
fcis-25702	156	32	trajectories	trajectory	NOUN
fcis-25702	156	33	and	and	CCONJ
fcis-25702	156	34	stabilizing	stabilize	VERB
fcis-25702	156	35	rl	rl	ADP
fcis-25702	156	36	training	training	NOUN
fcis-25702	156	37	.	.	PUNCT
fcis-25702	157	1	objective	objective	ADJ
fcis-25702	157	2	function	function	VERB
fcis-25702	157	3	the	the	DET
fcis-25702	157	4	goal	goal	NOUN
fcis-25702	157	5	of	of	ADP
fcis-25702	157	6	reinforcement	reinforcement	NOUN
fcis-25702	157	7	learning	learning	NOUN
fcis-25702	157	8	is	be	AUX
fcis-25702	157	9	to	to	PART
fcis-25702	157	10	optimize	optimize	VERB
fcis-25702	157	11	the	the	DET
fcis-25702	157	12	policy	policy	NOUN
fcis-25702	157	13	network	network	NOUN
fcis-25702	157	14	to	to	PART
fcis-25702	157	15	maximize	maximize	VERB
fcis-25702	157	16	the	the	DET
fcis-25702	157	17	expected	expect	VERB
fcis-25702	157	18	reward	reward	NOUN
fcis-25702	157	19	.	.	PUNCT
fcis-25702	158	1	formally	formally	ADV
fcis-25702	158	2	,	,	PUNCT
fcis-25702	158	3	the	the	DET
fcis-25702	158	4	objective	objective	ADJ
fcis-25702	158	5	function	function	NOUN
fcis-25702	158	6	is	be	AUX
fcis-25702	158	7	defined	define	VERB
fcis-25702	158	8	as	as	SCONJ
fcis-25702	158	9	follows	follow	VERB
fcis-25702	158	10	:	:	PUNCT
fcis-25702	158	11	j	j	PROPN
fcis-25702	158	12	θ	θ	PROPN
fcis-25702	158	13	e	e	NOUN
fcis-25702	158	14	,	,	PUNCT
fcis-25702	158	15	∼	∼	NOUN
fcis-25702	158	16	∣	∣	ADJ
fcis-25702	158	17	;	;	PUNCT
fcis-25702	158	18	r	r	NOUN
fcis-25702	158	19	s	s	PROPN
fcis-25702	158	20	,	,	PUNCT
fcis-25702	158	21	a	a	PRON
fcis-25702	158	22	…	…	PUNCT
fcis-25702	158	23	s	s	NOUN
fcis-25702	158	24	,	,	PUNCT
fcis-25702	158	25	a	a	DET
fcis-25702	158	26	(	(	PUNCT
fcis-25702	158	27	12	12	NUM
fcis-25702	158	28	)	)	PUNCT
fcis-25702	158	29	where	where	SCONJ
fcis-25702	158	30	l	l	NOUN
fcis-25702	158	31	is	be	AUX
fcis-25702	158	32	the	the	DET
fcis-25702	158	33	total	total	ADJ
fcis-25702	158	34	number	number	NOUN
fcis-25702	158	35	of	of	ADP
fcis-25702	158	36	layers	layer	NOUN
fcis-25702	158	37	,	,	PUNCT
fcis-25702	158	38	representing	represent	VERB
fcis-25702	158	39	the	the	DET
fcis-25702	158	40	number	number	NOUN
fcis-25702	158	41	of	of	ADP
fcis-25702	158	42	states	state	NOUN
fcis-25702	158	43	.	.	PUNCT
fcis-25702	159	1	according	accord	VERB
fcis-25702	159	2	to	to	ADP
fcis-25702	159	3	the	the	DET
fcis-25702	159	4	reinforce	reinforce	NOUN
fcis-25702	159	5	algorithm	algorithm	NOUN
fcis-25702	159	6	[	[	X
fcis-25702	159	7	39	39	NUM
fcis-25702	159	8	]	]	PUNCT
fcis-25702	159	9	and	and	CCONJ
fcis-25702	159	10	the	the	DET
fcis-25702	159	11	policy	policy	NOUN
fcis-25702	159	12	gradient	gradient	NOUN
fcis-25702	159	13	method	method	NOUN
fcis-25702	159	14	[	[	X
fcis-25702	159	15	40	40	NUM
fcis-25702	159	16	]	]	PUNCT
fcis-25702	159	17	,	,	PUNCT
fcis-25702	159	18	we	we	PRON
fcis-25702	159	19	update	update	VERB
fcis-25702	159	20	the	the	DET
fcis-25702	159	21	network	network	NOUN
fcis-25702	159	22	with	with	ADP
fcis-25702	159	23	the	the	DET
fcis-25702	159	24	policy	policy	NOUN
fcis-25702	159	25	gradient	gradient	NOUN
fcis-25702	159	26	as	as	SCONJ
fcis-25702	159	27	follows	follow	VERB
fcis-25702	159	28	:	:	PUNCT
fcis-25702	159	29	j	j	PROPN
fcis-25702	159	30	θ	θ	INTJ
fcis-25702	159	31	∑	∑	PUNCT
fcis-25702	159	32	   	   	SPACE
fcis-25702	159	33	r	r	NOUN
fcis-25702	159	34	⋅	⋅	PROPN
fcis-25702	159	35	log	log	NOUN
fcis-25702	159	36	π	π	PROPN
fcis-25702	159	37	a	a	DET
fcis-25702	159	38	∣	∣	PROPN
fcis-25702	159	39	s	s	X
fcis-25702	159	40	(	(	PUNCT
fcis-25702	159	41	13	13	NUM
fcis-25702	159	42	)	)	PUNCT
fcis-25702	159	43	3.3	3.3	NUM
fcis-25702	159	44	.	.	PUNCT
fcis-25702	160	1	model	model	NOUN
fcis-25702	160	2	training	training	NOUN
fcis-25702	160	3	here	here	ADV
fcis-25702	160	4	we	we	PRON
fcis-25702	160	5	present	present	VERB
fcis-25702	160	6	our	our	PRON
fcis-25702	160	7	entire	entire	ADJ
fcis-25702	160	8	training	training	NOUN
fcis-25702	160	9	procedure	procedure	NOUN
fcis-25702	160	10	:	:	PUNCT
fcis-25702	160	11	(	(	PUNCT
fcis-25702	160	12	1	1	X
fcis-25702	160	13	)	)	PUNCT
fcis-25702	160	14	adapter	adapter	NOUN
fcis-25702	160	15	initialization	initialization	NOUN
fcis-25702	160	16	:	:	PUNCT
fcis-25702	160	17	fine	fine	ADJ
fcis-25702	160	18	-	-	PUNCT
fcis-25702	160	19	tune	tune	NOUN
fcis-25702	160	20	the	the	DET
fcis-25702	160	21	adapter	adapter	NOUN
fcis-25702	160	22	layer	layer	NOUN
fcis-25702	160	23	on	on	ADP
fcis-25702	160	24	instruction	instruction	NOUN
fcis-25702	160	25	-	-	PUNCT
fcis-25702	160	26	tuning	tune	VERB
fcis-25702	160	27	datasets	dataset	NOUN
fcis-25702	160	28	while	while	SCONJ
fcis-25702	160	29	freezing	freeze	VERB
fcis-25702	160	30	the	the	DET
fcis-25702	160	31	parameters	parameter	NOUN
fcis-25702	160	32	of	of	ADP
fcis-25702	160	33	the	the	DET
fcis-25702	160	34	llm	llm	PROPN
fcis-25702	160	35	model	model	NOUN
fcis-25702	160	36	.	.	PUNCT
fcis-25702	161	1	this	this	DET
fcis-25702	161	2	step	step	NOUN
fcis-25702	161	3	helps	help	VERB
fcis-25702	161	4	initialize	initialize	VERB
fcis-25702	161	5	the	the	DET
fcis-25702	161	6	parameters	parameter	NOUN
fcis-25702	161	7	of	of	ADP
fcis-25702	161	8	the	the	DET
fcis-25702	161	9	adapter	adapter	NOUN
fcis-25702	161	10	layers	layer	NOUN
fcis-25702	161	11	.	.	PUNCT
fcis-25702	162	1	(	(	PUNCT
fcis-25702	162	2	2	2	X
fcis-25702	162	3	)	)	PUNCT
fcis-25702	162	4	skip	skip	ADJ
fcis-25702	162	5	pretraining	pretraine	VERB
fcis-25702	162	6	:	:	PUNCT
fcis-25702	162	7	train	train	VERB
fcis-25702	162	8	the	the	DET
fcis-25702	162	9	parameters	parameter	NOUN
fcis-25702	162	10	of	of	ADP
fcis-25702	162	11	the	the	DET
fcis-25702	162	12	llm	llm	NOUN
fcis-25702	162	13	and	and	CCONJ
fcis-25702	162	14	adapter	adapter	NOUN
fcis-25702	162	15	layers	layer	NOUN
fcis-25702	162	16	on	on	ADP
fcis-25702	162	17	nlu	nlu	NOUN
fcis-25702	162	18	and	and	CCONJ
fcis-25702	162	19	mt	mt	PROPN
fcis-25702	162	20	datasets	dataset	NOUN
fcis-25702	162	21	using	use	VERB
fcis-25702	162	22	pretrain	pretrain	NOUN
fcis-25702	162	23	.	.	PUNCT
fcis-25702	162	24	.	.	PUNCT
fcis-25702	163	1	one	one	PRON
fcis-25702	163	2	can	can	AUX
fcis-25702	163	3	choose	choose	VERB
fcis-25702	163	4	to	to	PART
fcis-25702	163	5	update	update	VERB
fcis-25702	163	6	the	the	DET
fcis-25702	163	7	full	full	ADJ
fcis-25702	163	8	parameters	parameter	NOUN
fcis-25702	163	9	of	of	ADP
fcis-25702	163	10	the	the	DET
fcis-25702	163	11	llms	llm	NOUN
fcis-25702	163	12	or	or	CCONJ
fcis-25702	163	13	use	use	VERB
fcis-25702	163	14	parameter	parameter	NOUN
fcis-25702	163	15	-	-	PUNCT
fcis-25702	163	16	efficient	efficient	ADJ
fcis-25702	163	17	tuning	tuning	NOUN
fcis-25702	163	18	such	such	ADJ
fcis-25702	163	19	as	as	ADP
fcis-25702	163	20	lora	lora	PROPN
fcis-25702	163	21	[	[	X
fcis-25702	163	22	37	37	NUM
fcis-25702	163	23	]	]	PUNCT
fcis-25702	163	24	and	and	CCONJ
fcis-25702	163	25	qlora	qlora	X
fcis-25702	164	1	[	[	X
fcis-25702	164	2	41	41	NUM
fcis-25702	164	3	]	]	PUNCT
fcis-25702	164	4	.	.	PUNCT
fcis-25702	165	1	this	this	DET
fcis-25702	165	2	step	step	NOUN
fcis-25702	165	3	enhances	enhance	VERB
fcis-25702	165	4	the	the	DET
fcis-25702	165	5	model	model	NOUN
fcis-25702	165	6	’s	’s	PART
fcis-25702	165	7	ability	ability	NOUN
fcis-25702	165	8	to	to	PART
fcis-25702	165	9	perform	perform	VERB
fcis-25702	165	10	layer	layer	NOUN
fcis-25702	165	11	skipping	skip	VERB
fcis-25702	165	12	and	and	CCONJ
fcis-25702	165	13	also	also	ADV
fcis-25702	165	14	provides	provide	VERB
fcis-25702	165	15	a	a	DET
fcis-25702	165	16	good	good	ADJ
fcis-25702	165	17	initialization	initialization	NOUN
fcis-25702	165	18	for	for	ADP
fcis-25702	165	19	reinforcement	reinforcement	NOUN
fcis-25702	165	20	learning	learning	NOUN
fcis-25702	165	21	.	.	PUNCT
fcis-25702	166	1	(	(	PUNCT
fcis-25702	166	2	3	3	X
fcis-25702	166	3	)	)	PUNCT
fcis-25702	166	4	reinforcement	reinforcement	NOUN
fcis-25702	166	5	learning	learning	NOUN
fcis-25702	166	6	:	:	PUNCT
fcis-25702	166	7	conduct	conduct	VERB
fcis-25702	166	8	rl	rl	ADP
fcis-25702	166	9	on	on	ADP
fcis-25702	166	10	nlu	nlu	NOUN
fcis-25702	166	11	and	and	CCONJ
fcis-25702	166	12	mt	mt	PROPN
fcis-25702	166	13	datasets	dataset	NOUN
fcis-25702	166	14	.	.	PUNCT
fcis-25702	167	1	freeze	freeze	VERB
fcis-25702	167	2	the	the	DET
fcis-25702	167	3	parameters	parameter	NOUN
fcis-25702	167	4	of	of	ADP
fcis-25702	167	5	the	the	DET
fcis-25702	167	6	llms	llm	NOUN
fcis-25702	167	7	and	and	CCONJ
fcis-25702	167	8	lora	lora	PROPN
fcis-25702	167	9	,	,	PUNCT
fcis-25702	167	10	and	and	CCONJ
fcis-25702	167	11	update	update	VERB
fcis-25702	167	12	the	the	DET
fcis-25702	167	13	parameters	parameter	NOUN
fcis-25702	167	14	of	of	ADP
fcis-25702	167	15	the	the	DET
fcis-25702	167	16	controller	controller	NOUN
fcis-25702	167	17	,	,	PUNCT
fcis-25702	167	18	which	which	PRON
fcis-25702	167	19	includes	include	VERB
fcis-25702	167	20	the	the	DET
fcis-25702	167	21	adapter	adapter	NOUN
fcis-25702	167	22	layer	layer	NOUN
fcis-25702	167	23	and	and	CCONJ
fcis-25702	167	24	rl	rl	NOUN
fcis-25702	167	25	-	-	PUNCT
fcis-25702	167	26	specific	specific	ADJ
fcis-25702	167	27	head	head	NOUN
fcis-25702	167	28	.	.	PUNCT
fcis-25702	168	1	this	this	DET
fcis-25702	168	2	step	step	NOUN
fcis-25702	168	3	trains	train	VERB
fcis-25702	168	4	the	the	DET
fcis-25702	168	5	controller	controller	NOUN
fcis-25702	168	6	’s	’s	PART
fcis-25702	168	7	ability	ability	NOUN
fcis-25702	168	8	to	to	PART
fcis-25702	168	9	perform	perform	VERB
fcis-25702	168	10	layer	layer	NOUN
fcis-25702	168	11	skipping	skipping	NOUN
fcis-25702	168	12	.	.	PUNCT
fcis-25702	169	1	(	(	PUNCT
fcis-25702	169	2	4	4	X
fcis-25702	169	3	)	)	PUNCT
fcis-25702	169	4	fine	fine	ADV
fcis-25702	169	5	-	-	PUNCT
fcis-25702	169	6	tuning	tuning	NOUN
fcis-25702	169	7	:	:	PUNCT
fcis-25702	169	8	unfreeze	unfreeze	VERB
fcis-25702	169	9	the	the	DET
fcis-25702	169	10	parameters	parameter	NOUN
fcis-25702	169	11	of	of	ADP
fcis-25702	169	12	the	the	DET
fcis-25702	169	13	llms	llm	NOUN
fcis-25702	169	14	or	or	CCONJ
fcis-25702	169	15	lora	lora	PROPN
fcis-25702	169	16	and	and	CCONJ
fcis-25702	169	17	the	the	DET
fcis-25702	169	18	adapter	adapter	NOUN
fcis-25702	169	19	,	,	PUNCT
fcis-25702	169	20	and	and	CCONJ
fcis-25702	169	21	train	train	VERB
fcis-25702	169	22	the	the	DET
fcis-25702	169	23	entire	entire	ADJ
fcis-25702	169	24	model	model	NOUN
fcis-25702	169	25	with	with	ADP
fcis-25702	169	26	the	the	DET
fcis-25702	169	27	task	task	NOUN
fcis-25702	169	28	-	-	PUNCT
fcis-25702	169	29	specific	specific	ADJ
fcis-25702	169	30	objective	objective	NOUN
fcis-25702	169	31	and	and	CCONJ
fcis-25702	169	32	the	the	DET
fcis-25702	169	33	rl	rl	PROPN
fcis-25702	169	34	objective	objective	NOUN
fcis-25702	169	35	simultaneously	simultaneously	ADV
fcis-25702	169	36	.	.	PUNCT
fcis-25702	170	1	specifically	specifically	ADV
fcis-25702	170	2	,	,	PUNCT
fcis-25702	170	3	we	we	PRON
fcis-25702	170	4	adopt	adopt	VERB
fcis-25702	170	5	the	the	DET
fcis-25702	170	6	training	training	NOUN
fcis-25702	170	7	method	method	NOUN
fcis-25702	170	8	used	use	VERB
fcis-25702	170	9	in	in	ADP
fcis-25702	170	10	mixer	mixer	NOUN
fcis-25702	170	11	[	[	X
fcis-25702	170	12	42	42	NUM
fcis-25702	170	13	]	]	PUNCT
fcis-25702	170	14	,	,	PUNCT
fcis-25702	170	15	gradually	gradually	ADV
fcis-25702	170	16	reducing	reduce	VERB
fcis-25702	170	17	the	the	DET
fcis-25702	170	18	weight	weight	NOUN
fcis-25702	170	19	of	of	ADP
fcis-25702	170	20	the	the	DET
fcis-25702	170	21	task	task	NOUN
fcis-25702	170	22	-	-	PUNCT
fcis-25702	170	23	specific	specific	ADJ
fcis-25702	170	24	objective	objective	NOUN
fcis-25702	170	25	.	.	PUNCT
fcis-25702	171	1	3.4	3.4	NUM
fcis-25702	171	2	.	.	PUNCT
fcis-25702	171	3	inference	inference	NOUN
fcis-25702	171	4	with	with	ADP
fcis-25702	171	5	adaptive	adaptive	ADJ
fcis-25702	171	6	layer	layer	NOUN
fcis-25702	171	7	skipping	skip	VERB
fcis-25702	171	8	during	during	ADP
fcis-25702	171	9	inference	inference	NOUN
fcis-25702	171	10	,	,	PUNCT
fcis-25702	171	11	at	at	ADP
fcis-25702	171	12	layer	layer	NOUN
fcis-25702	171	13	l	l	NOUN
fcis-25702	171	14	1	1	NUM
fcis-25702	171	15	,	,	PUNCT
fcis-25702	171	16	we	we	PRON
fcis-25702	171	17	use	use	VERB
fcis-25702	171	18	the	the	DET
fcis-25702	171	19	controller	controller	NOUN
fcis-25702	171	20	installed	instal	VERB
fcis-25702	171	21	at	at	ADP
fcis-25702	171	22	this	this	DET
fcis-25702	171	23	layer	layer	NOUN
fcis-25702	171	24	to	to	PART
fcis-25702	171	25	predict	predict	VERB
fcis-25702	171	26	whether	whether	SCONJ
fcis-25702	171	27	to	to	PART
fcis-25702	171	28	skip	skip	VERB
fcis-25702	171	29	the	the	DET
fcis-25702	171	30	next	next	ADJ
fcis-25702	171	31	layer	layer	NOUN
fcis-25702	171	32	l.	l.	NOUN
fcis-25702	171	33	we	we	PRON
fcis-25702	171	34	use	use	VERB
fcis-25702	171	35	the	the	DET
fcis-25702	171	36	hidden	hide	VERB
fcis-25702	171	37	states	state	NOUN
fcis-25702	171	38	of	of	ADP
fcis-25702	171	39	layer	layer	NOUN
fcis-25702	171	40	l	l	NOUN
fcis-25702	171	41	1	1	NUM
fcis-25702	171	42	as	as	ADP
fcis-25702	171	43	the	the	DET
fcis-25702	171	44	hidden	hidden	ADJ
fcis-25702	171	45	states	state	NOUN
fcis-25702	171	46	for	for	ADP
fcis-25702	171	47	layer	layer	NOUN
fcis-25702	171	48	l	l	NOUN
fcis-25702	171	49	.	.	PUNCT
fcis-25702	172	1	additionally	additionally	ADV
fcis-25702	172	2	,	,	PUNCT
fcis-25702	172	3	as	as	SCONJ
fcis-25702	172	4	discussed	discuss	VERB
fcis-25702	172	5	previously	previously	ADV
fcis-25702	172	6	,	,	PUNCT
fcis-25702	172	7	with	with	ADP
fcis-25702	172	8	kv	kv	PROPN
fcis-25702	172	9	caching	cache	VERB
fcis-25702	172	10	,	,	PUNCT
fcis-25702	172	11	the	the	DET
fcis-25702	172	12	next	next	ADJ
fcis-25702	172	13	token	token	NOUN
fcis-25702	172	14	depends	depend	VERB
fcis-25702	172	15	on	on	ADP
fcis-25702	172	16	the	the	DET
fcis-25702	172	17	kv	kv	PROPN
fcis-25702	172	18	caches	cache	NOUN
fcis-25702	172	19	of	of	ADP
fcis-25702	172	20	the	the	DET
fcis-25702	172	21	current	current	ADJ
fcis-25702	172	22	and	and	CCONJ
fcis-25702	172	23	previous	previous	ADJ
fcis-25702	172	24	tokens	token	NOUN
fcis-25702	172	25	.	.	PUNCT
fcis-25702	173	1	skip	skip	VERB
fcis-25702	173	2	-	-	PUNCT
fcis-25702	173	3	decode	decode	NOUN
fcis-25702	173	4	[	[	X
fcis-25702	173	5	33	33	NUM
fcis-25702	173	6	]	]	PUNCT
fcis-25702	173	7	monotonically	monotonically	ADV
fcis-25702	173	8	decreases	decrease	VERB
fcis-25702	173	9	exit	exit	NOUN
fcis-25702	173	10	points	point	NOUN
fcis-25702	173	11	,	,	PUNCT
fcis-25702	173	12	but	but	CCONJ
fcis-25702	173	13	we	we	PRON
fcis-25702	173	14	believe	believe	VERB
fcis-25702	173	15	this	this	DET
fcis-25702	173	16	approach	approach	NOUN
fcis-25702	173	17	may	may	AUX
fcis-25702	173	18	restrict	restrict	VERB
fcis-25702	173	19	layer	layer	NOUN
fcis-25702	173	20	skipping	skip	VERB
fcis-25702	173	21	decisions	decision	NOUN
fcis-25702	173	22	and	and	CCONJ
fcis-25702	173	23	result	result	VERB
fcis-25702	173	24	in	in	ADP
fcis-25702	173	25	performance	performance	NOUN
fcis-25702	173	26	degradation	degradation	NOUN
fcis-25702	173	27	.	.	PUNCT
fcis-25702	174	1	thus	thus	ADV
fcis-25702	174	2	,	,	PUNCT
fcis-25702	174	3	we	we	PRON
fcis-25702	174	4	propose	propose	VERB
fcis-25702	174	5	:	:	PUNCT
fcis-25702	174	6	when	when	SCONJ
fcis-25702	174	7	a	a	DET
fcis-25702	174	8	layer	layer	NOUN
fcis-25702	174	9	i	i	PRON
fcis-25702	174	10	is	be	AUX
fcis-25702	174	11	skipped	skip	VERB
fcis-25702	174	12	for	for	ADP
fcis-25702	174	13	predicting	predict	VERB
fcis-25702	174	14	w	w	PROPN
fcis-25702	174	15	,	,	PUNCT
fcis-25702	174	16	the	the	DET
fcis-25702	174	17	kv	kv	PROPN
fcis-25702	174	18	caches	cache	NOUN
fcis-25702	174	19	of	of	ADP
fcis-25702	174	20	w	w	NOUN
fcis-25702	174	21	at	at	ADP
fcis-25702	174	22	layer	layer	NOUN
fcis-25702	174	23	l	l	NOUN
fcis-25702	174	24	are	be	AUX
fcis-25702	174	25	filled	fill	VERB
fcis-25702	174	26	with	with	ADP
fcis-25702	174	27	those	those	PRON
fcis-25702	174	28	from	from	ADP
fcis-25702	174	29	layer	layer	NOUN
fcis-25702	174	30	l	l	NOUN
fcis-25702	174	31	1	1	X
fcis-25702	174	32	.	.	PUNCT
fcis-25702	174	33	to	to	PART
fcis-25702	174	34	improve	improve	VERB
fcis-25702	174	35	the	the	DET
fcis-25702	174	36	efficiency	efficiency	NOUN
fcis-25702	174	37	of	of	ADP
fcis-25702	174	38	inference	inference	NOUN
fcis-25702	174	39	,	,	PUNCT
fcis-25702	174	40	one	one	PRON
fcis-25702	174	41	can	can	AUX
fcis-25702	174	42	also	also	ADV
fcis-25702	174	43	perform	perform	VERB
fcis-25702	174	44	layer	layer	NOUN
fcis-25702	174	45	skipping	skip	VERB
fcis-25702	174	46	at	at	ADP
fcis-25702	174	47	the	the	DET
fcis-25702	174	48	batch	batch	NOUN
fcis-25702	174	49	level	level	NOUN
fcis-25702	174	50	.	.	PUNCT
fcis-25702	175	1	specifically	specifically	ADV
fcis-25702	175	2	,	,	PUNCT
fcis-25702	175	3	a	a	DET
fcis-25702	175	4	layer	layer	NOUN
fcis-25702	175	5	will	will	AUX
fcis-25702	175	6	be	be	AUX
fcis-25702	175	7	skipped	skip	VERB
fcis-25702	175	8	if	if	SCONJ
fcis-25702	175	9	the	the	DET
fcis-25702	175	10	predicted	predict	VERB
fcis-25702	175	11	probabilities	probability	NOUN
fcis-25702	175	12	for	for	ADP
fcis-25702	175	13	most	most	ADJ
fcis-25702	175	14	of	of	ADP
fcis-25702	175	15	the	the	DET
fcis-25702	175	16	samples	sample	NOUN
fcis-25702	175	17	in	in	ADP
fcis-25702	175	18	the	the	DET
fcis-25702	175	19	batch	batch	NOUN
fcis-25702	175	20	(	(	PUNCT
fcis-25702	175	21	such	such	ADJ
fcis-25702	175	22	as	as	ADP
fcis-25702	175	23	more	more	ADJ
fcis-25702	175	24	than	than	ADP
fcis-25702	175	25	half	half	NOUN
fcis-25702	175	26	)	)	PUNCT
fcis-25702	175	27	exceed	exceed	VERB
fcis-25702	175	28	a	a	DET
fcis-25702	175	29	5	5	NUM
fcis-25702	175	30	threshold	threshold	NOUN
fcis-25702	175	31	of	of	ADP
fcis-25702	175	32	0.5	0.5	NUM
fcis-25702	175	33	.	.	PUNCT
fcis-25702	176	1	thus	thus	ADV
fcis-25702	176	2	,	,	PUNCT
fcis-25702	176	3	for	for	ADP
fcis-25702	176	4	inference	inference	NOUN
fcis-25702	176	5	,	,	PUNCT
fcis-25702	176	6	we	we	PRON
fcis-25702	176	7	can	can	AUX
fcis-25702	176	8	strike	strike	VERB
fcis-25702	176	9	a	a	DET
fcis-25702	176	10	balance	balance	NOUN
fcis-25702	176	11	:	:	PUNCT
fcis-25702	176	12	process	process	NOUN
fcis-25702	176	13	one	one	NUM
fcis-25702	176	14	sample	sample	NOUN
fcis-25702	176	15	at	at	ADP
fcis-25702	176	16	a	a	DET
fcis-25702	176	17	time	time	NOUN
fcis-25702	176	18	to	to	PART
fcis-25702	176	19	assign	assign	VERB
fcis-25702	176	20	different	different	ADJ
fcis-25702	176	21	skip	skip	ADJ
fcis-25702	176	22	layers	layer	NOUN
fcis-25702	176	23	for	for	ADP
fcis-25702	176	24	different	different	ADJ
fcis-25702	176	25	samples	sample	NOUN
fcis-25702	176	26	to	to	PART
fcis-25702	176	27	achieve	achieve	VERB
fcis-25702	176	28	more	more	ADV
fcis-25702	176	29	accurate	accurate	ADJ
fcis-25702	176	30	results	result	NOUN
fcis-25702	176	31	,	,	PUNCT
fcis-25702	176	32	or	or	CCONJ
fcis-25702	176	33	process	process	VERB
fcis-25702	176	34	a	a	DET
fcis-25702	176	35	batch	batch	NOUN
fcis-25702	176	36	at	at	ADP
fcis-25702	176	37	a	a	DET
fcis-25702	176	38	time	time	NOUN
fcis-25702	176	39	to	to	PART
fcis-25702	176	40	share	share	VERB
fcis-25702	176	41	the	the	DET
fcis-25702	176	42	same	same	ADJ
fcis-25702	176	43	skip	skip	ADJ
fcis-25702	176	44	layer	layer	NOUN
fcis-25702	176	45	for	for	ADP
fcis-25702	176	46	a	a	DET
fcis-25702	176	47	batch	batch	NOUN
fcis-25702	176	48	of	of	ADP
fcis-25702	176	49	samples	sample	NOUN
fcis-25702	176	50	to	to	PART
fcis-25702	176	51	infer	infer	VERB
fcis-25702	176	52	in	in	ADP
fcis-25702	176	53	parallel	parallel	NOUN
fcis-25702	176	54	.	.	PUNCT
fcis-25702	177	1	moreover	moreover	ADV
fcis-25702	177	2	,	,	PUNCT
fcis-25702	177	3	we	we	PRON
fcis-25702	177	4	can	can	AUX
fcis-25702	177	5	adjust	adjust	VERB
fcis-25702	177	6	the	the	DET
fcis-25702	177	7	threshold	threshold	NOUN
fcis-25702	177	8	to	to	PART
fcis-25702	177	9	skip	skip	VERB
fcis-25702	177	10	more	more	ADJ
fcis-25702	177	11	or	or	CCONJ
fcis-25702	177	12	fewer	few	ADJ
fcis-25702	177	13	layers	layer	NOUN
fcis-25702	177	14	to	to	PART
fcis-25702	177	15	achieve	achieve	VERB
fcis-25702	177	16	different	different	ADJ
fcis-25702	177	17	speed	speed	NOUN
fcis-25702	177	18	-	-	PUNCT
fcis-25702	177	19	up	up	ADP
fcis-25702	177	20	ratios	ratio	NOUN
fcis-25702	177	21	.	.	PUNCT
fcis-25702	178	1	4	4	X
fcis-25702	178	2	.	.	X
fcis-25702	178	3	experiments	experiment	NOUN
fcis-25702	178	4	4.1	4.1	NUM
fcis-25702	178	5	.	.	PUNCT
fcis-25702	179	1	datasets	dataset	NOUN
fcis-25702	179	2	we	we	PRON
fcis-25702	179	3	use	use	VERB
fcis-25702	179	4	four	four	NUM
fcis-25702	179	5	nlu	nlu	NOUN
fcis-25702	179	6	datasets	dataset	NOUN
fcis-25702	179	7	from	from	ADP
fcis-25702	179	8	the	the	DET
fcis-25702	179	9	glue	glue	NOUN
fcis-25702	179	10	benchmark	benchmark	NOUN
fcis-25702	180	1	[	[	X
fcis-25702	180	2	43	43	NUM
fcis-25702	180	3	]	]	X
fcis-25702	180	4	:	:	PUNCT
fcis-25702	180	5	rte	rte	NOUN
fcis-25702	180	6	(	(	PUNCT
fcis-25702	180	7	recognizing	recognize	VERB
fcis-25702	180	8	textual	textual	ADJ
fcis-25702	180	9	entailment	entailment	NOUN
fcis-25702	180	10	):	):	PUNCT
fcis-25702	180	11	determines	determine	VERB
fcis-25702	180	12	whether	whether	SCONJ
fcis-25702	180	13	one	one	NUM
fcis-25702	180	14	sentence	sentence	NOUN
fcis-25702	180	15	logically	logically	ADV
fcis-25702	180	16	follows	follow	VERB
fcis-25702	180	17	from	from	ADP
fcis-25702	180	18	another	another	PRON
fcis-25702	180	19	.	.	PUNCT
fcis-25702	181	1	mrpc	mrpc	PROPN
fcis-25702	181	2	(	(	PUNCT
fcis-25702	181	3	microsoft	microsoft	PROPN
fcis-25702	181	4	research	research	PROPN
fcis-25702	181	5	paraphrase	paraphrase	PROPN
fcis-25702	181	6	corpus	corpus	PROPN
fcis-25702	181	7	):	):	PUNCT
fcis-25702	181	8	identifies	identify	VERB
fcis-25702	181	9	whether	whether	SCONJ
fcis-25702	181	10	two	two	NUM
fcis-25702	181	11	sentences	sentence	NOUN
fcis-25702	181	12	are	be	AUX
fcis-25702	181	13	semantically	semantically	ADV
fcis-25702	181	14	equivalent	equivalent	ADJ
fcis-25702	181	15	.	.	PUNCT
fcis-25702	182	1	col	col	PROPN
fcis-25702	182	2	a	a	DET
fcis-25702	182	3	(	(	PUNCT
fcis-25702	182	4	corpus	corpus	NOUN
fcis-25702	182	5	of	of	ADP
fcis-25702	182	6	linguistic	linguistic	ADJ
fcis-25702	182	7	acceptability	acceptability	NOUN
fcis-25702	182	8	):	):	PUNCT
fcis-25702	182	9	assesses	assesse	NOUN
fcis-25702	182	10	whether	whether	SCONJ
fcis-25702	182	11	an	an	DET
fcis-25702	182	12	english	english	ADJ
fcis-25702	182	13	sentence	sentence	NOUN
fcis-25702	182	14	is	be	AUX
fcis-25702	182	15	linguistically	linguistically	ADV
fcis-25702	182	16	acceptable	acceptable	ADJ
fcis-25702	182	17	.	.	PUNCT
fcis-25702	182	18	sst-2	sst-2	PUNCT
fcis-25702	182	19	(	(	PUNCT
fcis-25702	182	20	stanford	stanford	PROPN
fcis-25702	182	21	sentiment	sentiment	PROPN
fcis-25702	182	22	treebank	treebank	VERB
fcis-25702	182	23	):	):	PUNCT
fcis-25702	182	24	evaluates	evaluate	VERB
fcis-25702	182	25	the	the	DET
fcis-25702	182	26	sentiment	sentiment	NOUN
fcis-25702	182	27	of	of	ADP
fcis-25702	182	28	a	a	DET
fcis-25702	182	29	sentence	sentence	NOUN
fcis-25702	182	30	as	as	ADP
fcis-25702	182	31	positive	positive	ADJ
fcis-25702	182	32	or	or	CCONJ
fcis-25702	182	33	negative	negative	ADJ
fcis-25702	182	34	.	.	PUNCT
fcis-25702	183	1	we	we	PRON
fcis-25702	183	2	use	use	VERB
fcis-25702	183	3	three	three	NUM
fcis-25702	183	4	machine	machine	NOUN
fcis-25702	183	5	translation	translation	NOUN
fcis-25702	183	6	datasets	dataset	NOUN
fcis-25702	183	7	:	:	PUNCT
fcis-25702	183	8	wmt	wmt	PROPN
fcis-25702	183	9	2023	2023	NUM
fcis-25702	183	10	(	(	PUNCT
fcis-25702	183	11	english	english	PROPN
fcis-25702	183	12	-	-	PUNCT
fcis-25702	183	13	chinese	chinese	PROPN
fcis-25702	183	14	):	):	PUNCT
fcis-25702	183	15	focuses	focus	VERB
fcis-25702	183	16	on	on	ADP
fcis-25702	183	17	translating	translate	VERB
fcis-25702	183	18	between	between	ADP
fcis-25702	183	19	chinese	chinese	PROPN
fcis-25702	183	20	and	and	CCONJ
fcis-25702	183	21	english	english	PROPN
fcis-25702	183	22	,	,	PUNCT
fcis-25702	183	23	particularly	particularly	ADV
fcis-25702	183	24	in	in	ADP
fcis-25702	183	25	news	news	NOUN
fcis-25702	183	26	articles	article	NOUN
fcis-25702	183	27	and	and	CCONJ
fcis-25702	183	28	other	other	ADJ
fcis-25702	183	29	formal	formal	ADJ
fcis-25702	183	30	content	content	NOUN
fcis-25702	183	31	.	.	PUNCT
fcis-25702	184	1	it	it	PRON
fcis-25702	184	2	is	be	AUX
fcis-25702	184	3	a	a	DET
fcis-25702	184	4	key	key	ADJ
fcis-25702	184	5	benchmark	benchmark	NOUN
fcis-25702	184	6	for	for	ADP
fcis-25702	184	7	chinese	chinese	ADJ
fcis-25702	184	8	-	-	PUNCT
fcis-25702	184	9	english	english	ADJ
fcis-25702	184	10	translation	translation	NOUN
fcis-25702	184	11	quality	quality	NOUN
fcis-25702	184	12	.	.	PUNCT
fcis-25702	185	1	wmt	wmt	PROPN
fcis-25702	185	2	2023	2023	NUM
fcis-25702	185	3	(	(	PUNCT
fcis-25702	185	4	english	english	ADJ
fcis-25702	185	5	-	-	PUNCT
fcis-25702	185	6	german	german	NOUN
fcis-25702	185	7	):	):	PUNCT
fcis-25702	185	8	evaluates	evaluate	NOUN
fcis-25702	185	9	machine	machine	NOUN
fcis-25702	185	10	translation	translation	NOUN
fcis-25702	185	11	systems	system	NOUN
fcis-25702	185	12	on	on	ADP
fcis-25702	185	13	translating	translate	VERB
fcis-25702	185	14	news	news	NOUN
fcis-25702	185	15	articles	article	NOUN
fcis-25702	185	16	between	between	ADP
fcis-25702	185	17	english	english	PROPN
fcis-25702	185	18	and	and	CCONJ
fcis-25702	185	19	german	german	NOUN
fcis-25702	185	20	.	.	PUNCT
fcis-25702	186	1	it	it	PRON
fcis-25702	186	2	is	be	AUX
fcis-25702	186	3	a	a	DET
fcis-25702	186	4	major	major	ADJ
fcis-25702	186	5	benchmark	benchmark	NOUN
fcis-25702	186	6	in	in	ADP
fcis-25702	186	7	the	the	DET
fcis-25702	186	8	translation	translation	NOUN
fcis-25702	186	9	community	community	NOUN
fcis-25702	186	10	.	.	PUNCT
fcis-25702	187	1	iwslt	iwslt	PROPN
fcis-25702	187	2	2023	2023	NUM
fcis-25702	187	3	(	(	PUNCT
fcis-25702	187	4	german	german	ADJ
fcis-25702	187	5	-	-	PUNCT
fcis-25702	187	6	english	english	ADJ
fcis-25702	187	7	):	):	PUNCT
fcis-25702	187	8	focuses	focus	VERB
fcis-25702	187	9	on	on	ADP
fcis-25702	187	10	spoken	spoken	ADJ
fcis-25702	187	11	language	language	NOUN
fcis-25702	187	12	translation	translation	NOUN
fcis-25702	187	13	,	,	PUNCT
fcis-25702	187	14	particularly	particularly	ADV
fcis-25702	187	15	for	for	ADP
fcis-25702	187	16	translating	translate	VERB
fcis-25702	187	17	ted	te	VERB
fcis-25702	187	18	talks	talk	NOUN
fcis-25702	187	19	and	and	CCONJ
fcis-25702	187	20	other	other	ADJ
fcis-25702	187	21	speech	speech	NOUN
fcis-25702	187	22	data	datum	NOUN
fcis-25702	187	23	from	from	ADP
fcis-25702	187	24	german	german	NOUN
fcis-25702	187	25	to	to	ADP
fcis-25702	187	26	english	english	PROPN
fcis-25702	187	27	.	.	PUNCT
fcis-25702	188	1	4.2	4.2	NUM
fcis-25702	188	2	.	.	PUNCT
fcis-25702	188	3	baseline	baseline	PROPN
fcis-25702	188	4	methods	method	NOUN
fcis-25702	188	5	we	we	PRON
fcis-25702	188	6	compare	compare	VERB
fcis-25702	188	7	our	our	PRON
fcis-25702	188	8	method	method	NOUN
fcis-25702	188	9	with	with	ADP
fcis-25702	188	10	the	the	DET
fcis-25702	188	11	following	follow	VERB
fcis-25702	188	12	baselines	baseline	NOUN
fcis-25702	188	13	:	:	PUNCT
fcis-25702	188	14	vanilla	vanilla	NOUN
fcis-25702	188	15	model	model	NOUN
fcis-25702	188	16	:	:	PUNCT
fcis-25702	189	1	we	we	PRON
fcis-25702	189	2	directly	directly	ADV
fcis-25702	189	3	fine	fine	ADJ
fcis-25702	189	4	-	-	PUNCT
fcis-25702	189	5	tune	tune	NOUN
fcis-25702	189	6	the	the	DET
fcis-25702	189	7	llm	llm	NOUN
fcis-25702	189	8	on	on	ADP
fcis-25702	189	9	nlu	nlu	NOUN
fcis-25702	189	10	and	and	CCONJ
fcis-25702	189	11	mt	mt	PROPN
fcis-25702	189	12	datasets	dataset	NOUN
fcis-25702	189	13	without	without	ADP
fcis-25702	189	14	skipping	skip	VERB
fcis-25702	189	15	any	any	DET
fcis-25702	189	16	layers	layer	NOUN
fcis-25702	189	17	.	.	PUNCT
fcis-25702	190	1	skip	skip	VERB
fcis-25702	190	2	-	-	PUNCT
fcis-25702	190	3	decode	decode	NOUN
fcis-25702	191	1	[	[	X
fcis-25702	191	2	33	33	NUM
fcis-25702	191	3	]	]	PUNCT
fcis-25702	191	4	:	:	PUNCT
fcis-25702	191	5	following	follow	VERB
fcis-25702	191	6	its	its	PRON
fcis-25702	191	7	method	method	NOUN
fcis-25702	191	8	,	,	PUNCT
fcis-25702	191	9	we	we	PRON
fcis-25702	191	10	skip	skip	VERB
fcis-25702	191	11	the	the	DET
fcis-25702	191	12	intermediate	intermediate	ADJ
fcis-25702	191	13	layers	layer	NOUN
fcis-25702	191	14	of	of	ADP
fcis-25702	191	15	llms	llm	NOUN
fcis-25702	191	16	.	.	PUNCT
fcis-25702	192	1	as	as	SCONJ
fcis-25702	192	2	more	more	ADJ
fcis-25702	192	3	tokens	token	NOUN
fcis-25702	192	4	are	be	AUX
fcis-25702	192	5	generated	generate	VERB
fcis-25702	192	6	,	,	PUNCT
fcis-25702	192	7	the	the	DET
fcis-25702	192	8	number	number	NOUN
fcis-25702	192	9	of	of	ADP
fcis-25702	192	10	skipped	skip	VERB
fcis-25702	192	11	layers	layer	NOUN
fcis-25702	192	12	increases	increase	NOUN
fcis-25702	192	13	.	.	PUNCT
fcis-25702	193	1	the	the	DET
fcis-25702	193	2	method	method	NOUN
fcis-25702	193	3	does	do	AUX
fcis-25702	193	4	not	not	PART
fcis-25702	193	5	involve	involve	VERB
fcis-25702	193	6	training	train	VERB
fcis-25702	193	7	the	the	DET
fcis-25702	193	8	llms	llm	NOUN
fcis-25702	193	9	;	;	PUNCT
fcis-25702	193	10	we	we	PRON
fcis-25702	193	11	replicate	replicate	VERB
fcis-25702	193	12	its	its	PRON
fcis-25702	193	13	method	method	NOUN
fcis-25702	193	14	and	and	CCONJ
fcis-25702	193	15	test	test	VERB
fcis-25702	193	16	it	it	PRON
fcis-25702	193	17	on	on	ADP
fcis-25702	193	18	nlu	nlu	NOUN
fcis-25702	193	19	and	and	CCONJ
fcis-25702	193	20	mt	mt	PROPN
fcis-25702	193	21	datasets	dataset	NOUN
fcis-25702	193	22	.	.	PUNCT
fcis-25702	194	1	static	static	ADJ
fcis-25702	194	2	layer	layer	NOUN
fcis-25702	194	3	skipping	skip	VERB
fcis-25702	194	4	[	[	X
fcis-25702	194	5	34	34	NUM
fcis-25702	194	6	]	]	X
fcis-25702	194	7	:	:	PUNCT
fcis-25702	194	8	this	this	DET
fcis-25702	194	9	method	method	NOUN
fcis-25702	194	10	directly	directly	ADV
fcis-25702	194	11	skips	skip	VERB
fcis-25702	194	12	the	the	DET
fcis-25702	194	13	intermediate	intermediate	ADJ
fcis-25702	194	14	layers	layer	NOUN
fcis-25702	194	15	of	of	ADP
fcis-25702	194	16	llms	llm	NOUN
fcis-25702	194	17	while	while	SCONJ
fcis-25702	194	18	keeping	keep	VERB
fcis-25702	194	19	the	the	DET
fcis-25702	194	20	last	last	ADJ
fcis-25702	194	21	layer	layer	NOUN
fcis-25702	194	22	.	.	PUNCT
fcis-25702	195	1	it	it	PRON
fcis-25702	195	2	then	then	ADV
fcis-25702	195	3	finetunes	finetune	VERB
fcis-25702	195	4	the	the	DET
fcis-25702	195	5	llms	llm	NOUN
fcis-25702	195	6	on	on	ADP
fcis-25702	195	7	nlu	nlu	NOUN
fcis-25702	195	8	and	and	CCONJ
fcis-25702	195	9	mt	mt	PROPN
fcis-25702	195	10	datasets	dataset	NOUN
fcis-25702	195	11	.	.	PUNCT
fcis-25702	196	1	layer	layer	NOUN
fcis-25702	196	2	skip	skip	VERB
fcis-25702	196	3	pretraining	pretraine	VERB
fcis-25702	196	4	:	:	PUNCT
fcis-25702	196	5	we	we	PRON
fcis-25702	196	6	skip	skip	VERB
fcis-25702	196	7	the	the	DET
fcis-25702	196	8	same	same	ADJ
fcis-25702	196	9	set	set	NOUN
fcis-25702	196	10	of	of	ADP
fcis-25702	196	11	layers	layer	NOUN
fcis-25702	196	12	as	as	SCONJ
fcis-25702	196	13	the	the	DET
fcis-25702	196	14	static	static	ADJ
fcis-25702	196	15	layer	layer	NOUN
fcis-25702	196	16	skipping	skip	VERB
fcis-25702	196	17	baseline	baseline	NOUN
fcis-25702	196	18	method	method	NOUN
fcis-25702	196	19	and	and	CCONJ
fcis-25702	196	20	fine	fine	ADJ
fcis-25702	196	21	-	-	PUNCT
fcis-25702	196	22	tune	tune	NOUN
fcis-25702	196	23	the	the	DET
fcis-25702	196	24	llms	llm	NOUN
fcis-25702	196	25	on	on	ADP
fcis-25702	196	26	nlu	nlu	NOUN
fcis-25702	196	27	and	and	CCONJ
fcis-25702	196	28	mt	mt	PROPN
fcis-25702	196	29	datasets	dataset	NOUN
fcis-25702	196	30	using	use	VERB
fcis-25702	196	31	the	the	DET
fcis-25702	196	32	layer	layer	NOUN
fcis-25702	196	33	skip	skip	VERB
fcis-25702	196	34	pretraining	pretraine	VERB
fcis-25702	196	35	method	method	NOUN
fcis-25702	196	36	proposed	propose	VERB
fcis-25702	196	37	in	in	ADP
fcis-25702	196	38	this	this	DET
fcis-25702	196	39	work	work	NOUN
fcis-25702	196	40	.	.	PUNCT
fcis-25702	197	1	4.3	4.3	NUM
fcis-25702	197	2	.	.	PUNCT
fcis-25702	198	1	experimental	experimental	ADJ
fcis-25702	198	2	settings	setting	NOUN
fcis-25702	198	3	devices	device	NOUN
fcis-25702	198	4	:	:	PUNCT
fcis-25702	198	5	we	we	PRON
fcis-25702	198	6	implement	implement	VERB
fcis-25702	198	7	our	our	PRON
fcis-25702	198	8	method	method	NOUN
fcis-25702	198	9	based	base	VERB
fcis-25702	198	10	on	on	ADP
fcis-25702	198	11	hugging	hug	VERB
fcis-25702	198	12	face	face	NOUN
fcis-25702	198	13	’s	’s	PART
fcis-25702	198	14	transformers	transformer	NOUN
fcis-25702	198	15	[	[	X
fcis-25702	198	16	44	44	NUM
fcis-25702	198	17	]	]	PUNCT
fcis-25702	198	18	.	.	PUNCT
fcis-25702	199	1	we	we	PRON
fcis-25702	199	2	conduct	conduct	VERB
fcis-25702	199	3	our	our	PRON
fcis-25702	199	4	experiments	experiment	NOUN
fcis-25702	199	5	on	on	ADP
fcis-25702	199	6	two	two	NUM
fcis-25702	199	7	nvidia	nvidia	NOUN
fcis-25702	199	8	v100	v100	PROPN
fcis-25702	199	9	32	32	NUM
fcis-25702	199	10	gb	gb	PROPN
fcis-25702	199	11	gpus	gpu	NOUN
fcis-25702	199	12	.	.	PUNCT
fcis-25702	200	1	ptm	ptm	NOUN
fcis-25702	200	2	models	model	NOUN
fcis-25702	200	3	we	we	PRON
fcis-25702	200	4	adopt	adopt	VERB
fcis-25702	200	5	the	the	DET
fcis-25702	200	6	following	follow	VERB
fcis-25702	200	7	llms	llm	NOUN
fcis-25702	200	8	for	for	ADP
fcis-25702	200	9	our	our	PRON
fcis-25702	200	10	experiments	experiment	NOUN
fcis-25702	200	11	:	:	PUNCT
fcis-25702	200	12	gpt2	gpt2	PROPN
fcis-25702	200	13	-	-	PUNCT
fcis-25702	200	14	xl	xl	PROPN
fcis-25702	201	1	[	[	X
fcis-25702	201	2	45	45	NUM
fcis-25702	201	3	]	]	PUNCT
fcis-25702	201	4	,	,	PUNCT
fcis-25702	201	5	opt-1.3b	opt-1.3b	VERB
fcis-25702	201	6	[	[	X
fcis-25702	201	7	46	46	NUM
fcis-25702	201	8	]	]	PUNCT
fcis-25702	201	9	,	,	PUNCT
fcis-25702	201	10	falcon-7b	falcon-7b	X
fcis-25702	202	1	[	[	PUNCT
fcis-25702	202	2	47	47	NUM
fcis-25702	202	3	]	]	PUNCT
fcis-25702	202	4	,	,	PUNCT
fcis-25702	202	5	llama2	llama2	PROPN
fcis-25702	202	6	-	-	PUNCT
fcis-25702	202	7	7b	7b	NOUN
fcis-25702	202	8	-	-	PUNCT
fcis-25702	202	9	chat	chat	NOUN
fcis-25702	203	1	[	[	X
fcis-25702	203	2	32	32	NUM
fcis-25702	203	3	]	]	PUNCT
fcis-25702	203	4	.	.	PUNCT
fcis-25702	204	1	readers	reader	NOUN
fcis-25702	204	2	are	be	AUX
fcis-25702	204	3	referred	refer	VERB
fcis-25702	204	4	to	to	AUX
fcis-25702	204	5	appendix	appendix	VERB
fcis-25702	204	6	a.3	a.3	PROPN
fcis-25702	204	7	for	for	ADP
fcis-25702	204	8	the	the	DET
fcis-25702	204	9	details	detail	NOUN
fcis-25702	204	10	of	of	ADP
fcis-25702	204	11	these	these	DET
fcis-25702	204	12	llms	llm	NOUN
fcis-25702	204	13	.	.	PUNCT
fcis-25702	205	1	training	training	NOUN
fcis-25702	205	2	settings	setting	NOUN
fcis-25702	205	3	for	for	ADP
fcis-25702	205	4	all	all	DET
fcis-25702	205	5	training	training	NOUN
fcis-25702	205	6	stages	stage	NOUN
fcis-25702	205	7	of	of	ADP
fcis-25702	205	8	our	our	PRON
fcis-25702	205	9	experiments	experiment	NOUN
fcis-25702	205	10	,	,	PUNCT
fcis-25702	205	11	we	we	PRON
fcis-25702	205	12	adopt	adopt	VERB
fcis-25702	205	13	the	the	DET
fcis-25702	205	14	following	follow	VERB
fcis-25702	205	15	settings	setting	NOUN
fcis-25702	205	16	:	:	PUNCT
fcis-25702	205	17	we	we	PRON
fcis-25702	205	18	use	use	VERB
fcis-25702	205	19	qlora	qlora	PROPN
fcis-25702	206	1	[	[	X
fcis-25702	206	2	41	41	NUM
fcis-25702	206	3	]	]	PUNCT
fcis-25702	206	4	to	to	PART
fcis-25702	206	5	train	train	VERB
fcis-25702	206	6	the	the	DET
fcis-25702	206	7	llm	llm	PROPN
fcis-25702	206	8	backbone	backbone	NOUN
fcis-25702	206	9	model	model	NOUN
fcis-25702	206	10	,	,	PUNCT
fcis-25702	206	11	with	with	SCONJ
fcis-25702	206	12	the	the	DET
fcis-25702	206	13	lora	lora	PROPN
fcis-25702	206	14	rank	rank	NOUN
fcis-25702	206	15	set	set	VERB
fcis-25702	206	16	to	to	ADP
fcis-25702	206	17	16	16	NUM
fcis-25702	206	18	.	.	PUNCT
fcis-25702	207	1	the	the	DET
fcis-25702	207	2	maximum	maximum	ADJ
fcis-25702	207	3	sequence	sequence	NOUN
fcis-25702	207	4	length	length	NOUN
fcis-25702	207	5	is	be	AUX
fcis-25702	207	6	1024	1024	NUM
fcis-25702	207	7	,	,	PUNCT
fcis-25702	207	8	and	and	CCONJ
fcis-25702	207	9	we	we	PRON
fcis-25702	207	10	adopt	adopt	VERB
fcis-25702	207	11	the	the	DET
fcis-25702	207	12	efficient	efficient	ADJ
fcis-25702	207	13	sequence	sequence	NOUN
fcis-25702	207	14	packing	pack	VERB
fcis-25702	207	15	strategy	strategy	NOUN
fcis-25702	208	1	[	[	X
fcis-25702	208	2	48	48	NUM
fcis-25702	208	3	]	]	PUNCT
fcis-25702	208	4	to	to	PART
fcis-25702	208	5	accelerate	accelerate	VERB
fcis-25702	208	6	training	training	NOUN
fcis-25702	208	7	.	.	PUNCT
fcis-25702	209	1	the	the	DET
fcis-25702	209	2	batch	batch	NOUN
fcis-25702	209	3	size	size	NOUN
fcis-25702	209	4	is	be	AUX
fcis-25702	209	5	set	set	VERB
fcis-25702	209	6	to	to	ADP
fcis-25702	209	7	64	64	NUM
fcis-25702	209	8	.	.	PUNCT
fcis-25702	210	1	we	we	PRON
fcis-25702	210	2	employ	employ	VERB
fcis-25702	210	3	adamw	adamw	NOUN
fcis-25702	210	4	[	[	X
fcis-25702	210	5	49	49	NUM
fcis-25702	210	6	]	]	PUNCT
fcis-25702	210	7	as	as	ADP
fcis-25702	210	8	the	the	DET
fcis-25702	210	9	optimizer	optimizer	NOUN
fcis-25702	210	10	.	.	PUNCT
fcis-25702	211	1	for	for	ADP
fcis-25702	211	2	different	different	ADJ
fcis-25702	211	3	training	training	NOUN
fcis-25702	211	4	stages	stage	NOUN
fcis-25702	211	5	,	,	PUNCT
fcis-25702	211	6	we	we	PRON
fcis-25702	211	7	adopt	adopt	VERB
fcis-25702	211	8	different	different	ADJ
fcis-25702	211	9	learning	learning	NOUN
fcis-25702	211	10	rates	rate	NOUN
fcis-25702	211	11	and	and	CCONJ
fcis-25702	211	12	epochs	epoch	NOUN
fcis-25702	211	13	:	:	PUNCT
fcis-25702	211	14	in	in	ADP
fcis-25702	211	15	the	the	DET
fcis-25702	211	16	adapter	adapter	NOUN
fcis-25702	211	17	initialization	initialization	NOUN
fcis-25702	211	18	stage	stage	NOUN
fcis-25702	211	19	,	,	PUNCT
fcis-25702	211	20	the	the	DET
fcis-25702	211	21	epoch	epoch	NOUN
fcis-25702	211	22	is	be	AUX
fcis-25702	211	23	set	set	VERB
fcis-25702	211	24	to	to	ADP
fcis-25702	211	25	1	1	NUM
fcis-25702	211	26	and	and	CCONJ
fcis-25702	211	27	the	the	DET
fcis-25702	211	28	learning	learning	NOUN
fcis-25702	211	29	rate	rate	NOUN
fcis-25702	211	30	to	to	ADP
fcis-25702	211	31	1e	1e	PROPN
fcis-25702	211	32	4	4	NUM
fcis-25702	211	33	.	.	PUNCT
fcis-25702	212	1	in	in	ADP
fcis-25702	212	2	the	the	DET
fcis-25702	212	3	skip	skip	NOUN
fcis-25702	212	4	pretraining	pretraine	VERB
fcis-25702	212	5	stage	stage	NOUN
fcis-25702	212	6	,	,	PUNCT
fcis-25702	212	7	the	the	DET
fcis-25702	212	8	epoch	epoch	NOUN
fcis-25702	212	9	is	be	AUX
fcis-25702	212	10	set	set	VERB
fcis-25702	212	11	to	to	ADP
fcis-25702	212	12	5	5	NUM
fcis-25702	212	13	and	and	CCONJ
fcis-25702	212	14	the	the	DET
fcis-25702	212	15	learning	learning	NOUN
fcis-25702	212	16	rate	rate	NOUN
fcis-25702	212	17	to	to	ADP
fcis-25702	212	18	5e-5	5e-5	NUM
fcis-25702	212	19	.	.	PUNCT
fcis-25702	213	1	in	in	ADP
fcis-25702	213	2	the	the	DET
fcis-25702	213	3	reinforcement	reinforcement	NOUN
fcis-25702	213	4	learning	learning	NOUN
fcis-25702	213	5	stage	stage	NOUN
fcis-25702	213	6	,	,	PUNCT
fcis-25702	213	7	the	the	DET
fcis-25702	213	8	epoch	epoch	NOUN
fcis-25702	213	9	is	be	AUX
fcis-25702	213	10	set	set	VERB
fcis-25702	213	11	to	to	ADP
fcis-25702	213	12	10	10	NUM
fcis-25702	213	13	and	and	CCONJ
fcis-25702	213	14	the	the	DET
fcis-25702	213	15	learning	learning	NOUN
fcis-25702	213	16	rate	rate	NOUN
fcis-25702	213	17	to	to	ADP
fcis-25702	213	18	2e	2e	PROPN
fcis-25702	213	19	5	5	NUM
fcis-25702	213	20	.	.	PUNCT
fcis-25702	214	1	in	in	ADP
fcis-25702	214	2	the	the	DET
fcis-25702	214	3	fine	fine	ADV
fcis-25702	214	4	-	-	PUNCT
fcis-25702	214	5	tuning	tune	VERB
fcis-25702	214	6	stage	stage	NOUN
fcis-25702	214	7	,	,	PUNCT
fcis-25702	214	8	the	the	DET
fcis-25702	214	9	epoch	epoch	NOUN
fcis-25702	214	10	is	be	AUX
fcis-25702	214	11	set	set	VERB
fcis-25702	214	12	to	to	ADP
fcis-25702	214	13	5	5	NUM
fcis-25702	214	14	and	and	CCONJ
fcis-25702	214	15	the	the	DET
fcis-25702	214	16	learning	learning	NOUN
fcis-25702	214	17	rate	rate	NOUN
fcis-25702	214	18	to	to	ADP
fcis-25702	214	19	1e	1e	PROPN
fcis-25702	214	20	5	5	NUM
fcis-25702	214	21	.	.	PUNCT
fcis-25702	215	1	for	for	ADP
fcis-25702	215	2	each	each	DET
fcis-25702	215	3	method	method	NOUN
fcis-25702	215	4	,	,	PUNCT
fcis-25702	215	5	we	we	PRON
fcis-25702	215	6	use	use	VERB
fcis-25702	215	7	different	different	ADJ
fcis-25702	215	8	settings	setting	NOUN
fcis-25702	215	9	to	to	PART
fcis-25702	215	10	instruct	instruct	VERB
fcis-25702	215	11	the	the	DET
fcis-25702	215	12	model	model	NOUN
fcis-25702	215	13	to	to	PART
fcis-25702	215	14	skip	skip	VERB
fcis-25702	215	15	more	more	ADJ
fcis-25702	215	16	or	or	CCONJ
fcis-25702	215	17	fewer	few	ADJ
fcis-25702	215	18	layers	layer	NOUN
fcis-25702	215	19	,	,	PUNCT
fcis-25702	215	20	and	and	CCONJ
fcis-25702	215	21	we	we	PRON
fcis-25702	215	22	obtain	obtain	VERB
fcis-25702	215	23	the	the	DET
fcis-25702	215	24	results	result	NOUN
fcis-25702	215	25	of	of	ADP
fcis-25702	215	26	these	these	DET
fcis-25702	215	27	different	different	ADJ
fcis-25702	215	28	settings	setting	NOUN
fcis-25702	215	29	.	.	PUNCT
fcis-25702	216	1	in	in	ADP
fcis-25702	216	2	detail	detail	NOUN
fcis-25702	216	3	,	,	PUNCT
fcis-25702	216	4	for	for	ADP
fcis-25702	216	5	skip	skip	NOUN
fcis-25702	216	6	-	-	PUNCT
fcis-25702	216	7	decode	decode	NOUN
fcis-25702	216	8	,	,	PUNCT
fcis-25702	216	9	we	we	PRON
fcis-25702	216	10	set	set	VERB
fcis-25702	216	11	the	the	DET
fcis-25702	216	12	maximum	maximum	ADJ
fcis-25702	216	13	ratio	ratio	NOUN
fcis-25702	216	14	of	of	ADP
fcis-25702	216	15	skipped	skip	VERB
fcis-25702	216	16	layers	layer	NOUN
fcis-25702	216	17	to	to	AUX
fcis-25702	216	18	total	total	ADJ
fcis-25702	216	19	layers	layer	NOUN
fcis-25702	216	20	at	at	ADP
fcis-25702	216	21	90	90	NUM
fcis-25702	216	22	%	%	NOUN
fcis-25702	216	23	,	,	PUNCT
fcis-25702	216	24	70	70	NUM
fcis-25702	216	25	%	%	NOUN
fcis-25702	216	26	,	,	PUNCT
fcis-25702	216	27	50	50	NUM
fcis-25702	216	28	%	%	NOUN
fcis-25702	216	29	,	,	PUNCT
fcis-25702	216	30	30	30	NUM
fcis-25702	216	31	%	%	NOUN
fcis-25702	216	32	,	,	PUNCT
fcis-25702	216	33	and	and	CCONJ
fcis-25702	216	34	10	10	NUM
fcis-25702	216	35	%	%	NOUN
fcis-25702	216	36	;	;	PUNCT
fcis-25702	216	37	for	for	ADP
fcis-25702	216	38	the	the	DET
fcis-25702	216	39	yuan	yuan	PROPN
fcis-25702	216	40	et	et	PROPN
fcis-25702	216	41	al	al	PROPN
fcis-25702	216	42	.	.	PUNCT
fcis-25702	217	1	[	[	X
fcis-25702	217	2	34	34	NUM
fcis-25702	217	3	]	]	SYM
fcis-25702	217	4	method	method	NOUN
fcis-25702	217	5	,	,	PUNCT
fcis-25702	217	6	we	we	PRON
fcis-25702	217	7	follow	follow	VERB
fcis-25702	217	8	the	the	DET
fcis-25702	217	9	same	same	ADJ
fcis-25702	217	10	settings	setting	NOUN
fcis-25702	217	11	as	as	ADP
fcis-25702	217	12	skip	skip	NOUN
fcis-25702	217	13	-	-	PUNCT
fcis-25702	217	14	decode	decode	NOUN
fcis-25702	217	15	;	;	PUNCT
fcis-25702	217	16	for	for	ADP
fcis-25702	217	17	our	our	PRON
fcis-25702	217	18	method	method	NOUN
fcis-25702	217	19	,	,	PUNCT
fcis-25702	217	20	we	we	PRON
fcis-25702	217	21	set	set	VERB
fcis-25702	217	22	the	the	DET
fcis-25702	217	23	thresholds	threshold	NOUN
fcis-25702	217	24	for	for	ADP
fcis-25702	217	25	skipping	skip	VERB
fcis-25702	217	26	layers	layer	NOUN
fcis-25702	217	27	at	at	ADP
fcis-25702	217	28	0.8,0.7,0.5,0.4	0.8,0.7,0.5,0.4	NOUN
fcis-25702	217	29	,	,	PUNCT
fcis-25702	217	30	and	and	CCONJ
fcis-25702	217	31	0.3	0.3	NUM
fcis-25702	217	32	.	.	PUNCT
fcis-25702	218	1	4.4	4.4	NUM
fcis-25702	218	2	.	.	PUNCT
fcis-25702	219	1	main	main	ADJ
fcis-25702	219	2	results	result	NOUN
fcis-25702	219	3	table	table	NOUN
fcis-25702	219	4	1	1	NUM
fcis-25702	219	5	.	.	PUNCT
fcis-25702	220	1	f1	f1	NOUN
fcis-25702	220	2	-	-	PUNCT
fcis-25702	220	3	score	score	NOUN
fcis-25702	220	4	(	(	PUNCT
fcis-25702	220	5	%	%	NOUN
fcis-25702	220	6	)	)	PUNCT
fcis-25702	220	7	of	of	ADP
fcis-25702	220	8	methods	method	NOUN
fcis-25702	220	9	using	use	VERB
fcis-25702	220	10	the	the	DET
fcis-25702	220	11	llama-7b	llama-7b	NOUN
fcis-25702	220	12	-	-	NOUN
fcis-25702	220	13	chat	chat	ADJ
fcis-25702	220	14	backbone	backbone	NOUN
fcis-25702	220	15	on	on	ADP
fcis-25702	220	16	nlu	nlu	NOUN
fcis-25702	220	17	tasks	task	NOUN
fcis-25702	220	18	across	across	ADP
fcis-25702	220	19	five	five	NUM
fcis-25702	220	20	random	random	ADJ
fcis-25702	220	21	seeds	seed	NOUN
fcis-25702	220	22	.	.	PUNCT
fcis-25702	221	1	avg	avg	PROPN
fcis-25702	221	2	represents	represent	VERB
fcis-25702	221	3	the	the	DET
fcis-25702	221	4	average	average	ADJ
fcis-25702	221	5	score	score	NOUN
fcis-25702	221	6	achieved	achieve	VERB
fcis-25702	221	7	by	by	ADP
fcis-25702	221	8	skipping	skip	VERB
fcis-25702	221	9	different	different	ADJ
fcis-25702	221	10	numbers	number	NOUN
fcis-25702	221	11	of	of	ADP
fcis-25702	221	12	layers	layer	NOUN
fcis-25702	221	13	,	,	PUNCT
fcis-25702	221	14	and	and	CCONJ
fcis-25702	221	15	best	good	ADJ
fcis-25702	221	16	represents	represent	VERB
fcis-25702	221	17	the	the	DET
fcis-25702	221	18	best	good	ADJ
fcis-25702	221	19	score	score	NOUN
fcis-25702	221	20	among	among	ADP
fcis-25702	221	21	all	all	DET
fcis-25702	221	22	settings	setting	NOUN
fcis-25702	221	23	.	.	PUNCT
fcis-25702	222	1	rte	rte	PROPN
fcis-25702	222	2	mrpc	mrpc	PROPN
fcis-25702	222	3	cola	cola	PROPN
fcis-25702	222	4	sst-2	sst-2	PUNCT
fcis-25702	222	5	avg	avg	PROPN
fcis-25702	222	6	best	good	ADJ
fcis-25702	222	7	avg	avg	PROPN
fcis-25702	222	8	best	good	ADJ
fcis-25702	222	9	avg	avg	PROPN
fcis-25702	222	10	best	good	ADJ
fcis-25702	222	11	avg	avg	PROPN
fcis-25702	222	12	best	good	ADJ
fcis-25702	222	13	vanilla	vanilla	NOUN
fcis-25702	222	14	model	model	NOUN
fcis-25702	222	15	77.1	77.1	NUM
fcis-25702	222	16	77.1	77.1	NUM
fcis-25702	222	17	89.8	89.8	NUM
fcis-25702	222	18	89.8	89.8	NUM
fcis-25702	222	19	57.7	57.7	NUM
fcis-25702	222	20	57.7	57.7	NUM
fcis-25702	222	21	91.9	91.9	NUM
fcis-25702	222	22	91.9	91.9	NUM
fcis-25702	222	23	skip	skip	ADJ
fcis-25702	222	24	-	-	PUNCT
fcis-25702	222	25	decode	decode	NOUN
fcis-25702	222	26	44.7	44.7	NUM
fcis-25702	222	27	48.9	48.9	NUM
fcis-25702	222	28	69.1	69.1	NUM
fcis-25702	222	29	73.2	73.2	NUM
fcis-25702	222	30	41.6	41.6	NUM
fcis-25702	222	31	47.5	47.5	NUM
fcis-25702	222	32	81.0	81.0	NUM
fcis-25702	222	33	82.5	82.5	NUM
fcis-25702	222	34	yuan	yuan	NOUN
fcis-25702	222	35	et	et	NOUN
fcis-25702	222	36	al	al	PROPN
fcis-25702	222	37	.	.	PUNCT
fcis-25702	223	1	[	[	X
fcis-25702	223	2	34	34	NUM
fcis-25702	223	3	]	]	SYM
fcis-25702	223	4	62.8	62.8	NUM
fcis-25702	223	5	68.6	68.6	NUM
fcis-25702	223	6	76.7	76.7	NUM
fcis-25702	223	7	82.9	82.9	NUM
fcis-25702	223	8	48.7	48.7	NUM
fcis-25702	223	9	51.3	51.3	NUM
fcis-25702	223	10	86.2	86.2	NUM
fcis-25702	223	11	89.1	89.1	NUM
fcis-25702	223	12	layer	layer	NOUN
fcis-25702	223	13	skip	skip	NOUN
fcis-25702	223	14	pretraining	pretraine	VERB
fcis-25702	223	15	64.7	64.7	NUM
fcis-25702	223	16	73.1	73.1	NUM
fcis-25702	223	17	79.0	79.0	NUM
fcis-25702	223	18	85.2	85.2	NUM
fcis-25702	223	19	51.2	51.2	NUM
fcis-25702	223	20	56.3	56.3	NUM
fcis-25702	223	21	88.7	88.7	NUM
fcis-25702	223	22	91.3	91.3	NUM
fcis-25702	223	23	ours(bs=1	ours(bs=1	ADJ
fcis-25702	223	24	)	)	PUNCT
fcis-25702	223	25	68.8	68.8	NUM
fcis-25702	223	26	77.5	77.5	NUM
fcis-25702	223	27	85.6	85.6	NUM
fcis-25702	223	28	90.5	90.5	NUM
fcis-25702	223	29	49.4	49.4	NUM
fcis-25702	223	30	58.1	58.1	NUM
fcis-25702	223	31	90.6	90.6	NUM
fcis-25702	223	32	92.7	92.7	NUM
fcis-25702	223	33	ours(bs=8	ours(bs=8	NOUN
fcis-25702	223	34	)	)	PUNCT
fcis-25702	223	35	67.7	67.7	NUM
fcis-25702	223	36	77.1	77.1	NUM
fcis-25702	223	37	85.0	85.0	NUM
fcis-25702	223	38	90.3	90.3	NUM
fcis-25702	223	39	46.7	46.7	NUM
fcis-25702	223	40	57.2	57.2	NUM
fcis-25702	223	41	90.3	90.3	NUM
fcis-25702	223	42	92.5	92.5	NUM
fcis-25702	223	43	ours(bs=16	ours(bs=16	PROPN
fcis-25702	223	44	)	)	PUNCT
fcis-25702	223	45	66.5	66.5	NUM
fcis-25702	223	46	76.5	76.5	NUM
fcis-25702	223	47	84.7	84.7	NUM
fcis-25702	223	48	89.7	89.7	NUM
fcis-25702	223	49	45.8	45.8	NUM
fcis-25702	223	50	56.5	56.5	NUM
fcis-25702	223	51	90.0	90.0	NUM
fcis-25702	223	52	92.1	92.1	NUM
fcis-25702	223	53	to	to	PART
fcis-25702	223	54	validate	validate	VERB
fcis-25702	223	55	the	the	DET
fcis-25702	223	56	efficacy	efficacy	NOUN
fcis-25702	223	57	of	of	ADP
fcis-25702	223	58	our	our	PRON
fcis-25702	223	59	proposed	propose	VERB
fcis-25702	223	60	method	method	NOUN
fcis-25702	223	61	,	,	PUNCT
fcis-25702	223	62	we	we	PRON
fcis-25702	223	63	conducted	conduct	VERB
fcis-25702	223	64	comparative	comparative	ADJ
fcis-25702	223	65	experiments	experiment	NOUN
fcis-25702	223	66	against	against	ADP
fcis-25702	223	67	baseline	baseline	NOUN
fcis-25702	223	68	layer	layer	NOUN
fcis-25702	223	69	skipping	skip	VERB
fcis-25702	223	70	or	or	CCONJ
fcis-25702	223	71	pruning	prune	VERB
fcis-25702	223	72	methods	method	NOUN
fcis-25702	223	73	.	.	PUNCT
fcis-25702	224	1	the	the	DET
fcis-25702	224	2	experimental	experimental	ADJ
fcis-25702	224	3	results	result	NOUN
fcis-25702	224	4	are	be	AUX
fcis-25702	224	5	presented	present	VERB
fcis-25702	224	6	in	in	ADP
fcis-25702	224	7	table	table	NOUN
fcis-25702	224	8	1	1	NUM
fcis-25702	224	9	and	and	CCONJ
fcis-25702	224	10	table	table	NOUN
fcis-25702	224	11	2	2	NUM
fcis-25702	224	12	.	.	PUNCT
fcis-25702	225	1	‘	'	PUNCT
fcis-25702	225	2	bs	bs	X
fcis-25702	225	3	1	1	NUM
fcis-25702	225	4	’	'	PUNCT
fcis-25702	225	5	denotes	denote	NOUN
fcis-25702	225	6	we	we	PRON
fcis-25702	225	7	do	do	VERB
fcis-25702	225	8	dynamic	dynamic	ADJ
fcis-25702	225	9	layer	layer	NOUN
fcis-25702	225	10	skipping	skip	VERB
fcis-25702	225	11	for	for	ADP
fcis-25702	225	12	each	each	DET
fcis-25702	225	13	sample	sample	NOUN
fcis-25702	225	14	,	,	PUNCT
fcis-25702	225	15	‘	'	PUNCT
fcis-25702	225	16	bs	bs	X
fcis-25702	225	17	8	8	NUM
fcis-25702	225	18	or	or	CCONJ
fcis-25702	225	19	bs	bs	PRON
fcis-25702	225	20	16	16	NUM
fcis-25702	225	21	’	'	PUNCT
fcis-25702	225	22	means	mean	VERB
fcis-25702	225	23	we	we	PRON
fcis-25702	225	24	use	use	VERB
fcis-25702	225	25	batch	batch	NOUN
fcis-25702	225	26	-	-	PUNCT
fcis-25702	225	27	level	level	NOUN
fcis-25702	225	28	skipping	skipping	NOUN
fcis-25702	225	29	:	:	PUNCT
fcis-25702	225	30	all	all	DET
fcis-25702	225	31	samples	sample	NOUN
fcis-25702	225	32	in	in	ADP
fcis-25702	225	33	a	a	DET
fcis-25702	225	34	batch	batch	NOUN
fcis-25702	225	35	skip	skip	VERB
fcis-25702	225	36	6	6	NUM
fcis-25702	225	37	the	the	DET
fcis-25702	225	38	same	same	ADJ
fcis-25702	225	39	layers	layer	NOUN
fcis-25702	225	40	.	.	PUNCT
fcis-25702	226	1	avg	avg	PROPN
fcis-25702	226	2	represents	represent	VERB
fcis-25702	226	3	the	the	DET
fcis-25702	226	4	average	average	ADJ
fcis-25702	226	5	score	score	NOUN
fcis-25702	226	6	from	from	ADP
fcis-25702	226	7	different	different	ADJ
fcis-25702	226	8	layer	layer	NOUN
fcis-25702	226	9	skipping	skip	VERB
fcis-25702	226	10	settings	setting	NOUN
fcis-25702	226	11	mentioned	mention	VERB
fcis-25702	226	12	in	in	ADP
fcis-25702	226	13	the	the	DET
fcis-25702	226	14	training	training	NOUN
fcis-25702	226	15	settings	setting	NOUN
fcis-25702	226	16	section	section	NOUN
fcis-25702	226	17	,	,	PUNCT
fcis-25702	226	18	and	and	CCONJ
fcis-25702	226	19	best	good	ADJ
fcis-25702	226	20	represents	represent	VERB
fcis-25702	226	21	the	the	DET
fcis-25702	226	22	best	good	ADJ
fcis-25702	226	23	score	score	NOUN
fcis-25702	226	24	among	among	ADP
fcis-25702	226	25	all	all	DET
fcis-25702	226	26	settings	setting	NOUN
fcis-25702	226	27	.	.	PUNCT
fcis-25702	227	1	we	we	PRON
fcis-25702	227	2	set	set	VERB
fcis-25702	227	3	different	different	ADJ
fcis-25702	227	4	thresholds	threshold	NOUN
fcis-25702	227	5	to	to	PART
fcis-25702	227	6	control	control	VERB
fcis-25702	227	7	the	the	DET
fcis-25702	227	8	number	number	NOUN
fcis-25702	227	9	of	of	ADP
fcis-25702	227	10	layers	layer	NOUN
fcis-25702	227	11	to	to	PART
fcis-25702	227	12	be	be	AUX
fcis-25702	227	13	skipped	skip	VERB
fcis-25702	227	14	.	.	PUNCT
fcis-25702	228	1	we	we	PRON
fcis-25702	228	2	present	present	VERB
fcis-25702	228	3	the	the	DET
fcis-25702	228	4	results	result	NOUN
fcis-25702	228	5	of	of	ADP
fcis-25702	228	6	other	other	ADJ
fcis-25702	228	7	llms	llm	NOUN
fcis-25702	228	8	in	in	ADP
fcis-25702	228	9	appendix	appendix	ADJ
fcis-25702	228	10	a.3.1	a.3.1	NOUN
fcis-25702	228	11	.	.	PUNCT
fcis-25702	229	1	as	as	SCONJ
fcis-25702	229	2	we	we	PRON
fcis-25702	229	3	can	can	AUX
fcis-25702	229	4	see	see	VERB
fcis-25702	229	5	,	,	PUNCT
fcis-25702	229	6	by	by	ADP
fcis-25702	229	7	fine	fine	ADV
fcis-25702	229	8	-	-	PUNCT
fcis-25702	229	9	tuning	tuning	NOUN
fcis-25702	229	10	on	on	ADP
fcis-25702	229	11	a	a	DET
fcis-25702	229	12	text	text	NOUN
fcis-25702	229	13	-	-	PUNCT
fcis-25702	229	14	classification	classification	NOUN
fcis-25702	229	15	dataset	dataset	NOUN
fcis-25702	229	16	,	,	PUNCT
fcis-25702	229	17	yuan	yuan	PROPN
fcis-25702	229	18	et	et	NOUN
fcis-25702	229	19	al	al	PROPN
fcis-25702	229	20	.	.	PUNCT
fcis-25702	230	1	[	[	X
fcis-25702	230	2	34	34	NUM
fcis-25702	230	3	]	]	PUNCT
fcis-25702	230	4	achieves	achieve	VERB
fcis-25702	230	5	better	well	ADJ
fcis-25702	230	6	results	result	NOUN
fcis-25702	230	7	with	with	ADP
fcis-25702	230	8	more	more	ADJ
fcis-25702	230	9	skipped	skip	VERB
fcis-25702	230	10	layers	layer	NOUN
fcis-25702	230	11	compared	compare	VERB
fcis-25702	230	12	to	to	ADP
fcis-25702	230	13	skip	skip	VERB
fcis-25702	230	14	-	-	PUNCT
fcis-25702	230	15	decode	decode	NOUN
fcis-25702	230	16	.	.	PUNCT
fcis-25702	231	1	the	the	DET
fcis-25702	231	2	layer	layer	NOUN
fcis-25702	231	3	skip	skip	VERB
fcis-25702	231	4	pretraining	pretraine	VERB
fcis-25702	231	5	method	method	NOUN
fcis-25702	231	6	also	also	ADV
fcis-25702	231	7	boosts	boost	VERB
fcis-25702	231	8	the	the	DET
fcis-25702	231	9	performance	performance	NOUN
fcis-25702	231	10	of	of	ADP
fcis-25702	231	11	llms	llm	NOUN
fcis-25702	231	12	.	.	PUNCT
fcis-25702	232	1	our	our	PRON
fcis-25702	232	2	method	method	NOUN
fcis-25702	232	3	significantly	significantly	ADV
fcis-25702	232	4	outperforms	outperform	VERB
fcis-25702	232	5	the	the	DET
fcis-25702	232	6	yuan	yuan	NOUN
fcis-25702	232	7	et	et	NOUN
fcis-25702	232	8	al	al	PROPN
fcis-25702	232	9	.	.	PUNCT
fcis-25702	233	1	[	[	X
fcis-25702	233	2	34	34	NUM
fcis-25702	233	3	]	]	SYM
fcis-25702	233	4	method	method	NOUN
fcis-25702	233	5	,	,	PUNCT
fcis-25702	233	6	which	which	PRON
fcis-25702	233	7	applies	apply	VERB
fcis-25702	233	8	the	the	DET
fcis-25702	233	9	same	same	ADJ
fcis-25702	233	10	layer	layer	NOUN
fcis-25702	233	11	settings	setting	NOUN
fcis-25702	233	12	for	for	ADP
fcis-25702	233	13	all	all	DET
fcis-25702	233	14	samples	sample	NOUN
fcis-25702	233	15	,	,	PUNCT
fcis-25702	233	16	validating	validate	VERB
fcis-25702	233	17	the	the	DET
fcis-25702	233	18	effectiveness	effectiveness	NOUN
fcis-25702	233	19	of	of	ADP
fcis-25702	233	20	dynamic	dynamic	ADJ
fcis-25702	233	21	layer	layer	NOUN
fcis-25702	233	22	skipping	skip	VERB
fcis-25702	233	23	.	.	PUNCT
fcis-25702	234	1	table	table	NOUN
fcis-25702	234	2	2	2	NUM
fcis-25702	234	3	.	.	PUNCT
fcis-25702	234	4	bleu	bleu	NOUN
fcis-25702	234	5	score	score	NOUN
fcis-25702	234	6	(	(	PUNCT
fcis-25702	234	7	%	%	INTJ
fcis-25702	234	8	)	)	PUNCT
fcis-25702	234	9	of	of	ADP
fcis-25702	234	10	methods	method	NOUN
fcis-25702	234	11	using	use	VERB
fcis-25702	234	12	the	the	DET
fcis-25702	234	13	llama-7b	llama-7b	NOUN
fcis-25702	234	14	-	-	NOUN
fcis-25702	234	15	chat	chat	ADJ
fcis-25702	234	16	backbone	backbone	NOUN
fcis-25702	234	17	on	on	ADP
fcis-25702	234	18	mt	mt	PROPN
fcis-25702	234	19	tasks	task	NOUN
fcis-25702	234	20	across	across	ADP
fcis-25702	234	21	five	five	NUM
fcis-25702	234	22	random	random	ADJ
fcis-25702	234	23	seeds	seed	NOUN
fcis-25702	234	24	.	.	PUNCT
fcis-25702	235	1	avg	avg	PROPN
fcis-25702	235	2	represents	represent	VERB
fcis-25702	235	3	the	the	DET
fcis-25702	235	4	average	average	ADJ
fcis-25702	235	5	score	score	NOUN
fcis-25702	235	6	achieved	achieve	VERB
fcis-25702	235	7	by	by	ADP
fcis-25702	235	8	skipping	skip	VERB
fcis-25702	235	9	different	different	ADJ
fcis-25702	235	10	numbers	number	NOUN
fcis-25702	235	11	of	of	ADP
fcis-25702	235	12	layers	layer	NOUN
fcis-25702	235	13	,	,	PUNCT
fcis-25702	235	14	and	and	CCONJ
fcis-25702	235	15	best	good	ADJ
fcis-25702	235	16	represents	represent	VERB
fcis-25702	235	17	the	the	DET
fcis-25702	235	18	best	good	ADJ
fcis-25702	235	19	score	score	NOUN
fcis-25702	235	20	among	among	ADP
fcis-25702	235	21	all	all	DET
fcis-25702	235	22	settings	setting	NOUN
fcis-25702	235	23	.	.	PUNCT
fcis-25702	236	1	wmt	wmt	PROPN
fcis-25702	236	2	2023	2023	NUM
fcis-25702	236	3	wmt	wmt	PROPN
fcis-25702	236	4	2023	2023	NUM
fcis-25702	236	5	iwslt	iwslt	ADJ
fcis-25702	236	6	2023	2023	NUM
fcis-25702	236	7	language	language	NOUN
fcis-25702	236	8	pair	pair	NOUN
fcis-25702	236	9	english	english	PROPN
fcis-25702	236	10	-	-	PUNCT
fcis-25702	236	11	chinese	chinese	ADJ
fcis-25702	236	12	english	english	ADJ
fcis-25702	236	13	-	-	PUNCT
fcis-25702	236	14	german	german	ADJ
fcis-25702	236	15	german	german	ADJ
fcis-25702	236	16	-	-	PUNCT
fcis-25702	236	17	english	english	ADJ
fcis-25702	236	18	avg	avg	PROPN
fcis-25702	236	19	best	good	ADJ
fcis-25702	236	20	avg	avg	PROPN
fcis-25702	236	21	best	good	ADJ
fcis-25702	236	22	avg	avg	PROPN
fcis-25702	236	23	best	good	ADJ
fcis-25702	236	24	vanilla	vanilla	PROPN
fcis-25702	236	25	model	model	NOUN
fcis-25702	236	26	38.7	38.7	NUM
fcis-25702	236	27	38.7	38.7	NUM
fcis-25702	236	28	37.8	37.8	NUM
fcis-25702	236	29	37.8	37.8	NUM
fcis-25702	236	30	33.2	33.2	NUM
fcis-25702	236	31	33.2	33.2	NUM
fcis-25702	236	32	skip	skip	NOUN
fcis-25702	236	33	-	-	PUNCT
fcis-25702	236	34	decode	decode	NOUN
fcis-25702	236	35	32.8	32.8	NUM
fcis-25702	236	36	34.6	34.6	NUM
fcis-25702	236	37	36.4	36.4	NUM
fcis-25702	236	38	37.1	37.1	NUM
fcis-25702	236	39	30.8	30.8	NUM
fcis-25702	236	40	33.1	33.1	NUM
fcis-25702	236	41	yuan	yuan	NOUN
fcis-25702	236	42	et	et	NOUN
fcis-25702	236	43	al	al	PROPN
fcis-25702	236	44	.	.	PUNCT
fcis-25702	237	1	[	[	X
fcis-25702	237	2	34	34	NUM
fcis-25702	237	3	]	]	SYM
fcis-25702	237	4	35.6	35.6	NUM
fcis-25702	237	5	37.3	37.3	NUM
fcis-25702	237	6	38.2	38.2	NUM
fcis-25702	237	7	33.2	33.2	NUM
fcis-25702	237	8	33.0	33.0	NUM
fcis-25702	237	9	36.1	36.1	NUM
fcis-25702	237	10	layer	layer	NOUN
fcis-25702	237	11	skip	skip	NOUN
fcis-25702	237	12	pretraining	pretraine	VERB
fcis-25702	237	13	37.5	37.5	NUM
fcis-25702	237	14	37.9	37.9	NUM
fcis-25702	237	15	38.5	38.5	NUM
fcis-25702	237	16	34.1	34.1	NUM
fcis-25702	237	17	33.3	33.3	NUM
fcis-25702	237	18	34.1	34.1	NUM
fcis-25702	237	19	ours(bs=1	ours(bs=1	ADJ
fcis-25702	237	20	)	)	PUNCT
fcis-25702	237	21	39.1	39.1	NUM
fcis-25702	237	22	39.8	39.8	NUM
fcis-25702	237	23	38.2	38.2	NUM
fcis-25702	237	24	38.9	38.9	NUM
fcis-25702	237	25	35.0	35.0	NUM
fcis-25702	237	26	35.9	35.9	NUM
fcis-25702	237	27	ours(bs=8	ours(bs=8	PROPN
fcis-25702	237	28	)	)	PUNCT
fcis-25702	237	29	38.5	38.5	NUM
fcis-25702	237	30	38.9	38.9	NUM
fcis-25702	237	31	37.8	37.8	NUM
fcis-25702	237	32	38.0	38.0	NUM
fcis-25702	237	33	34.7	34.7	NUM
fcis-25702	237	34	35.2	35.2	NUM
fcis-25702	237	35	ours(bs=16	ours(bs=16	PROPN
fcis-25702	237	36	)	)	PUNCT
fcis-25702	237	37	38.0	38.0	NUM
fcis-25702	237	38	38.4	38.4	NUM
fcis-25702	237	39	37.7	37.7	NUM
fcis-25702	237	40	38.1	38.1	NUM
fcis-25702	237	41	34.7	34.7	NUM
fcis-25702	237	42	35.0	35.0	NUM
fcis-25702	237	43	4.5	4.5	NUM
fcis-25702	237	44	.	.	PUNCT
fcis-25702	238	1	discussions	discussion	NOUN
fcis-25702	238	2	and	and	CCONJ
fcis-25702	238	3	ablation	ablation	NOUN
fcis-25702	238	4	studies	study	NOUN
fcis-25702	238	5	we	we	PRON
fcis-25702	238	6	present	present	VERB
fcis-25702	238	7	the	the	DET
fcis-25702	238	8	f1	f1	NOUN
fcis-25702	238	9	-	-	PUNCT
fcis-25702	238	10	score	score	NOUN
fcis-25702	238	11	of	of	ADP
fcis-25702	238	12	the	the	DET
fcis-25702	238	13	llama2	llama2	PROPN
fcis-25702	238	14	-	-	PUNCT
fcis-25702	238	15	7b	7b	NUM
fcis-25702	238	16	-	-	PUNCT
fcis-25702	238	17	chat	chat	NOUN
fcis-25702	238	18	model	model	NOUN
fcis-25702	238	19	with	with	ADP
fcis-25702	238	20	different	different	ADJ
fcis-25702	238	21	layer	layer	NOUN
fcis-25702	238	22	skipping	skip	VERB
fcis-25702	238	23	ratios	ratio	NOUN
fcis-25702	238	24	on	on	ADP
fcis-25702	238	25	the	the	DET
fcis-25702	238	26	qnli	qnli	PROPN
fcis-25702	238	27	dataset	dataset	VERB
fcis-25702	238	28	in	in	ADP
fcis-25702	238	29	figure	figure	NOUN
fcis-25702	238	30	3	3	NUM
fcis-25702	238	31	.	.	PUNCT
fcis-25702	239	1	the	the	DET
fcis-25702	239	2	layer	layer	NOUN
fcis-25702	239	3	skipping	skip	VERB
fcis-25702	239	4	ratio	ratio	NOUN
fcis-25702	239	5	is	be	AUX
fcis-25702	239	6	computed	compute	VERB
fcis-25702	239	7	by	by	ADP
fcis-25702	239	8	dividing	divide	VERB
fcis-25702	239	9	the	the	DET
fcis-25702	239	10	number	number	NOUN
fcis-25702	239	11	of	of	ADP
fcis-25702	239	12	skipped	skip	VERB
fcis-25702	239	13	layers	layer	NOUN
fcis-25702	239	14	by	by	ADP
fcis-25702	239	15	the	the	DET
fcis-25702	239	16	total	total	ADJ
fcis-25702	239	17	number	number	NOUN
fcis-25702	239	18	of	of	ADP
fcis-25702	239	19	layers	layer	NOUN
fcis-25702	239	20	.	.	PUNCT
fcis-25702	240	1	the	the	DET
fcis-25702	240	2	results	result	NOUN
fcis-25702	240	3	show	show	VERB
fcis-25702	240	4	that	that	SCONJ
fcis-25702	240	5	skipping	skip	VERB
fcis-25702	240	6	nearly	nearly	ADV
fcis-25702	240	7	30	30	NUM
fcis-25702	240	8	%	%	NOUN
fcis-25702	240	9	of	of	ADP
fcis-25702	240	10	the	the	DET
fcis-25702	240	11	layers	layer	NOUN
fcis-25702	240	12	can	can	AUX
fcis-25702	240	13	even	even	ADV
fcis-25702	240	14	improve	improve	VERB
fcis-25702	240	15	performance	performance	NOUN
fcis-25702	240	16	,	,	PUNCT
fcis-25702	240	17	demonstrating	demonstrate	VERB
fcis-25702	240	18	the	the	DET
fcis-25702	240	19	redundancy	redundancy	NOUN
fcis-25702	240	20	in	in	ADP
fcis-25702	240	21	llm	llm	PROPN
fcis-25702	240	22	layers	layer	NOUN
fcis-25702	240	23	.	.	PUNCT
fcis-25702	241	1	the	the	DET
fcis-25702	241	2	reasons	reason	NOUN
fcis-25702	241	3	why	why	SCONJ
fcis-25702	241	4	skipping	skip	VERB
fcis-25702	241	5	many	many	ADJ
fcis-25702	241	6	layers	layer	NOUN
fcis-25702	241	7	allows	allow	VERB
fcis-25702	241	8	llms	llm	NOUN
fcis-25702	241	9	to	to	PART
fcis-25702	241	10	still	still	ADV
fcis-25702	241	11	achieve	achieve	VERB
fcis-25702	241	12	comparable	comparable	ADJ
fcis-25702	241	13	results	result	NOUN
fcis-25702	241	14	are	be	AUX
fcis-25702	241	15	twofold	twofold	ADJ
fcis-25702	241	16	:	:	PUNCT
fcis-25702	241	17	(	(	PUNCT
fcis-25702	241	18	1	1	X
fcis-25702	241	19	)	)	PUNCT
fcis-25702	241	20	much	much	ADJ
fcis-25702	241	21	work	work	NOUN
fcis-25702	242	1	[	[	X
fcis-25702	242	2	29][50][16	29][50][16	NUM
fcis-25702	242	3	]	]	X
fcis-25702	242	4	proves	prove	VERB
fcis-25702	242	5	that	that	SCONJ
fcis-25702	242	6	the	the	DET
fcis-25702	242	7	parameters	parameter	NOUN
fcis-25702	242	8	of	of	ADP
fcis-25702	242	9	llms	llm	NOUN
fcis-25702	242	10	are	be	AUX
fcis-25702	242	11	redundant	redundant	ADJ
fcis-25702	242	12	.	.	PUNCT
fcis-25702	243	1	(	(	PUNCT
fcis-25702	243	2	2	2	X
fcis-25702	243	3	)	)	PUNCT
fcis-25702	243	4	the	the	DET
fcis-25702	243	5	parameters	parameter	NOUN
fcis-25702	243	6	of	of	ADP
fcis-25702	243	7	llms	llm	NOUN
fcis-25702	243	8	can	can	AUX
fcis-25702	243	9	be	be	AUX
fcis-25702	243	10	further	far	ADV
fcis-25702	243	11	reduced	reduce	VERB
fcis-25702	243	12	if	if	SCONJ
fcis-25702	243	13	they	they	PRON
fcis-25702	243	14	are	be	AUX
fcis-25702	243	15	used	use	VERB
fcis-25702	243	16	only	only	ADV
fcis-25702	243	17	for	for	ADP
fcis-25702	243	18	certain	certain	ADJ
fcis-25702	243	19	tasks	task	NOUN
fcis-25702	243	20	,	,	PUNCT
fcis-25702	243	21	such	such	ADJ
fcis-25702	243	22	as	as	ADP
fcis-25702	243	23	nlu	nlu	NOUN
fcis-25702	243	24	tasks	task	NOUN
fcis-25702	243	25	.	.	PUNCT
fcis-25702	244	1	llms	llm	NOUN
fcis-25702	244	2	do	do	AUX
fcis-25702	244	3	not	not	PART
fcis-25702	244	4	need	need	VERB
fcis-25702	244	5	the	the	DET
fcis-25702	244	6	extra	extra	ADJ
fcis-25702	244	7	parameters	parameter	NOUN
fcis-25702	244	8	to	to	PART
fcis-25702	244	9	store	store	VERB
fcis-25702	244	10	knowledge	knowledge	NOUN
fcis-25702	244	11	and	and	CCONJ
fcis-25702	244	12	make	make	VERB
fcis-25702	244	13	complex	complex	ADJ
fcis-25702	244	14	inferences	inference	NOUN
fcis-25702	244	15	for	for	ADP
fcis-25702	244	16	solving	solve	VERB
fcis-25702	244	17	other	other	ADJ
fcis-25702	244	18	tasks	task	NOUN
fcis-25702	244	19	such	such	ADJ
fcis-25702	244	20	as	as	ADP
fcis-25702	244	21	code	code	NOUN
fcis-25702	244	22	completion	completion	NOUN
fcis-25702	244	23	and	and	CCONJ
fcis-25702	244	24	multi	multi	ADJ
fcis-25702	244	25	-	-	ADJ
fcis-25702	244	26	turn	turn	ADJ
fcis-25702	244	27	question	question	NOUN
fcis-25702	244	28	answering	answering	NOUN
fcis-25702	244	29	.	.	PUNCT
fcis-25702	245	1	4.5.1	4.5.1	X
fcis-25702	245	2	.	.	PUNCT
fcis-25702	245	3	ablation	ablation	NOUN
fcis-25702	245	4	on	on	ADP
fcis-25702	245	5	dynamic	dynamic	ADJ
fcis-25702	245	6	layer	layer	NOUN
fcis-25702	245	7	skipping	skip	VERB
fcis-25702	245	8	for	for	ADP
fcis-25702	245	9	a	a	DET
fcis-25702	245	10	llm	llm	NOUN
fcis-25702	245	11	with	with	ADP
fcis-25702	245	12	l	l	NOUN
fcis-25702	245	13	layers	layer	NOUN
fcis-25702	245	14	,	,	PUNCT
fcis-25702	245	15	we	we	PRON
fcis-25702	245	16	initialize	initialize	VERB
fcis-25702	245	17	an	an	DET
fcis-25702	245	18	array	array	NOUN
fcis-25702	245	19	of	of	ADP
fcis-25702	245	20	length	length	NOUN
fcis-25702	245	21	l	l	NOUN
fcis-25702	245	22	and	and	CCONJ
fcis-25702	245	23	fill	fill	VERB
fcis-25702	245	24	it	it	PRON
fcis-25702	245	25	with	with	ADP
fcis-25702	245	26	0	0	NUM
fcis-25702	245	27	;	;	PUNCT
fcis-25702	245	28	for	for	ADP
fcis-25702	245	29	the	the	DET
fcis-25702	245	30	skipped	skip	VERB
fcis-25702	245	31	layer	layer	NOUN
fcis-25702	245	32	,	,	PUNCT
fcis-25702	245	33	we	we	PRON
fcis-25702	245	34	set	set	VERB
fcis-25702	245	35	its	its	PRON
fcis-25702	245	36	index	index	NOUN
fcis-25702	245	37	in	in	ADP
fcis-25702	245	38	the	the	DET
fcis-25702	245	39	array	array	NOUN
fcis-25702	245	40	to	to	ADP
fcis-25702	245	41	1	1	NUM
fcis-25702	245	42	.	.	PUNCT
fcis-25702	246	1	then	then	ADV
fcis-25702	246	2	,	,	PUNCT
fcis-25702	246	3	we	we	PRON
fcis-25702	246	4	can	can	AUX
fcis-25702	246	5	obtain	obtain	VERB
fcis-25702	246	6	arrays	array	VERB
fcis-25702	246	7	a	a	DET
fcis-25702	246	8	,	,	PUNCT
fcis-25702	246	9	a	a	DET
fcis-25702	246	10	,	,	PUNCT
fcis-25702	246	11	…	…	PUNCT
fcis-25702	246	12	,	,	PUNCT
fcis-25702	246	13	a	a	PRON
fcis-25702	246	14	,	,	PUNCT
fcis-25702	246	15	where	where	SCONJ
fcis-25702	246	16	n	n	PRON
fcis-25702	246	17	is	be	AUX
fcis-25702	246	18	the	the	DET
fcis-25702	246	19	total	total	ADJ
fcis-25702	246	20	number	number	NOUN
fcis-25702	246	21	of	of	ADP
fcis-25702	246	22	samples	sample	NOUN
fcis-25702	246	23	.	.	PUNCT
fcis-25702	247	1	since	since	SCONJ
fcis-25702	247	2	the	the	DET
fcis-25702	247	3	length	length	NOUN
fcis-25702	247	4	of	of	ADP
fcis-25702	247	5	an	an	DET
fcis-25702	247	6	array	array	NOUN
fcis-25702	247	7	a	a	PRON
fcis-25702	247	8	is	be	AUX
fcis-25702	247	9	l	l	NOUN
fcis-25702	247	10	,	,	PUNCT
fcis-25702	247	11	for	for	ADP
fcis-25702	247	12	each	each	DET
fcis-25702	247	13	position	position	NOUN
fcis-25702	247	14	l	l	NOUN
fcis-25702	247	15	in	in	ADP
fcis-25702	247	16	a	a	PRON
fcis-25702	247	17	,	,	PUNCT
fcis-25702	247	18	we	we	PRON
fcis-25702	247	19	can	can	AUX
fcis-25702	247	20	compute	compute	VERB
fcis-25702	247	21	the	the	DET
fcis-25702	247	22	variance	variance	NOUN
fcis-25702	247	23	for	for	ADP
fcis-25702	247	24	all	all	DET
fcis-25702	247	25	arrays	array	NOUN
fcis-25702	247	26	,	,	PUNCT
fcis-25702	247	27	and	and	CCONJ
fcis-25702	247	28	we	we	PRON
fcis-25702	247	29	average	average	VERB
fcis-25702	247	30	the	the	DET
fcis-25702	247	31	variance	variance	NOUN
fcis-25702	247	32	for	for	ADP
fcis-25702	247	33	all	all	DET
fcis-25702	247	34	positions	position	NOUN
fcis-25702	247	35	by	by	ADP
fcis-25702	247	36	:	:	PUNCT
fcis-25702	247	37	∑	∑	PUNCT
fcis-25702	247	38	    	    	SPACE
fcis-25702	247	39	(	(	PUNCT
fcis-25702	247	40	14	14	NUM
fcis-25702	247	41	)	)	PUNCT
fcis-25702	247	42	∑	∑	DET
fcis-25702	247	43	    	    	SPACE
fcis-25702	247	44	∑	∑	PROPN
fcis-25702	247	45	    	    	SPACE
fcis-25702	247	46	(	(	PUNCT
fcis-25702	247	47	15	15	NUM
fcis-25702	247	48	)	)	PUNCT
fcis-25702	247	49	where	where	SCONJ
fcis-25702	247	50	a	a	DET
fcis-25702	247	51	the	the	DET
fcis-25702	247	52	l	l	NOUN
fcis-25702	247	53	position	position	NOUN
fcis-25702	247	54	number	number	NOUN
fcis-25702	247	55	in	in	ADP
fcis-25702	247	56	i	i	PROPN
fcis-25702	247	57	-	-	PUNCT
fcis-25702	247	58	th	th	X
fcis-25702	247	59	array	array	NOUN
fcis-25702	247	60	,	,	PUNCT
fcis-25702	247	61	μ	μ	PROPN
fcis-25702	247	62	is	be	AUX
fcis-25702	247	63	average	average	ADJ
fcis-25702	247	64	of	of	ADP
fcis-25702	247	65	all	all	DET
fcis-25702	247	66	arrays	array	NOUN
fcis-25702	247	67	at	at	ADP
fcis-25702	247	68	position	position	NOUN
fcis-25702	247	69	l.	l.	PROPN
fcis-25702	247	70	σ	σ	PROPN
fcis-25702	247	71	denotes	denote	VERB
fcis-25702	247	72	the	the	DET
fcis-25702	247	73	overall	overall	ADJ
fcis-25702	247	74	difference	difference	NOUN
fcis-25702	247	75	among	among	ADP
fcis-25702	247	76	all	all	DET
fcis-25702	247	77	samples	sample	NOUN
fcis-25702	247	78	;	;	PUNCT
fcis-25702	247	79	a	a	DET
fcis-25702	247	80	large	large	ADJ
fcis-25702	247	81	variance	variance	NOUN
fcis-25702	247	82	indicates	indicate	VERB
fcis-25702	247	83	that	that	SCONJ
fcis-25702	247	84	the	the	DET
fcis-25702	247	85	differences	difference	NOUN
fcis-25702	247	86	in	in	ADP
fcis-25702	247	87	the	the	DET
fcis-25702	247	88	data	datum	NOUN
fcis-25702	247	89	are	be	AUX
fcis-25702	247	90	larger	large	ADJ
fcis-25702	247	91	and	and	CCONJ
fcis-25702	247	92	suggests	suggest	VERB
fcis-25702	247	93	that	that	SCONJ
fcis-25702	247	94	the	the	DET
fcis-25702	247	95	layer	layer	NOUN
fcis-25702	247	96	skipping	skip	VERB
fcis-25702	247	97	options	option	NOUN
fcis-25702	247	98	for	for	ADP
fcis-25702	247	99	different	different	ADJ
fcis-25702	247	100	samples	sample	NOUN
fcis-25702	247	101	vary	vary	VERB
fcis-25702	247	102	.	.	PUNCT
fcis-25702	248	1	we	we	PRON
fcis-25702	248	2	present	present	VERB
fcis-25702	248	3	the	the	DET
fcis-25702	248	4	variance	variance	NOUN
fcis-25702	248	5	for	for	ADP
fcis-25702	248	6	different	different	ADJ
fcis-25702	248	7	datasets	dataset	NOUN
fcis-25702	248	8	in	in	ADP
fcis-25702	248	9	table	table	NOUN
fcis-25702	248	10	3	3	NUM
fcis-25702	248	11	.	.	PUNCT
fcis-25702	249	1	as	as	SCONJ
fcis-25702	249	2	we	we	PRON
fcis-25702	249	3	can	can	AUX
fcis-25702	249	4	see	see	VERB
fcis-25702	249	5	,	,	PUNCT
fcis-25702	249	6	all	all	PRON
fcis-25702	249	7	have	have	VERB
fcis-25702	249	8	a	a	DET
fcis-25702	249	9	large	large	ADJ
fcis-25702	249	10	variance	variance	NOUN
fcis-25702	249	11	,	,	PUNCT
fcis-25702	249	12	indicating	indicate	VERB
fcis-25702	249	13	that	that	SCONJ
fcis-25702	249	14	the	the	DET
fcis-25702	249	15	skipped	skip	VERB
fcis-25702	249	16	layers	layer	NOUN
fcis-25702	249	17	for	for	ADP
fcis-25702	249	18	different	different	ADJ
fcis-25702	249	19	samples	sample	NOUN
fcis-25702	249	20	vary	vary	VERB
fcis-25702	249	21	significantly	significantly	ADV
fcis-25702	249	22	,	,	PUNCT
fcis-25702	249	23	which	which	PRON
fcis-25702	249	24	validates	validate	VERB
fcis-25702	249	25	the	the	DET
fcis-25702	249	26	importance	importance	NOUN
fcis-25702	249	27	of	of	ADP
fcis-25702	249	28	applying	apply	VERB
fcis-25702	249	29	dynamic	dynamic	ADJ
fcis-25702	249	30	layer	layer	NOUN
fcis-25702	249	31	skipping	skip	VERB
fcis-25702	249	32	.	.	PUNCT
fcis-25702	250	1	in	in	ADP
fcis-25702	250	2	contrast	contrast	NOUN
fcis-25702	250	3	,	,	PUNCT
fcis-25702	250	4	the	the	DET
fcis-25702	250	5	variance	variance	NOUN
fcis-25702	250	6	for	for	ADP
fcis-25702	250	7	static	static	ADJ
fcis-25702	250	8	layer	layer	NOUN
fcis-25702	250	9	skipping	skip	VERB
fcis-25702	250	10	is	be	AUX
fcis-25702	250	11	0	0	NUM
fcis-25702	250	12	,	,	PUNCT
fcis-25702	250	13	because	because	SCONJ
fcis-25702	250	14	the	the	DET
fcis-25702	250	15	same	same	ADJ
fcis-25702	250	16	settings	setting	NOUN
fcis-25702	250	17	are	be	AUX
fcis-25702	250	18	applied	apply	VERB
fcis-25702	250	19	to	to	ADP
fcis-25702	250	20	all	all	DET
fcis-25702	250	21	samples	sample	NOUN
fcis-25702	250	22	.	.	PUNCT
fcis-25702	251	1	table	table	NOUN
fcis-25702	251	2	3	3	NUM
fcis-25702	251	3	.	.	X
fcis-25702	251	4	variance	variance	NOUN
fcis-25702	251	5	of	of	ADP
fcis-25702	251	6	layer	layer	NOUN
fcis-25702	251	7	skipping	skip	VERB
fcis-25702	251	8	for	for	ADP
fcis-25702	251	9	different	different	ADJ
fcis-25702	251	10	datasets	dataset	NOUN
fcis-25702	251	11	dataset	dataset	NOUN
fcis-25702	251	12	variance	variance	NOUN
fcis-25702	251	13	rte	rte	NOUN
fcis-25702	251	14	0.35	0.35	NUM
fcis-25702	251	15	mrpc	mrpc	NOUN
fcis-25702	251	16	0.21	0.21	NUM
fcis-25702	251	17	cola	cola	NOUN
fcis-25702	251	18	0.43	0.43	NUM
fcis-25702	251	19	wmt2023	wmt2023	NOUN
fcis-25702	251	20	0.32	0.32	NUM
fcis-25702	251	21	iwslt2023	iwslt2023	PROPN
fcis-25702	251	22	0.41	0.41	NUM
fcis-25702	251	23	4.5.2	4.5.2	NUM
fcis-25702	251	24	.	.	PUNCT
fcis-25702	252	1	ablation	ablation	NOUN
fcis-25702	252	2	on	on	ADP
fcis-25702	252	3	layer	layer	NOUN
fcis-25702	252	4	skipping	skip	VERB
fcis-25702	252	5	pretraining	pretraine	VERB
fcis-25702	252	6	layer	layer	NOUN
fcis-25702	252	7	skip	skip	ADJ
fcis-25702	252	8	pretraining	pretraine	VERB
fcis-25702	252	9	enhances	enhance	NOUN
fcis-25702	252	10	the	the	DET
fcis-25702	252	11	model	model	NOUN
fcis-25702	252	12	’s	’s	PART
fcis-25702	252	13	ability	ability	NOUN
fcis-25702	252	14	to	to	PART
fcis-25702	252	15	infer	infer	VERB
fcis-25702	252	16	with	with	ADP
fcis-25702	252	17	fewer	few	ADJ
fcis-25702	252	18	layers	layer	NOUN
fcis-25702	252	19	,	,	PUNCT
fcis-25702	252	20	providing	provide	VERB
fcis-25702	252	21	a	a	DET
fcis-25702	252	22	good	good	ADJ
fcis-25702	252	23	initialization	initialization	NOUN
fcis-25702	252	24	for	for	SCONJ
fcis-25702	252	25	the	the	DET
fcis-25702	252	26	model	model	NOUN
fcis-25702	252	27	to	to	PART
fcis-25702	252	28	engage	engage	VERB
fcis-25702	252	29	in	in	ADP
fcis-25702	252	30	reinforcement	reinforcement	NOUN
fcis-25702	252	31	learning	learning	NOUN
fcis-25702	252	32	.	.	PUNCT
fcis-25702	253	1	as	as	ADP
fcis-25702	253	2	the	the	DET
fcis-25702	253	3	results	result	NOUN
fcis-25702	253	4	in	in	ADP
fcis-25702	253	5	table	table	NOUN
fcis-25702	253	6	1	1	NUM
fcis-25702	253	7	and	and	CCONJ
fcis-25702	253	8	table	table	NOUN
fcis-25702	253	9	2	2	NUM
fcis-25702	253	10	,	,	PUNCT
fcis-25702	253	11	yuan	yuan	NOUN
fcis-25702	253	12	et	et	NOUN
fcis-25702	253	13	al	al	PROPN
fcis-25702	253	14	.	.	PUNCT
fcis-25702	254	1	[	[	X
fcis-25702	254	2	34	34	NUM
fcis-25702	254	3	]	]	PUNCT
fcis-25702	254	4	shows	show	VERB
fcis-25702	254	5	better	well	ADJ
fcis-25702	254	6	performance	performance	NOUN
fcis-25702	254	7	than	than	ADP
fcis-25702	254	8	skip	skip	ADJ
fcis-25702	254	9	-	-	PUNCT
fcis-25702	254	10	decode	decode	NOUN
fcis-25702	254	11	[	[	X
fcis-25702	254	12	33	33	NUM
fcis-25702	254	13	]	]	PUNCT
fcis-25702	254	14	.	.	PUNCT
fcis-25702	255	1	this	this	PRON
fcis-25702	255	2	indicates	indicate	VERB
fcis-25702	255	3	that	that	SCONJ
fcis-25702	255	4	further	further	ADJ
fcis-25702	255	5	training	training	NOUN
fcis-25702	255	6	can	can	AUX
fcis-25702	255	7	significantly	significantly	ADV
fcis-25702	255	8	recover	recover	VERB
fcis-25702	255	9	the	the	DET
fcis-25702	255	10	model	model	NOUN
fcis-25702	255	11	’s	’s	PART
fcis-25702	255	12	performance	performance	NOUN
fcis-25702	255	13	lost	lose	VERB
fcis-25702	255	14	due	due	ADJ
fcis-25702	255	15	to	to	ADP
fcis-25702	255	16	skipping	skip	VERB
fcis-25702	255	17	layers	layer	NOUN
fcis-25702	255	18	.	.	PUNCT
fcis-25702	256	1	furthermore	furthermore	ADV
fcis-25702	256	2	,	,	PUNCT
fcis-25702	256	3	the	the	DET
fcis-25702	256	4	results	result	NOUN
fcis-25702	256	5	show	show	VERB
fcis-25702	256	6	that	that	SCONJ
fcis-25702	256	7	the	the	DET
fcis-25702	256	8	model	model	NOUN
fcis-25702	256	9	after	after	SCONJ
fcis-25702	256	10	layer	layer	NOUN
fcis-25702	256	11	skip	skip	NOUN
fcis-25702	256	12	pretraining	pretraine	VERB
fcis-25702	256	13	achieves	achieve	VERB
fcis-25702	256	14	even	even	ADV
fcis-25702	256	15	better	well	ADJ
fcis-25702	256	16	results	result	NOUN
fcis-25702	256	17	than	than	ADP
fcis-25702	256	18	yuan	yuan	NOUN
fcis-25702	256	19	et	et	NOUN
fcis-25702	256	20	al	al	PROPN
fcis-25702	256	21	.	.	PUNCT
fcis-25702	257	1	[	[	X
fcis-25702	257	2	34	34	NUM
fcis-25702	257	3	]	]	PUNCT
fcis-25702	257	4	,	,	PUNCT
fcis-25702	257	5	validating	validate	VERB
fcis-25702	257	6	the	the	DET
fcis-25702	257	7	effectiveness	effectiveness	NOUN
fcis-25702	257	8	of	of	ADP
fcis-25702	257	9	the	the	DET
fcis-25702	257	10	consistency	consistency	NOUN
fcis-25702	257	11	loss	loss	NOUN
fcis-25702	257	12	and	and	CCONJ
fcis-25702	257	13	distillation	distillation	NOUN
fcis-25702	257	14	loss	loss	NOUN
fcis-25702	257	15	from	from	ADP
fcis-25702	257	16	the	the	DET
fcis-25702	257	17	teacher	teacher	NOUN
fcis-25702	257	18	model	model	NOUN
fcis-25702	257	19	.	.	PUNCT
fcis-25702	258	1	4.5.3	4.5.3	X
fcis-25702	258	2	.	.	PUNCT
fcis-25702	258	3	ablation	ablation	NOUN
fcis-25702	258	4	on	on	ADP
fcis-25702	258	5	sampling	sample	VERB
fcis-25702	258	6	strategy	strategy	NOUN
fcis-25702	258	7	encouraging	encourage	VERB
fcis-25702	258	8	the	the	DET
fcis-25702	258	9	model	model	NOUN
fcis-25702	258	10	to	to	PART
fcis-25702	258	11	skip	skip	VERB
fcis-25702	258	12	more	more	ADV
fcis-25702	258	13	confident	confident	ADJ
fcis-25702	258	14	layers	layer	NOUN
fcis-25702	258	15	(	(	PUNCT
fcis-25702	258	16	measuring	measure	VERB
fcis-25702	258	17	their	their	PRON
fcis-25702	258	18	similarities	similarity	NOUN
fcis-25702	258	19	with	with	ADP
fcis-25702	258	20	previous	previous	ADJ
fcis-25702	258	21	layers	layer	NOUN
fcis-25702	258	22	)	)	PUNCT
fcis-25702	258	23	at	at	ADP
fcis-25702	258	24	initial	initial	ADJ
fcis-25702	258	25	steps	step	NOUN
fcis-25702	258	26	can	can	AUX
fcis-25702	258	27	help	help	VERB
fcis-25702	258	28	the	the	DET
fcis-25702	258	29	model	model	NOUN
fcis-25702	258	30	find	find	VERB
fcis-25702	258	31	a	a	DET
fcis-25702	258	32	better	well	ADJ
fcis-25702	258	33	policy	policy	NOUN
fcis-25702	258	34	early	early	ADV
fcis-25702	258	35	on	on	ADV
fcis-25702	258	36	and	and	CCONJ
fcis-25702	258	37	converge	converge	VERB
fcis-25702	258	38	to	to	ADP
fcis-25702	258	39	a	a	DET
fcis-25702	258	40	better	well	ADJ
fcis-25702	258	41	overall	overall	ADJ
fcis-25702	258	42	policy	policy	NOUN
fcis-25702	258	43	.	.	PUNCT
fcis-25702	259	1	as	as	SCONJ
fcis-25702	259	2	shown	show	VERB
fcis-25702	259	3	in	in	ADP
fcis-25702	259	4	figure	figure	NOUN
fcis-25702	259	5	4	4	NUM
fcis-25702	259	6	,	,	PUNCT
fcis-25702	259	7	we	we	PRON
fcis-25702	259	8	7	7	NUM
fcis-25702	259	9	can	can	AUX
fcis-25702	259	10	see	see	VERB
fcis-25702	259	11	that	that	SCONJ
fcis-25702	259	12	without	without	ADP
fcis-25702	259	13	the	the	DET
fcis-25702	259	14	sampling	sample	VERB
fcis-25702	259	15	strategy	strategy	NOUN
fcis-25702	259	16	,	,	PUNCT
fcis-25702	259	17	the	the	DET
fcis-25702	259	18	loss	loss	NOUN
fcis-25702	259	19	at	at	ADP
fcis-25702	259	20	initial	initial	ADJ
fcis-25702	259	21	stages	stage	NOUN
fcis-25702	259	22	is	be	AUX
fcis-25702	259	23	very	very	ADV
fcis-25702	259	24	unstable	unstable	ADJ
fcis-25702	259	25	and	and	CCONJ
fcis-25702	259	26	may	may	AUX
fcis-25702	259	27	even	even	ADV
fcis-25702	259	28	increase	increase	VERB
fcis-25702	259	29	significantly	significantly	ADV
fcis-25702	259	30	.	.	PUNCT
fcis-25702	260	1	this	this	PRON
fcis-25702	260	2	occurs	occur	VERB
fcis-25702	260	3	because	because	SCONJ
fcis-25702	260	4	the	the	DET
fcis-25702	260	5	model	model	NOUN
fcis-25702	260	6	may	may	AUX
fcis-25702	260	7	skip	skip	VERB
fcis-25702	260	8	some	some	DET
fcis-25702	260	9	vital	vital	ADJ
fcis-25702	260	10	layers	layer	NOUN
fcis-25702	260	11	,	,	PUNCT
fcis-25702	260	12	and	and	CCONJ
fcis-25702	260	13	each	each	DET
fcis-25702	260	14	skipping	skip	VERB
fcis-25702	260	15	decision	decision	NOUN
fcis-25702	260	16	does	do	AUX
fcis-25702	260	17	not	not	PART
fcis-25702	260	18	directly	directly	ADV
fcis-25702	260	19	contribute	contribute	VERB
fcis-25702	260	20	to	to	ADP
fcis-25702	260	21	the	the	DET
fcis-25702	260	22	final	final	ADJ
fcis-25702	260	23	token	token	ADJ
fcis-25702	260	24	prediction	prediction	NOUN
fcis-25702	260	25	,	,	PUNCT
fcis-25702	260	26	leading	lead	VERB
fcis-25702	260	27	the	the	DET
fcis-25702	260	28	model	model	NOUN
fcis-25702	260	29	to	to	PART
fcis-25702	260	30	potentially	potentially	ADV
fcis-25702	260	31	learn	learn	VERB
fcis-25702	260	32	a	a	DET
fcis-25702	260	33	false	false	ADJ
fcis-25702	260	34	policy	policy	NOUN
fcis-25702	260	35	.	.	PUNCT
fcis-25702	261	1	thus	thus	ADV
fcis-25702	261	2	,	,	PUNCT
fcis-25702	261	3	the	the	DET
fcis-25702	261	4	sampling	sample	VERB
fcis-25702	261	5	strategy	strategy	NOUN
fcis-25702	261	6	helps	help	VERB
fcis-25702	261	7	make	make	VERB
fcis-25702	261	8	the	the	DET
fcis-25702	261	9	training	training	NOUN
fcis-25702	261	10	of	of	ADP
fcis-25702	261	11	rl	rl	X
fcis-25702	261	12	more	more	ADV
fcis-25702	261	13	stable	stable	ADJ
fcis-25702	261	14	.	.	PUNCT
fcis-25702	262	1	since	since	SCONJ
fcis-25702	262	2	the	the	DET
fcis-25702	262	3	total	total	ADJ
fcis-25702	262	4	number	number	NOUN
fcis-25702	262	5	of	of	ADP
fcis-25702	262	6	sampling	sample	VERB
fcis-25702	262	7	trajectories	trajectory	NOUN
fcis-25702	262	8	is	be	AUX
fcis-25702	262	9	2	2	NUM
fcis-25702	262	10	,	,	PUNCT
fcis-25702	262	11	adding	add	VERB
fcis-25702	262	12	a	a	DET
fcis-25702	262	13	sampling	sample	VERB
fcis-25702	262	14	strategy	strategy	NOUN
fcis-25702	262	15	can	can	AUX
fcis-25702	262	16	significantly	significantly	ADV
fcis-25702	262	17	reduce	reduce	VERB
fcis-25702	262	18	the	the	DET
fcis-25702	262	19	number	number	NOUN
fcis-25702	262	20	of	of	ADP
fcis-25702	262	21	low	low	ADJ
fcis-25702	262	22	-	-	PUNCT
fcis-25702	262	23	quality	quality	NOUN
fcis-25702	262	24	trajectories	trajectory	NOUN
fcis-25702	262	25	,	,	PUNCT
fcis-25702	262	26	ensuring	ensure	VERB
fcis-25702	262	27	the	the	DET
fcis-25702	262	28	model	model	NOUN
fcis-25702	262	29	remains	remain	VERB
fcis-25702	262	30	at	at	ADP
fcis-25702	262	31	a	a	DET
fcis-25702	262	32	relatively	relatively	ADV
fcis-25702	262	33	good	good	ADJ
fcis-25702	262	34	policy	policy	NOUN
fcis-25702	262	35	.	.	PUNCT
fcis-25702	263	1	we	we	PRON
fcis-25702	263	2	can	can	AUX
fcis-25702	263	3	also	also	ADV
fcis-25702	263	4	see	see	VERB
fcis-25702	263	5	that	that	SCONJ
fcis-25702	263	6	the	the	DET
fcis-25702	263	7	final	final	ADJ
fcis-25702	263	8	loss	loss	NOUN
fcis-25702	263	9	is	be	AUX
fcis-25702	263	10	lower	low	ADJ
fcis-25702	263	11	,	,	PUNCT
fcis-25702	263	12	indicating	indicate	VERB
fcis-25702	263	13	the	the	DET
fcis-25702	263	14	effectiveness	effectiveness	NOUN
fcis-25702	263	15	of	of	ADP
fcis-25702	263	16	the	the	DET
fcis-25702	263	17	proposed	propose	VERB
fcis-25702	263	18	sampling	sampling	NOUN
fcis-25702	263	19	strategy	strategy	NOUN
fcis-25702	263	20	.	.	PUNCT
fcis-25702	264	1	figure	figure	VERB
fcis-25702	264	2	3	3	NUM
fcis-25702	264	3	.	.	PUNCT
fcis-25702	264	4	f1	f1	NOUN
fcis-25702	264	5	-	-	PUNCT
fcis-25702	264	6	score	score	NOUN
fcis-25702	264	7	of	of	ADP
fcis-25702	264	8	llama-7b	llama-7b	NOUN
fcis-25702	264	9	-	-	ADJ
fcis-25702	264	10	chat	chat	ADJ
fcis-25702	264	11	model	model	NOUN
fcis-25702	264	12	on	on	ADP
fcis-25702	264	13	qnli	qnli	PROPN
fcis-25702	264	14	dataset	dataset	VERB
fcis-25702	264	15	with	with	ADP
fcis-25702	264	16	different	different	ADJ
fcis-25702	264	17	layer	layer	NOUN
fcis-25702	264	18	skipping	skip	VERB
fcis-25702	264	19	ratios	ratio	NOUN
fcis-25702	264	20	.	.	PUNCT
fcis-25702	265	1	a	a	X
fcis-25702	265	2	)	)	PUNCT
fcis-25702	265	3	loss	loss	NOUN
fcis-25702	265	4	without	without	ADP
fcis-25702	265	5	sampling	sample	VERB
fcis-25702	265	6	strategy	strategy	NOUN
fcis-25702	265	7	b	b	NOUN
fcis-25702	265	8	)	)	PUNCT
fcis-25702	265	9	loss	loss	NOUN
fcis-25702	265	10	with	with	ADP
fcis-25702	265	11	sampling	sample	VERB
fcis-25702	265	12	strategy	strategy	NOUN
fcis-25702	265	13	figure	figure	NOUN
fcis-25702	265	14	4	4	NUM
fcis-25702	265	15	.	.	PUNCT
fcis-25702	266	1	the	the	DET
fcis-25702	266	2	variation	variation	NOUN
fcis-25702	266	3	diagram	diagram	NOUN
fcis-25702	266	4	of	of	ADP
fcis-25702	266	5	rl	rl	ADP
fcis-25702	266	6	loss	loss	NOUN
fcis-25702	266	7	with	with	ADP
fcis-25702	266	8	the	the	DET
fcis-25702	266	9	increase	increase	NOUN
fcis-25702	266	10	of	of	ADP
fcis-25702	266	11	training	training	NOUN
fcis-25702	266	12	steps	step	NOUN
fcis-25702	266	13	8	8	NUM
fcis-25702	266	14	5	5	NUM
fcis-25702	266	15	.	.	PUNCT
fcis-25702	266	16	conclusion	conclusion	NOUN
fcis-25702	266	17	in	in	ADP
fcis-25702	266	18	this	this	DET
fcis-25702	266	19	work	work	NOUN
fcis-25702	266	20	,	,	PUNCT
fcis-25702	266	21	we	we	PRON
fcis-25702	266	22	present	present	VERB
fcis-25702	266	23	a	a	DET
fcis-25702	266	24	novel	novel	ADJ
fcis-25702	266	25	approach	approach	NOUN
fcis-25702	266	26	for	for	ADP
fcis-25702	266	27	speeding	speed	VERB
fcis-25702	266	28	up	up	ADP
fcis-25702	266	29	large	large	ADJ
fcis-25702	266	30	language	language	NOUN
fcis-25702	266	31	model	model	NOUN
fcis-25702	266	32	(	(	PUNCT
fcis-25702	266	33	llm	llm	NOUN
fcis-25702	266	34	)	)	PUNCT
fcis-25702	266	35	inference	inference	NOUN
fcis-25702	266	36	through	through	ADP
fcis-25702	266	37	the	the	DET
fcis-25702	266	38	application	application	NOUN
fcis-25702	266	39	of	of	ADP
fcis-25702	266	40	dynamic	dynamic	ADJ
fcis-25702	266	41	inference	inference	NOUN
fcis-25702	266	42	methods	method	NOUN
fcis-25702	266	43	.	.	PUNCT
fcis-25702	267	1	to	to	ADP
fcis-25702	267	2	the	the	DET
fcis-25702	267	3	best	good	ADJ
fcis-25702	267	4	of	of	ADP
fcis-25702	267	5	our	our	PRON
fcis-25702	267	6	knowledge	knowledge	NOUN
fcis-25702	267	7	,	,	PUNCT
fcis-25702	267	8	we	we	PRON
fcis-25702	267	9	are	be	AUX
fcis-25702	267	10	the	the	DET
fcis-25702	267	11	first	first	ADJ
fcis-25702	267	12	to	to	PART
fcis-25702	267	13	explore	explore	VERB
fcis-25702	267	14	the	the	DET
fcis-25702	267	15	dynamic	dynamic	ADJ
fcis-25702	267	16	layer	layer	NOUN
fcis-25702	267	17	skipping	skip	VERB
fcis-25702	267	18	technique	technique	NOUN
fcis-25702	267	19	in	in	ADP
fcis-25702	267	20	the	the	DET
fcis-25702	267	21	context	context	NOUN
fcis-25702	267	22	of	of	ADP
fcis-25702	267	23	llms	llm	NOUN
fcis-25702	267	24	.	.	PUNCT
fcis-25702	268	1	our	our	PRON
fcis-25702	268	2	contributions	contribution	NOUN
fcis-25702	268	3	include	include	VERB
fcis-25702	268	4	the	the	DET
fcis-25702	268	5	introduction	introduction	NOUN
fcis-25702	268	6	of	of	ADP
fcis-25702	268	7	a	a	DET
fcis-25702	268	8	framework	framework	NOUN
fcis-25702	268	9	that	that	PRON
fcis-25702	268	10	utilizes	utilize	VERB
fcis-25702	268	11	an	an	DET
fcis-25702	268	12	adapter	adapter	NOUN
fcis-25702	268	13	layer	layer	NOUN
fcis-25702	268	14	to	to	PART
fcis-25702	268	15	dynamically	dynamically	ADV
fcis-25702	268	16	skip	skip	VERB
fcis-25702	268	17	llm	llm	ADJ
fcis-25702	268	18	layers	layer	NOUN
fcis-25702	268	19	without	without	ADP
fcis-25702	268	20	introducing	introduce	VERB
fcis-25702	268	21	significant	significant	ADJ
fcis-25702	268	22	computational	computational	ADJ
fcis-25702	268	23	costs	cost	NOUN
fcis-25702	268	24	.	.	PUNCT
fcis-25702	269	1	furthermore	furthermore	ADV
fcis-25702	269	2	,	,	PUNCT
fcis-25702	269	3	we	we	PRON
fcis-25702	269	4	optimize	optimize	VERB
fcis-25702	269	5	our	our	PRON
fcis-25702	269	6	framework	framework	NOUN
fcis-25702	269	7	using	use	VERB
fcis-25702	269	8	reinforcement	reinforcement	NOUN
fcis-25702	269	9	learning	learning	NOUN
fcis-25702	269	10	,	,	PUNCT
fcis-25702	269	11	incorporating	incorporate	VERB
fcis-25702	269	12	sampling	sampling	NOUN
fcis-25702	269	13	strategies	strategy	NOUN
fcis-25702	269	14	to	to	PART
fcis-25702	269	15	enhance	enhance	VERB
fcis-25702	269	16	the	the	DET
fcis-25702	269	17	training	training	NOUN
fcis-25702	269	18	process	process	NOUN
fcis-25702	269	19	.	.	PUNCT
fcis-25702	270	1	experimental	experimental	ADJ
fcis-25702	270	2	results	result	NOUN
fcis-25702	270	3	demonstrate	demonstrate	VERB
fcis-25702	270	4	that	that	SCONJ
fcis-25702	270	5	our	our	PRON
fcis-25702	270	6	method	method	NOUN
fcis-25702	270	7	achieves	achieve	VERB
fcis-25702	270	8	state	state	NOUN
fcis-25702	270	9	-	-	PUNCT
fcis-25702	270	10	of	of	ADP
fcis-25702	270	11	-	-	PUNCT
fcis-25702	270	12	the	the	DET
fcis-25702	270	13	-	-	PUNCT
fcis-25702	270	14	art	art	NOUN
fcis-25702	270	15	performance	performance	NOUN
fcis-25702	270	16	in	in	ADP
fcis-25702	270	17	terms	term	NOUN
fcis-25702	270	18	of	of	ADP
fcis-25702	270	19	llm	llm	PROPN
fcis-25702	270	20	inference	inference	NOUN
fcis-25702	270	21	speedup	speedup	NOUN
fcis-25702	270	22	.	.	PUNCT
fcis-25702	271	1	additional	additional	ADJ
fcis-25702	271	2	ablation	ablation	NOUN
fcis-25702	271	3	studies	study	NOUN
fcis-25702	271	4	and	and	CCONJ
fcis-25702	271	5	intrinsic	intrinsic	ADJ
fcis-25702	271	6	evaluations	evaluation	NOUN
fcis-25702	271	7	further	far	ADV
fcis-25702	271	8	validate	validate	VERB
fcis-25702	271	9	the	the	DET
fcis-25702	271	10	effectiveness	effectiveness	NOUN
fcis-25702	271	11	of	of	ADP
fcis-25702	271	12	our	our	PRON
fcis-25702	271	13	proposed	propose	VERB
fcis-25702	271	14	technique	technique	NOUN
fcis-25702	271	15	.	.	PUNCT
fcis-25702	272	1	references	reference	NOUN
fcis-25702	272	2	[	[	X
fcis-25702	272	3	1	1	NUM
fcis-25702	272	4	]	]	PUNCT
fcis-25702	272	5	josh	josh	PROPN
fcis-25702	272	6	achiam	achiam	PROPN
fcis-25702	272	7	,	,	PUNCT
fcis-25702	272	8	steven	steven	PROPN
fcis-25702	272	9	adler	adler	PROPN
fcis-25702	272	10	,	,	PUNCT
fcis-25702	272	11	sandhini	sandhini	PROPN
fcis-25702	272	12	agarwal	agarwal	PROPN
fcis-25702	272	13	,	,	PUNCT
fcis-25702	272	14	lama	lama	PROPN
fcis-25702	272	15	ahmad	ahmad	PROPN
fcis-25702	272	16	,	,	PUNCT
fcis-25702	272	17	ilge	ilge	PROPN
fcis-25702	272	18	akkaya	akkaya	PROPN
fcis-25702	272	19	,	,	PUNCT
fcis-25702	272	20	florencia	florencia	PROPN
fcis-25702	272	21	leoni	leoni	PROPN
fcis-25702	272	22	aleman	aleman	PROPN
fcis-25702	272	23	,	,	PUNCT
fcis-25702	272	24	diogo	diogo	PROPN
fcis-25702	272	25	almeida	almeida	PROPN
fcis-25702	272	26	,	,	PUNCT
fcis-25702	272	27	janko	janko	PROPN
fcis-25702	272	28	altenschmidt	altenschmidt	PROPN
fcis-25702	272	29	,	,	PUNCT
fcis-25702	272	30	sam	sam	PROPN
fcis-25702	272	31	altman	altman	PROPN
fcis-25702	272	32	,	,	PUNCT
fcis-25702	272	33	shyamal	shyamal	ADJ
fcis-25702	272	34	anadkat	anadkat	PROPN
fcis-25702	272	35	,	,	PUNCT
fcis-25702	272	36	et	et	PROPN
fcis-25702	272	37	al	al	PROPN
fcis-25702	272	38	.	.	PROPN
fcis-25702	272	39	2023	2023	NUM
fcis-25702	272	40	.	.	PUNCT
fcis-25702	273	1	gpt-4	gpt-4	VERB
fcis-25702	273	2	technical	technical	ADJ
fcis-25702	273	3	report	report	NOUN
fcis-25702	273	4	.	.	PUNCT
fcis-25702	274	1	arxiv	arxiv	PROPN
fcis-25702	274	2	preprint	preprint	VERB
fcis-25702	274	3	arxiv:2303.08774	arxiv:2303.08774	NOUN
fcis-25702	274	4	.	.	PUNCT
fcis-25702	275	1	[	[	X
fcis-25702	275	2	2	2	NUM
fcis-25702	275	3	]	]	PUNCT
fcis-25702	275	4	tom	tom	PROPN
fcis-25702	275	5	brown	brown	PROPN
fcis-25702	275	6	,	,	PUNCT
fcis-25702	275	7	benjamin	benjamin	PROPN
fcis-25702	275	8	mann	mann	PROPN
fcis-25702	275	9	,	,	PUNCT
fcis-25702	275	10	nick	nick	PROPN
fcis-25702	275	11	ryder	ryder	PROPN
fcis-25702	275	12	,	,	PUNCT
fcis-25702	275	13	melanie	melanie	PROPN
fcis-25702	275	14	subbiah	subbiah	PROPN
fcis-25702	275	15	,	,	PUNCT
fcis-25702	275	16	jared	jared	PROPN
fcis-25702	275	17	d	d	PROPN
fcis-25702	275	18	kaplan	kaplan	PROPN
fcis-25702	275	19	,	,	PUNCT
fcis-25702	275	20	prafulla	prafulla	NOUN
fcis-25702	275	21	dhariwal	dhariwal	NOUN
fcis-25702	275	22	,	,	PUNCT
fcis-25702	275	23	arvind	arvind	PROPN
fcis-25702	275	24	neelakantan	neelakantan	PROPN
fcis-25702	275	25	,	,	PUNCT
fcis-25702	275	26	pranav	pranav	PROPN
fcis-25702	275	27	shyam	shyam	PROPN
fcis-25702	275	28	,	,	PUNCT
fcis-25702	275	29	girish	girish	PROPN
fcis-25702	275	30	sastry	sastry	PROPN
fcis-25702	275	31	,	,	PUNCT
fcis-25702	275	32	amanda	amanda	PROPN
fcis-25702	275	33	askell	askell	PROPN
fcis-25702	275	34	,	,	PUNCT
fcis-25702	275	35	et	et	PROPN
fcis-25702	275	36	al	al	PROPN
fcis-25702	275	37	.	.	PROPN
fcis-25702	275	38	2020	2020	NUM
fcis-25702	275	39	.	.	PUNCT
fcis-25702	276	1	language	language	NOUN
fcis-25702	276	2	models	model	NOUN
fcis-25702	276	3	are	be	AUX
fcis-25702	276	4	few	few	ADJ
fcis-25702	276	5	-	-	PUNCT
fcis-25702	276	6	shot	shot	NOUN
fcis-25702	276	7	learners	learner	NOUN
fcis-25702	276	8	.	.	PUNCT
fcis-25702	277	1	advances	advance	NOUN
fcis-25702	277	2	in	in	ADP
fcis-25702	277	3	neural	neural	ADJ
fcis-25702	277	4	information	information	NOUN
fcis-25702	277	5	processing	processing	NOUN
fcis-25702	277	6	systems	system	NOUN
fcis-25702	277	7	,	,	PUNCT
fcis-25702	277	8	33:1877	33:1877	NUM
fcis-25702	277	9	-	-	SYM
fcis-25702	277	10	1901	1901	NUM
fcis-25702	277	11	.	.	PUNCT
fcis-25702	278	1	[	[	X
fcis-25702	278	2	3	3	NUM
fcis-25702	278	3	]	]	X
fcis-25702	278	4	jason	jason	PROPN
fcis-25702	278	5	wei	wei	PROPN
fcis-25702	278	6	,	,	PUNCT
fcis-25702	278	7	yi	yi	PROPN
fcis-25702	278	8	tay	tay	PROPN
fcis-25702	278	9	,	,	PUNCT
fcis-25702	278	10	rishi	rishi	PROPN
fcis-25702	278	11	bommasani	bommasani	NOUN
fcis-25702	278	12	,	,	PUNCT
fcis-25702	278	13	colin	colin	PROPN
fcis-25702	278	14	raffel	raffel	PROPN
fcis-25702	278	15	,	,	PUNCT
fcis-25702	278	16	barret	barret	PROPN
fcis-25702	278	17	zoph	zoph	NOUN
fcis-25702	278	18	,	,	PUNCT
fcis-25702	278	19	sebastian	sebastian	PROPN
fcis-25702	278	20	borgeaud	borgeaud	PROPN
fcis-25702	278	21	,	,	PUNCT
fcis-25702	278	22	dani	dani	PROPN
fcis-25702	278	23	yogatama	yogatama	PROPN
fcis-25702	278	24	,	,	PUNCT
fcis-25702	278	25	maarten	maarten	PROPN
fcis-25702	278	26	bosma	bosma	PROPN
fcis-25702	278	27	,	,	PUNCT
fcis-25702	278	28	denny	denny	PROPN
fcis-25702	278	29	zhou	zhou	PROPN
fcis-25702	278	30	,	,	PUNCT
fcis-25702	278	31	donald	donald	PROPN
fcis-25702	278	32	metzler	metzler	PROPN
fcis-25702	278	33	,	,	PUNCT
fcis-25702	278	34	et	et	PROPN
fcis-25702	278	35	al	al	PROPN
fcis-25702	278	36	.	.	PROPN
fcis-25702	278	37	2022	2022	NUM
fcis-25702	278	38	.	.	PUNCT
fcis-25702	279	1	emergent	emergent	ADJ
fcis-25702	279	2	abilities	ability	NOUN
fcis-25702	279	3	of	of	ADP
fcis-25702	279	4	large	large	ADJ
fcis-25702	279	5	language	language	NOUN
fcis-25702	279	6	models	model	NOUN
fcis-25702	279	7	.	.	PUNCT
fcis-25702	280	1	arxiv	arxiv	PROPN
fcis-25702	280	2	preprint	preprint	VERB
fcis-25702	280	3	arxiv:2206.07682	arxiv:2206.07682	ADV
fcis-25702	280	4	.	.	PUNCT
fcis-25702	281	1	[	[	X
fcis-25702	281	2	4	4	X
fcis-25702	281	3	]	]	X
fcis-25702	281	4	jason	jason	PROPN
fcis-25702	281	5	wei	wei	PROPN
fcis-25702	281	6	,	,	PUNCT
fcis-25702	281	7	xuezhi	xuezhi	PROPN
fcis-25702	281	8	wang	wang	PROPN
fcis-25702	281	9	,	,	PUNCT
fcis-25702	281	10	dale	dale	PROPN
fcis-25702	281	11	schuurmans	schuurmans	PROPN
fcis-25702	281	12	,	,	PUNCT
fcis-25702	281	13	maarten	maarten	PROPN
fcis-25702	281	14	bosma	bosma	PROPN
fcis-25702	281	15	,	,	PUNCT
fcis-25702	281	16	fei	fei	PROPN
fcis-25702	281	17	xia	xia	PROPN
fcis-25702	281	18	,	,	PUNCT
fcis-25702	281	19	ed	ed	PROPN
fcis-25702	281	20	chi	chi	PROPN
fcis-25702	281	21	,	,	PUNCT
fcis-25702	281	22	quoc	quoc	VERB
fcis-25702	281	23	v	v	NUM
fcis-25702	281	24	le	le	X
fcis-25702	281	25	,	,	PUNCT
fcis-25702	281	26	denny	denny	PROPN
fcis-25702	281	27	zhou	zhou	PROPN
fcis-25702	281	28	,	,	PUNCT
fcis-25702	281	29	et	et	PROPN
fcis-25702	281	30	al	al	PROPN
fcis-25702	281	31	.	.	PROPN
fcis-25702	281	32	2022	2022	NUM
fcis-25702	281	33	.	.	PUNCT
fcis-25702	282	1	chainof	chainof	NOUN
fcis-25702	282	2	-	-	PUNCT
fcis-25702	282	3	thought	thought	NOUN
fcis-25702	282	4	prompting	prompt	VERB
fcis-25702	282	5	elicits	elicit	NOUN
fcis-25702	282	6	rea	rea	VERB
fcis-25702	282	7	soning	soning	NOUN
fcis-25702	282	8	in	in	ADP
fcis-25702	282	9	large	large	ADJ
fcis-25702	282	10	language	language	NOUN
fcis-25702	282	11	models	model	NOUN
fcis-25702	282	12	.	.	PUNCT
fcis-25702	283	1	advances	advance	NOUN
fcis-25702	283	2	in	in	ADP
fcis-25702	283	3	neural	neural	ADJ
fcis-25702	283	4	information	information	NOUN
fcis-25702	283	5	processing	processing	NOUN
fcis-25702	283	6	systems	system	NOUN
fcis-25702	283	7	,	,	PUNCT
fcis-25702	283	8	35:24824	35:24824	NUM
fcis-25702	283	9	-	-	NOUN
fcis-25702	283	10	24837	24837	NUM
fcis-25702	283	11	.	.	PUNCT
fcis-25702	284	1	[	[	X
fcis-25702	284	2	5	5	NUM
fcis-25702	284	3	]	]	X
fcis-25702	284	4	hyung	hyung	PROPN
fcis-25702	284	5	won	win	VERB
fcis-25702	284	6	chung	chung	PROPN
fcis-25702	284	7	,	,	PUNCT
fcis-25702	284	8	le	le	X
fcis-25702	284	9	hou	hou	PROPN
fcis-25702	284	10	,	,	PUNCT
fcis-25702	284	11	shayne	shayne	PROPN
fcis-25702	284	12	longpre	longpre	PROPN
fcis-25702	284	13	,	,	PUNCT
fcis-25702	284	14	barretzoph	barretzoph	PROPN
fcis-25702	284	15	,	,	PUNCT
fcis-25702	284	16	yi	yi	PROPN
fcis-25702	284	17	tay	tay	PROPN
fcis-25702	284	18	,	,	PUNCT
fcis-25702	284	19	william	william	PROPN
fcis-25702	284	20	fedus	fedus	PROPN
fcis-25702	284	21	,	,	PUNCT
fcis-25702	284	22	yunxuan	yunxuan	PROPN
fcis-25702	284	23	li	li	PROPN
fcis-25702	284	24	,	,	PUNCT
fcis-25702	284	25	xuezhi	xuezhi	PROPN
fcis-25702	284	26	wang	wang	PROPN
fcis-25702	284	27	,	,	PUNCT
fcis-25702	284	28	mostafa	mostafa	PROPN
fcis-25702	284	29	dehghani	dehghani	PROPN
fcis-25702	284	30	,	,	PUNCT
fcis-25702	284	31	siddhartha	siddhartha	PROPN
fcis-25702	284	32	brahma	brahma	NOUN
fcis-25702	284	33	,	,	PUNCT
fcis-25702	284	34	et	et	PROPN
fcis-25702	284	35	al	al	PROPN
fcis-25702	284	36	.	.	PROPN
fcis-25702	284	37	2022	2022	NUM
fcis-25702	284	38	.	.	PUNCT
fcis-25702	285	1	scaling	scale	VERB
fcis-25702	285	2	instructionfinetuned	instructionfinetune	VERB
fcis-25702	285	3	language	language	NOUN
fcis-25702	285	4	models	model	NOUN
fcis-25702	285	5	.	.	PUNCT
fcis-25702	286	1	arxiv	arxiv	PROPN
fcis-25702	286	2	preprint	preprint	NOUN
fcis-25702	286	3	arxiv:2210.11416	arxiv:2210.11416	PROPN
fcis-25702	286	4	.	.	PUNCT
fcis-25702	287	1	[	[	X
fcis-25702	287	2	6	6	NUM
fcis-25702	287	3	]	]	SYM
fcis-25702	287	4	xuanli	xuanli	NOUN
fcis-25702	288	1	he	he	PRON
fcis-25702	288	2	,	,	PUNCT
fcis-25702	288	3	yuxiang	yuxiang	PROPN
fcis-25702	288	4	wu	wu	PROPN
fcis-25702	288	5	,	,	PUNCT
fcis-25702	288	6	oana	oana	PROPN
fcis-25702	288	7	-	-	PROPN
fcis-25702	288	8	maria	maria	PROPN
fcis-25702	288	9	camburu	camburu	PROPN
fcis-25702	288	10	,	,	PUNCT
fcis-25702	288	11	pasquale	pasquale	NOUN
fcis-25702	288	12	minervini	minervini	NOUN
fcis-25702	288	13	,	,	PUNCT
fcis-25702	288	14	and	and	CCONJ
fcis-25702	288	15	pontus	pontus	PROPN
fcis-25702	288	16	stenetorp	stenetorp	PROPN
fcis-25702	288	17	.	.	PROPN
fcis-25702	288	18	2023	2023	NUM
fcis-25702	288	19	.	.	PUNCT
fcis-25702	289	1	using	use	VERB
fcis-25702	289	2	natural	natural	ADJ
fcis-25702	289	3	language	language	NOUN
fcis-25702	289	4	explanations	explanation	NOUN
fcis-25702	289	5	to	to	PART
fcis-25702	289	6	improve	improve	VERB
fcis-25702	289	7	robust	robust	ADJ
fcis-25702	289	8	ness	ness	NOUN
fcis-25702	289	9	of	of	ADP
fcis-25702	289	10	in	in	ADP
fcis-25702	289	11	-	-	PUNCT
fcis-25702	289	12	context	context	NOUN
fcis-25702	289	13	learning	learning	NOUN
fcis-25702	289	14	for	for	ADP
fcis-25702	289	15	natural	natural	ADJ
fcis-25702	289	16	language	language	NOUN
fcis-25702	289	17	infer	infer	NOUN
fcis-25702	289	18	ence	ence	NOUN
fcis-25702	289	19	.	.	PUNCT
fcis-25702	290	1	arxiv	arxiv	PROPN
fcis-25702	290	2	preprint	preprint	VERB
fcis-25702	290	3	arxiv:2311.07556	arxiv:2311.07556	NOUN
fcis-25702	290	4	.	.	PUNCT
fcis-25702	291	1	[	[	X
fcis-25702	291	2	7	7	X
fcis-25702	291	3	]	]	X
fcis-25702	291	4	ankur	ankur	PROPN
fcis-25702	291	5	goswami	goswami	PROPN
fcis-25702	291	6	,	,	PUNCT
fcis-25702	291	7	akshata	akshata	NOUN
fcis-25702	291	8	bhat	bhat	PROPN
fcis-25702	291	9	,	,	PUNCT
fcis-25702	291	10	hadar	hadar	PROPN
fcis-25702	291	11	ohana	ohana	PROPN
fcis-25702	291	12	,	,	PUNCT
fcis-25702	291	13	and	and	CCONJ
fcis-25702	291	14	theodoros	theodoros	NOUN
fcis-25702	291	15	rekatsinas	rekatsina	NOUN
fcis-25702	291	16	.	.	PUNCT
fcis-25702	292	1	2020	2020	NUM
fcis-25702	292	2	.	.	PUNCT
fcis-25702	293	1	unsupervised	unsupervised	ADJ
fcis-25702	293	2	relation	relation	NOUN
fcis-25702	293	3	extraction	extraction	NOUN
fcis-25702	293	4	from	from	ADP
fcis-25702	293	5	language	language	NOUN
fcis-25702	293	6	models	model	NOUN
fcis-25702	293	7	using	use	VERB
fcis-25702	293	8	constrained	constrained	ADJ
fcis-25702	293	9	cloze	cloze	NOUN
fcis-25702	293	10	completion	completion	NOUN
fcis-25702	293	11	.	.	PUNCT
fcis-25702	294	1	arxiv	arxiv	PROPN
fcis-25702	294	2	preprint	preprint	NOUN
fcis-25702	294	3	arxiv:2010.06804	arxiv:2010.06804	NOUN
fcis-25702	294	4	.	.	PUNCT
fcis-25702	295	1	[	[	X
fcis-25702	295	2	8	8	NUM
fcis-25702	295	3	]	]	X
fcis-25702	295	4	xin	xin	PROPN
fcis-25702	295	5	xu	xu	PROPN
fcis-25702	295	6	,	,	PUNCT
fcis-25702	295	7	yuqi	yuqi	PROPN
fcis-25702	295	8	zhu	zhu	PROPN
fcis-25702	295	9	,	,	PUNCT
fcis-25702	295	10	xiaohan	xiaohan	PROPN
fcis-25702	295	11	wang	wang	PROPN
fcis-25702	295	12	,	,	PUNCT
fcis-25702	295	13	and	and	CCONJ
fcis-25702	295	14	ningyu	ningyu	PROPN
fcis-25702	295	15	zhang	zhang	PROPN
fcis-25702	295	16	.	.	PROPN
fcis-25702	295	17	2023	2023	NUM
fcis-25702	295	18	.	.	PUNCT
fcis-25702	296	1	how	how	SCONJ
fcis-25702	296	2	to	to	PART
fcis-25702	296	3	unleash	unleash	VERB
fcis-25702	296	4	the	the	DET
fcis-25702	296	5	power	power	NOUN
fcis-25702	296	6	of	of	ADP
fcis-25702	296	7	large	large	ADJ
fcis-25702	296	8	lan	lan	NOUN
fcis-25702	296	9	guage	guage	NOUN
fcis-25702	296	10	models	model	NOUN
fcis-25702	296	11	for	for	ADP
fcis-25702	296	12	fewshot	fewshot	ADJ
fcis-25702	296	13	relation	relation	NOUN
fcis-25702	296	14	extraction	extraction	NOUN
fcis-25702	296	15	?	?	PUNCT
fcis-25702	297	1	arxiv	arxiv	PROPN
fcis-25702	297	2	preprint	preprint	NOUN
fcis-25702	297	3	arxiv:2305.01555	arxiv:2305.01555	NOUN
fcis-25702	297	4	.	.	PUNCT
fcis-25702	298	1	[	[	X
fcis-25702	298	2	9	9	NUM
fcis-25702	298	3	]	]	PUNCT
fcis-25702	298	4	derong	derong	PROPN
fcis-25702	298	5	xu	xu	PROPN
fcis-25702	298	6	,	,	PUNCT
fcis-25702	298	7	wei	wei	PROPN
fcis-25702	298	8	chen	chen	PROPN
fcis-25702	298	9	,	,	PUNCT
fcis-25702	298	10	wenjun	wenjun	PROPN
fcis-25702	298	11	peng	peng	PROPN
fcis-25702	298	12	,	,	PUNCT
fcis-25702	298	13	chao	chao	PROPN
fcis-25702	298	14	zhang	zhang	PROPN
fcis-25702	298	15	,	,	PUNCT
fcis-25702	298	16	tong	tong	PROPN
fcis-25702	298	17	xu	xu	PROPN
fcis-25702	298	18	,	,	PUNCT
fcis-25702	298	19	xiangyu	xiangyu	PROPN
fcis-25702	298	20	zhao	zhao	PROPN
fcis-25702	298	21	,	,	PUNCT
fcis-25702	298	22	xian	xian	PROPN
fcis-25702	298	23	wu	wu	PROPN
fcis-25702	298	24	,	,	PUNCT
fcis-25702	298	25	yefeng	yefeng	PROPN
fcis-25702	298	26	zheng	zheng	PROPN
fcis-25702	298	27	,	,	PUNCT
fcis-25702	298	28	and	and	CCONJ
fcis-25702	298	29	enhong	enhong	PROPN
fcis-25702	298	30	chen	chen	PROPN
fcis-25702	298	31	.	.	PUNCT
fcis-25702	299	1	2023	2023	NUM
fcis-25702	299	2	.	.	PUNCT
fcis-25702	300	1	large	large	ADJ
fcis-25702	300	2	language	language	NOUN
fcis-25702	300	3	models	model	NOUN
fcis-25702	300	4	for	for	ADP
fcis-25702	300	5	generative	generative	ADJ
fcis-25702	300	6	information	information	NOUN
fcis-25702	300	7	extraction	extraction	NOUN
fcis-25702	300	8	:	:	PUNCT
fcis-25702	300	9	a	a	DET
fcis-25702	300	10	survey	survey	NOUN
fcis-25702	300	11	.	.	PUNCT
fcis-25702	301	1	arxiv	arxiv	PROPN
fcis-25702	301	2	preprint	preprint	VERB
fcis-25702	301	3	arxiv:2312.17617	arxiv:2312.17617	ADV
fcis-25702	301	4	.	.	PUNCT
fcis-25702	302	1	[	[	X
fcis-25702	302	2	10	10	NUM
fcis-25702	302	3	]	]	X
fcis-25702	302	4	xiaofei	xiaofei	PROPN
fcis-25702	302	5	sun	sun	PROPN
fcis-25702	302	6	,	,	PUNCT
fcis-25702	302	7	xiaoya	xiaoya	PROPN
fcis-25702	302	8	li	li	PROPN
fcis-25702	302	9	,	,	PUNCT
fcis-25702	302	10	jiwei	jiwei	PROPN
fcis-25702	302	11	li	li	PROPN
fcis-25702	302	12	,	,	PUNCT
fcis-25702	302	13	fei	fei	PROPN
fcis-25702	302	14	wu	wu	PROPN
fcis-25702	302	15	,	,	PUNCT
fcis-25702	302	16	shangwei	shangwei	PROPN
fcis-25702	302	17	guo	guo	PROPN
fcis-25702	302	18	,	,	PUNCT
fcis-25702	302	19	tianwei	tianwei	PROPN
fcis-25702	302	20	zhang	zhang	PROPN
fcis-25702	302	21	,	,	PUNCT
fcis-25702	302	22	and	and	CCONJ
fcis-25702	302	23	guoyin	guoyin	PROPN
fcis-25702	302	24	wang	wang	PROPN
fcis-25702	302	25	.	.	PUNCT
fcis-25702	303	1	2023	2023	NUM
fcis-25702	303	2	.	.	PUNCT
fcis-25702	304	1	text	text	NOUN
fcis-25702	304	2	classification	classification	NOUN
fcis-25702	304	3	via	via	ADP
fcis-25702	304	4	large	large	ADJ
fcis-25702	304	5	language	language	NOUN
fcis-25702	304	6	models	model	NOUN
fcis-25702	304	7	.	.	PUNCT
fcis-25702	305	1	arxiv	arxiv	PROPN
fcis-25702	305	2	preprint	preprint	NOUN
fcis-25702	305	3	arxiv:2305.08377	arxiv:2305.08377	NOUN
fcis-25702	305	4	.	.	PUNCT
fcis-25702	306	1	[	[	X
fcis-25702	306	2	11	11	NUM
fcis-25702	306	3	]	]	PUNCT
fcis-25702	306	4	aristides	aristide	VERB
fcis-25702	306	5	milios	milio	NOUN
fcis-25702	306	6	,	,	PUNCT
fcis-25702	306	7	siva	siva	PROPN
fcis-25702	306	8	reddy	reddy	PROPN
fcis-25702	306	9	,	,	PUNCT
fcis-25702	306	10	and	and	CCONJ
fcis-25702	306	11	dzmitry	dzmitry	ADJ
fcis-25702	306	12	bahdanau	bahdanau	NOUN
fcis-25702	306	13	.	.	PUNCT
fcis-25702	307	1	2023	2023	NUM
fcis-25702	307	2	.	.	PUNCT
fcis-25702	308	1	in	in	ADP
fcis-25702	308	2	-	-	PUNCT
fcis-25702	308	3	context	context	NOUN
fcis-25702	308	4	learning	learning	NOUN
fcis-25702	308	5	for	for	ADP
fcis-25702	308	6	text	text	NOUN
fcis-25702	308	7	classification	classification	NOUN
fcis-25702	308	8	with	with	ADP
fcis-25702	308	9	many	many	ADJ
fcis-25702	308	10	labels	label	NOUN
fcis-25702	308	11	.	.	PUNCT
fcis-25702	309	1	in	in	ADP
fcis-25702	309	2	proceedings	proceeding	NOUN
fcis-25702	309	3	of	of	ADP
fcis-25702	309	4	the	the	DET
fcis-25702	309	5	1st	1st	ADJ
fcis-25702	309	6	genbench	genbench	NOUN
fcis-25702	309	7	workshop	workshop	NOUN
fcis-25702	309	8	on	on	ADP
fcis-25702	309	9	(	(	PUNCT
fcis-25702	309	10	benchmarking	benchmarke	VERB
fcis-25702	309	11	)	)	PUNCT
fcis-25702	309	12	generalisation	generalisation	NOUN
fcis-25702	309	13	in	in	ADP
fcis-25702	309	14	nlp	nlp	NOUN
fcis-25702	309	15	,	,	PUNCT
fcis-25702	309	16	pages	page	NOUN
fcis-25702	309	17	173	173	NUM
fcis-25702	309	18	-	-	SYM
fcis-25702	309	19	184	184	NUM
fcis-25702	309	20	.	.	PUNCT
fcis-25702	310	1	[	[	X
fcis-25702	310	2	12	12	NUM
fcis-25702	310	3	]	]	X
fcis-25702	310	4	gaurav	gaurav	PROPN
fcis-25702	310	5	sahu	sahu	PROPN
fcis-25702	310	6	,	,	PUNCT
fcis-25702	310	7	olga	olga	PROPN
fcis-25702	310	8	vechtomova	vechtomova	PROPN
fcis-25702	310	9	,	,	PUNCT
fcis-25702	310	10	dzmitry	dzmitry	PROPN
fcis-25702	310	11	bahdanau	bahdanau	NOUN
fcis-25702	310	12	,	,	PUNCT
fcis-25702	310	13	and	and	CCONJ
fcis-25702	310	14	issam	issam	PROPN
fcis-25702	310	15	h	h	PROPN
fcis-25702	310	16	laradji	laradji	PROPN
fcis-25702	310	17	.	.	PUNCT
fcis-25702	311	1	2023	2023	NUM
fcis-25702	311	2	.	.	PUNCT
fcis-25702	312	1	promptmix	promptmix	NOUN
fcis-25702	312	2	:	:	PUNCT
fcis-25702	312	3	a	a	DET
fcis-25702	312	4	class	class	NOUN
fcis-25702	312	5	boundary	boundary	ADJ
fcis-25702	312	6	augmentation	augmentation	NOUN
fcis-25702	312	7	method	method	NOUN
fcis-25702	312	8	for	for	ADP
fcis-25702	312	9	large	large	ADJ
fcis-25702	312	10	language	language	NOUN
fcis-25702	312	11	model	model	NOUN
fcis-25702	312	12	distillation	distillation	NOUN
fcis-25702	312	13	.	.	PUNCT
fcis-25702	313	1	arxiv	arxiv	PROPN
fcis-25702	313	2	preprint	preprint	NOUN
fcis-25702	313	3	arxiv:2310.14192	arxiv:2310.14192	NOUN
fcis-25702	313	4	.	.	PUNCT
fcis-25702	314	1	[	[	X
fcis-25702	314	2	13	13	NUM
fcis-25702	314	3	]	]	PUNCT
fcis-25702	314	4	lochan	lochan	PROPN
fcis-25702	314	5	basyal	basyal	PROPN
fcis-25702	314	6	and	and	CCONJ
fcis-25702	314	7	mihir	mihir	PROPN
fcis-25702	314	8	sanghvi	sanghvi	PROPN
fcis-25702	314	9	.	.	PUNCT
fcis-25702	315	1	2023	2023	NUM
fcis-25702	315	2	.	.	PUNCT
fcis-25702	316	1	text	text	NOUN
fcis-25702	316	2	summarization	summarization	NOUN
fcis-25702	316	3	using	use	VERB
fcis-25702	316	4	large	large	ADJ
fcis-25702	316	5	language	language	NOUN
fcis-25702	316	6	models	model	NOUN
fcis-25702	316	7	:	:	PUNCT
fcis-25702	316	8	acomparative	acomparative	ADJ
fcis-25702	316	9	study	study	NOUN
fcis-25702	316	10	of	of	ADP
fcis-25702	316	11	mpt-7binstruct	mpt-7binstruct	NOUN
fcis-25702	316	12	,	,	PUNCT
fcis-25702	316	13	falcon-7binstruct	falcon-7binstruct	NOUN
fcis-25702	316	14	,	,	PUNCT
fcis-25702	316	15	and	and	CCONJ
fcis-25702	316	16	openai	openai	VERB
fcis-25702	316	17	chat	chat	NOUN
fcis-25702	316	18	-	-	PUNCT
fcis-25702	316	19	gpt	gpt	NOUN
fcis-25702	316	20	models	model	NOUN
fcis-25702	316	21	.	.	PUNCT
fcis-25702	317	1	arxiv	arxiv	PROPN
fcis-25702	317	2	preprintarxiv:2310.10449	preprintarxiv:2310.10449	PROPN
fcis-25702	317	3	.	.	PUNCT
fcis-25702	318	1	[	[	X
fcis-25702	318	2	14	14	NUM
fcis-25702	318	3	]	]	X
fcis-25702	318	4	youngjin	youngjin	PROPN
fcis-25702	318	5	chae	chae	PROPN
fcis-25702	318	6	and	and	CCONJ
fcis-25702	318	7	thomas	thomas	PROPN
fcis-25702	318	8	davidson	davidson	PROPN
fcis-25702	318	9	.	.	PUNCT
fcis-25702	319	1	2023	2023	NUM
fcis-25702	319	2	.	.	PUNCT
fcis-25702	320	1	large	large	ADJ
fcis-25702	320	2	lan	lan	PROPN
fcis-25702	320	3	guage	guage	NOUN
fcis-25702	320	4	models	model	NOUN
fcis-25702	320	5	for	for	ADP
fcis-25702	320	6	text	text	NOUN
fcis-25702	320	7	classification	classification	NOUN
fcis-25702	320	8	:	:	PUNCT
fcis-25702	320	9	from	from	ADP
fcis-25702	320	10	zero	zero	NUM
fcis-25702	320	11	-	-	PUNCT
fcis-25702	320	12	shot	shot	NOUN
fcis-25702	320	13	learning	learning	NOUN
fcis-25702	320	14	to	to	ADP
fcis-25702	320	15	finetuning	finetune	VERB
fcis-25702	320	16	.	.	PUNCT
fcis-25702	321	1	open	open	ADJ
fcis-25702	321	2	science	science	PROPN
fcis-25702	321	3	foundation	foundation	PROPN
fcis-25702	321	4	.	.	PUNCT
fcis-25702	322	1	[	[	X
fcis-25702	322	2	15	15	NUM
fcis-25702	322	3	]	]	X
fcis-25702	322	4	jacob	jacob	PROPN
fcis-25702	322	5	devlin	devlin	PROPN
fcis-25702	322	6	,	,	PUNCT
fcis-25702	322	7	ming	ming	PROPN
fcis-25702	322	8	-	-	PUNCT
fcis-25702	322	9	wei	wei	PROPN
fcis-25702	322	10	chang	chang	PROPN
fcis-25702	322	11	,	,	PUNCT
fcis-25702	322	12	kenton	kenton	PROPN
fcis-25702	322	13	lee	lee	PROPN
fcis-25702	322	14	,	,	PUNCT
fcis-25702	322	15	and	and	CCONJ
fcis-25702	322	16	kristina	kristina	PROPN
fcis-25702	322	17	toutanova	toutanova	PROPN
fcis-25702	322	18	.	.	PROPN
fcis-25702	322	19	2018	2018	NUM
fcis-25702	322	20	.	.	PUNCT
fcis-25702	323	1	bert	bert	PROPN
fcis-25702	323	2	:	:	PUNCT
fcis-25702	323	3	pre	pre	ADJ
fcis-25702	323	4	-	-	NOUN
fcis-25702	323	5	training	training	NOUN
fcis-25702	323	6	of	of	ADP
fcis-25702	323	7	deep	deep	ADJ
fcis-25702	323	8	bidirectional	bidirectional	ADJ
fcis-25702	323	9	transformers	transformer	NOUN
fcis-25702	323	10	for	for	ADP
fcis-25702	323	11	language	language	NOUN
fcis-25702	323	12	understanding	understanding	NOUN
fcis-25702	323	13	.	.	PUNCT
fcis-25702	324	1	arxiv	arxiv	PROPN
fcis-25702	324	2	preprint	preprint	PROPN
fcis-25702	324	3	ar	ar	PROPN
fcis-25702	324	4	xiv	xiv	PROPN
fcis-25702	324	5	:	:	PUNCT
fcis-25702	324	6	1810.04805	1810.04805	NUM
fcis-25702	324	7	.	.	PUNCT
fcis-25702	325	1	[	[	X
fcis-25702	325	2	16	16	NUM
fcis-25702	325	3	]	]	X
fcis-25702	325	4	mengzhou	mengzhou	NOUN
fcis-25702	325	5	xia	xia	PROPN
fcis-25702	325	6	,	,	PUNCT
fcis-25702	325	7	tianyu	tianyu	PROPN
fcis-25702	325	8	gao	gao	PROPN
fcis-25702	325	9	,	,	PUNCT
fcis-25702	325	10	zhiyuan	zhiyuan	PROPN
fcis-25702	325	11	zeng	zeng	PROPN
fcis-25702	325	12	,	,	PUNCT
fcis-25702	325	13	and	and	CCONJ
fcis-25702	325	14	danqichen	danqichen	NOUN
fcis-25702	325	15	.	.	PUNCT
fcis-25702	326	1	2023	2023	NUM
fcis-25702	326	2	.	.	PUNCT
fcis-25702	327	1	sheared	shear	VERB
fcis-25702	327	2	llama	llama	NOUN
fcis-25702	327	3	:	:	PUNCT
fcis-25702	327	4	accelerating	accelerate	VERB
fcis-25702	327	5	language	language	NOUN
fcis-25702	327	6	model	model	NOUN
fcis-25702	327	7	pre	pre	NOUN
fcis-25702	327	8	-	-	NOUN
fcis-25702	327	9	training	training	NOUN
fcis-25702	327	10	via	via	ADP
fcis-25702	327	11	structured	structured	ADJ
fcis-25702	327	12	pruning	pruning	NOUN
fcis-25702	327	13	.	.	PUNCT
fcis-25702	328	1	arxiv	arxiv	PROPN
fcis-25702	328	2	preprint	preprint	VERB
fcis-25702	328	3	arxiv:2310.06694	arxiv:2310.06694	PROPN
fcis-25702	328	4	.	.	PUNCT
fcis-25702	329	1	[	[	X
fcis-25702	329	2	17	17	NUM
fcis-25702	329	3	]	]	PUNCT
fcis-25702	329	4	ziqing	ziqing	PROPN
fcis-25702	329	5	yang	yang	PROPN
fcis-25702	329	6	,	,	PUNCT
fcis-25702	329	7	yiming	yiming	NOUN
fcis-25702	329	8	cui	cui	PROPN
fcis-25702	329	9	,	,	PUNCT
fcis-25702	329	10	xin	xin	PROPN
fcis-25702	329	11	yao	yao	PROPN
fcis-25702	329	12	,	,	PUNCT
fcis-25702	329	13	and	and	CCONJ
fcis-25702	329	14	shijin	shijin	PROPN
fcis-25702	329	15	wang	wang	PROPN
fcis-25702	329	16	.	.	PUNCT
fcis-25702	330	1	2022	2022	NUM
fcis-25702	330	2	.	.	PUNCT
fcis-25702	331	1	gradient	gradient	NOUN
fcis-25702	331	2	-	-	PUNCT
fcis-25702	331	3	based	base	VERB
fcis-25702	331	4	intra	intra	ADJ
fcis-25702	331	5	-	-	ADJ
fcis-25702	331	6	attention	attention	NOUN
fcis-25702	331	7	pruning	prune	VERB
fcis-25702	331	8	on	on	ADP
fcis-25702	331	9	pre	pre	ADJ
fcis-25702	331	10	-	-	ADJ
fcis-25702	331	11	trained	train	VERB
fcis-25702	331	12	language	language	NOUN
fcis-25702	331	13	models	model	NOUN
fcis-25702	331	14	.	.	PUNCT
fcis-25702	332	1	arxiv	arxiv	PROPN
fcis-25702	332	2	preprint	preprint	NOUN
fcis-25702	332	3	arxiv:2212.07634	arxiv:2212.07634	NOUN
fcis-25702	332	4	.	.	PUNCT
fcis-25702	333	1	[	[	X
fcis-25702	333	2	18	18	NUM
fcis-25702	333	3	]	]	X
fcis-25702	333	4	victor	victor	PROPN
fcis-25702	333	5	sanh	sanh	PROPN
fcis-25702	333	6	,	,	PUNCT
fcis-25702	333	7	lysandre	lysandre	PROPN
fcis-25702	333	8	debut	debut	PROPN
fcis-25702	333	9	,	,	PUNCT
fcis-25702	333	10	julien	julien	PROPN
fcis-25702	333	11	chaumond	chaumond	PROPN
fcis-25702	333	12	,	,	PUNCT
fcis-25702	333	13	and	and	CCONJ
fcis-25702	333	14	thomas	thomas	PROPN
fcis-25702	333	15	wolf	wolf	PROPN
fcis-25702	333	16	.	.	PUNCT
fcis-25702	333	17	2019	2019	NUM
fcis-25702	333	18	.	.	PUNCT
fcis-25702	334	1	distilbert	distilbert	PROPN
fcis-25702	334	2	,	,	PUNCT
fcis-25702	334	3	a	a	DET
fcis-25702	334	4	distilled	distil	VERB
fcis-25702	334	5	version	version	NOUN
fcis-25702	334	6	of	of	ADP
fcis-25702	334	7	bert	bert	NOUN
fcis-25702	334	8	:	:	PUNCT
fcis-25702	334	9	smaller	small	ADJ
fcis-25702	334	10	,	,	PUNCT
fcis-25702	334	11	faster	fast	ADJ
fcis-25702	334	12	,	,	PUNCT
fcis-25702	334	13	cheaper	cheap	ADJ
fcis-25702	334	14	and	and	CCONJ
fcis-25702	334	15	lighter	light	ADJ
fcis-25702	334	16	.	.	PUNCT
fcis-25702	335	1	arxivpreprint	arxivpreprint	PROPN
fcis-25702	335	2	arxiv	arxiv	PROPN
fcis-25702	335	3	:	:	PUNCT
fcis-25702	335	4	1910	1910	NUM
fcis-25702	335	5	.	.	PUNCT
fcis-25702	336	1	01108	01108	NUM
fcis-25702	336	2	.	.	PUNCT
fcis-25702	337	1	[	[	X
fcis-25702	337	2	19	19	NUM
fcis-25702	337	3	]	]	PUNCT
fcis-25702	337	4	chenhe	chenhe	NOUN
fcis-25702	337	5	dong	dong	PROPN
fcis-25702	337	6	,	,	PUNCT
fcis-25702	337	7	guangrun	guangrun	PROPN
fcis-25702	337	8	wang	wang	PROPN
fcis-25702	337	9	,	,	PUNCT
fcis-25702	337	10	hang	hang	PROPN
fcis-25702	337	11	xu	xu	PROPN
fcis-25702	337	12	,	,	PUNCT
fcis-25702	337	13	jiefeng	jiefeng	PROPN
fcis-25702	337	14	peng	peng	PROPN
fcis-25702	337	15	,	,	PUNCT
fcis-25702	337	16	xiaozhe	xiaozhe	PROPN
fcis-25702	337	17	ren	ren	PROPN
fcis-25702	337	18	,	,	PUNCT
fcis-25702	337	19	and	and	CCONJ
fcis-25702	337	20	xiaodan	xiaodan	PROPN
fcis-25702	337	21	liang	liang	PROPN
fcis-25702	337	22	.	.	PROPN
fcis-25702	337	23	2021	2021	NUM
fcis-25702	337	24	.	.	PUNCT
fcis-25702	338	1	efficient	efficient	ADJ
fcis-25702	338	2	bert	bert	NOUN
fcis-25702	338	3	:	:	PUNCT
fcis-25702	338	4	progressively	progressively	ADV
fcis-25702	338	5	searching	search	VERB
fcis-25702	338	6	multilayer	multilayer	ADJ
fcis-25702	338	7	perceptron	perceptron	NOUN
fcis-25702	338	8	via	via	ADP
fcis-25702	338	9	warm	warm	ADJ
fcis-25702	338	10	-	-	PUNCT
fcis-25702	338	11	up	up	ADP
fcis-25702	338	12	knowledge	knowledge	NOUN
fcis-25702	338	13	distillation	distillation	NOUN
fcis-25702	338	14	.	.	PUNCT
fcis-25702	339	1	arxiv	arxiv	PROPN
fcis-25702	339	2	preprint	preprint	PROPN
fcis-25702	339	3	arxiv:2109.07222	arxiv:2109.07222	NOUN
fcis-25702	339	4	.	.	PUNCT
fcis-25702	340	1	[	[	X
fcis-25702	340	2	20	20	NUM
fcis-25702	340	3	]	]	X
fcis-25702	340	4	ji	ji	PROPN
fcis-25702	340	5	lin	lin	PROPN
fcis-25702	340	6	,	,	PUNCT
fcis-25702	340	7	jiaming	jiame	VERB
fcis-25702	340	8	tang	tang	PROPN
fcis-25702	340	9	,	,	PUNCT
fcis-25702	340	10	haotian	haotian	PROPN
fcis-25702	340	11	tang	tang	PROPN
fcis-25702	340	12	,	,	PUNCT
fcis-25702	340	13	shang	shang	PROPN
fcis-25702	340	14	yang	yang	PROPN
fcis-25702	340	15	,	,	PUNCT
fcis-25702	340	16	xingyu	xingyu	PROPN
fcis-25702	340	17	dang	dang	PROPN
fcis-25702	340	18	,	,	PUNCT
fcis-25702	340	19	and	and	CCONJ
fcis-25702	340	20	song	song	PROPN
fcis-25702	340	21	han	han	PROPN
fcis-25702	340	22	.	.	PROPN
fcis-25702	340	23	2023	2023	NUM
fcis-25702	340	24	.	.	PUNCT
fcis-25702	341	1	awq	awq	NOUN
fcis-25702	341	2	:	:	PUNCT
fcis-25702	341	3	activation	activation	NOUN
fcis-25702	341	4	aware	aware	ADJ
fcis-25702	341	5	weight	weight	NOUN
fcis-25702	341	6	quantization	quantization	NOUN
fcis-25702	341	7	for	for	ADP
fcis-25702	341	8	11	11	NUM
fcis-25702	341	9	lm	lm	NOUN
fcis-25702	341	10	compression	compression	NOUN
fcis-25702	341	11	and	and	CCONJ
fcis-25702	341	12	acceleration	acceleration	NOUN
fcis-25702	341	13	.	.	PUNCT
fcis-25702	342	1	arxiv	arxiv	PROPN
fcis-25702	342	2	preprint	preprint	VERB
fcis-25702	342	3	arxiv:2306.00978	arxiv:2306.00978	NOUN
fcis-25702	342	4	.	.	PUNCT
fcis-25702	343	1	[	[	X
fcis-25702	343	2	21	21	NUM
fcis-25702	343	3	]	]	PUNCT
fcis-25702	343	4	elias	elias	PROPN
fcis-25702	343	5	frantar	frantar	PROPN
fcis-25702	343	6	,	,	PUNCT
fcis-25702	343	7	saleh	saleh	PROPN
fcis-25702	343	8	ashkboos	ashkboos	PROPN
fcis-25702	343	9	,	,	PUNCT
fcis-25702	343	10	torsten	torsten	PROPN
fcis-25702	343	11	hoefler	hoefler	PROPN
fcis-25702	343	12	,	,	PUNCT
fcis-25702	343	13	and	and	CCONJ
fcis-25702	343	14	dan	dan	PROPN
fcis-25702	343	15	alistarh	alistarh	PROPN
fcis-25702	343	16	.	.	PUNCT
fcis-25702	343	17	2022	2022	NUM
fcis-25702	343	18	.	.	PUNCT
fcis-25702	344	1	gptq	gptq	NOUN
fcis-25702	344	2	:	:	PUNCT
fcis-25702	344	3	accurate	accurate	ADJ
fcis-25702	344	4	post	post	ADJ
fcis-25702	344	5	-	-	ADJ
fcis-25702	344	6	training	training	ADJ
fcis-25702	344	7	quantization	quantization	NOUN
fcis-25702	344	8	for	for	ADP
fcis-25702	344	9	generative	generative	ADJ
fcis-25702	344	10	pre	pre	ADJ
fcis-25702	344	11	-	-	ADJ
fcis-25702	344	12	trained	train	VERB
fcis-25702	344	13	transformers.arxiv	transformers.arxiv	PROPN
fcis-25702	344	14	preprint	preprint	NOUN
fcis-25702	344	15	arxiv	arxiv	NOUN
fcis-25702	344	16	:	:	PUNCT
fcis-25702	344	17	2210	2210	NUM
fcis-25702	344	18	.	.	PUNCT
fcis-25702	345	1	17323	17323	NUM
fcis-25702	345	2	.	.	PUNCT
fcis-25702	346	1	[	[	X
fcis-25702	346	2	22	22	NUM
fcis-25702	346	3	]	]	X
fcis-25702	346	4	surat	surat	PROPN
fcis-25702	346	5	teerapittayanon	teerapittayanon	PROPN
fcis-25702	346	6	,	,	PUNCT
fcis-25702	346	7	bradley	bradley	NOUN
fcis-25702	346	8	mcdanel	mcdanel	PROPN
fcis-25702	346	9	,	,	PUNCT
fcis-25702	346	10	and	and	CCONJ
fcis-25702	346	11	hsiang	hsiang	PROPN
fcis-25702	346	12	tsung	tsung	PROPN
fcis-25702	346	13	kung	kung	PROPN
fcis-25702	346	14	.	.	PROPN
fcis-25702	346	15	2016	2016	NUM
fcis-25702	346	16	.	.	PUNCT
fcis-25702	347	1	branchynet	branchynet	NOUN
fcis-25702	347	2	:	:	PUNCT
fcis-25702	347	3	fast	fast	ADJ
fcis-25702	347	4	inference	inference	NOUN
fcis-25702	347	5	via	via	ADP
fcis-25702	347	6	early	early	ADJ
fcis-25702	347	7	exiting	exit	VERB
fcis-25702	347	8	from	from	ADP
fcis-25702	347	9	deep	deep	ADJ
fcis-25702	347	10	neural	neural	ADJ
fcis-25702	347	11	networks	network	NOUN
fcis-25702	347	12	.	.	PUNCT
fcis-25702	348	1	in	in	ADP
fcis-25702	348	2	2016	2016	NUM
fcis-25702	348	3	23	23	NUM
fcis-25702	348	4	rd	rd	NOUN
fcis-25702	348	5	international	international	ADJ
fcis-25702	348	6	conference	conference	NOUN
fcis-25702	348	7	on	on	ADP
fcis-25702	348	8	pattern	pattern	NOUN
fcis-25702	348	9	recognition	recognition	NOUN
fcis-25702	348	10	,	,	PUNCT
fcis-25702	348	11	pages	page	NOUN
fcis-25702	348	12	2464	2464	NUM
fcis-25702	348	13	-	-	SYM
fcis-25702	348	14	2469	2469	NUM
fcis-25702	348	15	.	.	PUNCT
fcis-25702	348	16	ieee	ieee	NOUN
fcis-25702	348	17	.	.	PUNCT
fcis-25702	349	1	[	[	X
fcis-25702	349	2	23	23	NUM
fcis-25702	349	3	]	]	X
fcis-25702	349	4	weijie	weijie	PROPN
fcis-25702	349	5	liu	liu	PROPN
fcis-25702	349	6	,	,	PUNCT
fcis-25702	349	7	peng	peng	PROPN
fcis-25702	349	8	zhou	zhou	PROPN
fcis-25702	349	9	,	,	PUNCT
fcis-25702	349	10	zhe	zhe	PROPN
fcis-25702	349	11	zhao	zhao	PROPN
fcis-25702	349	12	,	,	PUNCT
fcis-25702	349	13	zhiruo	zhiruo	PROPN
fcis-25702	349	14	wang	wang	PROPN
fcis-25702	349	15	,	,	PUNCT
fcis-25702	349	16	haotang	haotang	PROPN
fcis-25702	349	17	deng	deng	PROPN
fcis-25702	349	18	,	,	PUNCT
fcis-25702	349	19	and	and	CCONJ
fcis-25702	349	20	qi	qi	PROPN
fcis-25702	349	21	ju	ju	PROPN
fcis-25702	349	22	.	.	PROPN
fcis-25702	349	23	2020	2020	NUM
fcis-25702	349	24	.	.	PUNCT
fcis-25702	350	1	fastbert	fastbert	ADJ
fcis-25702	350	2	:	:	PUNCT
fcis-25702	350	3	a	a	DET
fcis-25702	350	4	selfdistilling	selfdistille	VERB
fcis-25702	350	5	bert	bert	NOUN
fcis-25702	350	6	with	with	ADP
fcis-25702	350	7	adaptive	adaptive	ADJ
fcis-25702	350	8	inference	inference	NOUN
fcis-25702	350	9	time	time	NOUN
fcis-25702	350	10	.	.	PUNCT
fcis-25702	351	1	arxivpreprint	arxivpreprint	NOUN
fcis-25702	351	2	arxiv:2004.02178	arxiv:2004.02178	NOUN
fcis-25702	351	3	.	.	PUNCT
fcis-25702	352	1	[	[	X
fcis-25702	352	2	24	24	NUM
fcis-25702	352	3	]	]	PUNCT
fcis-25702	352	4	wangchunshu	wangchunshu	PROPN
fcis-25702	352	5	zhou	zhou	PROPN
fcis-25702	352	6	,	,	PUNCT
fcis-25702	352	7	canwen	canwen	PROPN
fcis-25702	352	8	xu	xu	PROPN
fcis-25702	352	9	,	,	PUNCT
fcis-25702	352	10	tao	tao	PROPN
fcis-25702	352	11	ge	ge	PROPN
fcis-25702	352	12	,	,	PUNCT
fcis-25702	352	13	julian	julian	PROPN
fcis-25702	352	14	mcauley	mcauley	PROPN
fcis-25702	352	15	,	,	PUNCT
fcis-25702	352	16	ke	ke	PROPN
fcis-25702	352	17	xu	xu	PROPN
fcis-25702	352	18	,	,	PUNCT
fcis-25702	352	19	and	and	CCONJ
fcis-25702	352	20	furu	furu	PROPN
fcis-25702	352	21	wei	wei	PROPN
fcis-25702	352	22	.	.	PUNCT
fcis-25702	352	23	2020	2020	NUM
fcis-25702	352	24	.	.	PUNCT
fcis-25702	353	1	bert	bert	PROPN
fcis-25702	353	2	loses	lose	VERB
fcis-25702	353	3	patience	patience	NOUN
fcis-25702	353	4	:	:	PUNCT
fcis-25702	353	5	fast	fast	ADJ
fcis-25702	353	6	and	and	CCONJ
fcis-25702	353	7	robust	robust	ADJ
fcis-25702	353	8	inference	inference	NOUN
fcis-25702	353	9	with	with	ADP
fcis-25702	353	10	early	early	ADJ
fcis-25702	353	11	exit	exit	NOUN
fcis-25702	353	12	.	.	PUNCT
fcis-25702	354	1	advances	advance	NOUN
fcis-25702	354	2	in	in	ADP
fcis-25702	354	3	neural	neural	ADJ
fcis-25702	354	4	information	information	NOUN
fcis-25702	354	5	processing	processing	NOUN
fcis-25702	354	6	systems	system	NOUN
fcis-25702	354	7	.	.	PUNCT
fcis-25702	355	1	arxiv:2006.04152	arxiv:2006.04152	NOUN
fcis-25702	355	2	.	.	PUNCT
fcis-25702	356	1	[	[	X
fcis-25702	356	2	25	25	NUM
fcis-25702	356	3	]	]	X
fcis-25702	356	4	zhen	zhen	PROPN
fcis-25702	356	5	zhang	zhang	PROPN
fcis-25702	356	6	,	,	PUNCT
fcis-25702	356	7	wei	wei	PROPN
fcis-25702	356	8	zhu	zhu	PROPN
fcis-25702	356	9	,	,	PUNCT
fcis-25702	356	10	jinfan	jinfan	PROPN
fcis-25702	356	11	zhang	zhang	PROPN
fcis-25702	356	12	,	,	PUNCT
fcis-25702	356	13	peng	peng	PROPN
fcis-25702	356	14	wang	wang	PROPN
fcis-25702	356	15	,	,	PUNCT
fcis-25702	356	16	rize	rize	PROPN
fcis-25702	356	17	jin	jin	PROPN
fcis-25702	356	18	,	,	PUNCT
fcis-25702	356	19	and	and	CCONJ
fcis-25702	356	20	tae	tae	NOUN
fcis-25702	356	21	-	-	PUNCT
fcis-25702	356	22	sun	sun	NOUN
fcis-25702	356	23	chung	chung	PROPN
fcis-25702	356	24	.	.	PUNCT
fcis-25702	356	25	2022	2022	NUM
fcis-25702	356	26	.	.	PUNCT
fcis-25702	357	1	pcee	pcee	NOUN
fcis-25702	357	2	-	-	PUNCT
fcis-25702	357	3	bert	bert	NOUN
fcis-25702	357	4	:	:	PUNCT
fcis-25702	357	5	accelerating	accelerate	VERB
fcis-25702	357	6	bert	bert	NOUN
fcis-25702	357	7	inference	inference	NOUN
fcis-25702	357	8	via	via	ADP
fcis-25702	357	9	patient	patient	NOUN
fcis-25702	357	10	and	and	CCONJ
fcis-25702	357	11	confident	confident	ADJ
fcis-25702	357	12	early	early	ADJ
fcis-25702	357	13	exiting	exiting	NOUN
fcis-25702	357	14	.	.	PUNCT
fcis-25702	358	1	in	in	ADP
fcis-25702	358	2	findings	finding	NOUN
fcis-25702	358	3	of	of	ADP
fcis-25702	358	4	the	the	DET
fcis-25702	358	5	association	association	NOUN
fcis-25702	358	6	for	for	ADP
fcis-25702	358	7	computational	computational	ADJ
fcis-25702	358	8	linguistics	linguistic	NOUN
fcis-25702	358	9	:	:	PUNCT
fcis-25702	358	10	naacl	naacl	PROPN
fcis-25702	358	11	2022	2022	NUM
fcis-25702	358	12	,	,	PUNCT
fcis-25702	358	13	pages	page	NOUN
fcis-25702	358	14	327	327	NUM
fcis-25702	358	15	-	-	SYM
fcis-25702	358	16	338	338	NUM
fcis-25702	358	17	.	.	PUNCT
fcis-25702	359	1	[	[	X
fcis-25702	359	2	26	26	NUM
fcis-25702	359	3	]	]	X
fcis-25702	359	4	jingfan	jingfan	PROPN
fcis-25702	359	5	zhang	zhang	PROPN
fcis-25702	359	6	,	,	PUNCT
fcis-25702	359	7	ming	ming	PROPN
fcis-25702	359	8	tan	tan	PROPN
fcis-25702	359	9	,	,	PUNCT
fcis-25702	359	10	pengyu	pengyu	PROPN
fcis-25702	359	11	dai	dai	PROPN
fcis-25702	359	12	,	,	PUNCT
fcis-25702	359	13	and	and	CCONJ
fcis-25702	359	14	wei	wei	PROPN
fcis-25702	359	15	zhu	zhu	PROPN
fcis-25702	359	16	.	.	PUNCT
fcis-25702	359	17	2023	2023	NUM
fcis-25702	359	18	.	.	PUNCT
fcis-25702	360	1	leco	leco	PROPN
fcis-25702	360	2	:	:	PUNCT
fcis-25702	360	3	improving	improve	VERB
fcis-25702	360	4	early	early	ADV
fcis-25702	360	5	exiting	exit	VERB
fcis-25702	360	6	via	via	ADP
fcis-25702	360	7	learned	learn	VERB
fcis-25702	360	8	exits	exit	NOUN
fcis-25702	360	9	and	and	CCONJ
fcis-25702	360	10	comparison	comparison	NOUN
fcis-25702	360	11	-	-	PUNCT
fcis-25702	360	12	based	base	VERB
fcis-25702	360	13	exiting	exiting	NOUN
fcis-25702	360	14	mechanism	mechanism	NOUN
fcis-25702	360	15	.	.	PUNCT
fcis-25702	361	1	in	in	ADP
fcis-25702	361	2	proceedings	proceeding	NOUN
fcis-25702	361	3	of	of	ADP
fcis-25702	361	4	the	the	DET
fcis-25702	361	5	61st	61st	ADJ
fcis-25702	361	6	annual	annual	ADJ
fcis-25702	361	7	meeting	meeting	NOUN
fcis-25702	361	8	of	of	ADP
fcis-25702	361	9	the	the	DET
fcis-25702	361	10	association	association	NOUN
fcis-25702	361	11	for	for	ADP
fcis-25702	361	12	computational	computational	ADJ
fcis-25702	361	13	linguistics	linguistic	NOUN
fcis-25702	361	14	(	(	PUNCT
fcis-25702	361	15	volume	volume	NOUN
fcis-25702	361	16	4	4	NUM
fcis-25702	361	17	:	:	PUNCT
fcis-25702	361	18	student	student	NOUN
fcis-25702	361	19	research	research	NOUN
fcis-25702	361	20	workshop	workshop	NOUN
fcis-25702	361	21	)	)	PUNCT
fcis-25702	361	22	,	,	PUNCT
fcis-25702	361	23	pages	page	NOUN
fcis-25702	361	24	298	298	NUM
fcis-25702	361	25	-	-	SYM
fcis-25702	361	26	309	309	NUM
fcis-25702	361	27	.	.	PUNCT
fcis-25702	362	1	[	[	X
fcis-25702	362	2	27	27	NUM
fcis-25702	362	3	]	]	X
fcis-25702	362	4	alec	alec	PROPN
fcis-25702	362	5	radford	radford	PROPN
fcis-25702	362	6	,	,	PUNCT
fcis-25702	362	7	karthik	karthik	PROPN
fcis-25702	362	8	narasimhan	narasimhan	PROPN
fcis-25702	362	9	,	,	PUNCT
fcis-25702	362	10	tim	tim	PROPN
fcis-25702	362	11	salimans	salimans	PROPN
fcis-25702	362	12	,	,	PUNCT
fcis-25702	362	13	ilya	ilya	PROPN
fcis-25702	362	14	sutskever	sutskever	PROPN
fcis-25702	362	15	,	,	PUNCT
fcis-25702	362	16	et	et	PROPN
fcis-25702	362	17	al	al	PROPN
fcis-25702	362	18	.	.	PROPN
fcis-25702	362	19	2018	2018	NUM
fcis-25702	362	20	.	.	PUNCT
fcis-25702	363	1	improving	improve	VERB
fcis-25702	363	2	language	language	NOUN
fcis-25702	363	3	under	under	ADV
fcis-25702	363	4	-	-	PUNCT
fcis-25702	363	5	standing	standing	NOUN
fcis-25702	363	6	by	by	ADP
fcis-25702	363	7	generative	generative	ADJ
fcis-25702	363	8	pre	pre	NOUN
fcis-25702	363	9	-	-	NOUN
fcis-25702	363	10	training	training	NOUN
fcis-25702	363	11	.	.	PUNCT
fcis-25702	364	1	[	[	X
fcis-25702	364	2	28	28	NUM
fcis-25702	364	3	]	]	X
fcis-25702	364	4	bowen	bowen	PROPN
fcis-25702	364	5	shen	shen	PROPN
fcis-25702	364	6	,	,	PUNCT
fcis-25702	364	7	zheng	zheng	PROPN
fcis-25702	364	8	lin	lin	PROPN
fcis-25702	364	9	,	,	PUNCT
fcis-25702	364	10	yuanxin	yuanxin	PROPN
fcis-25702	364	11	liu	liu	PROPN
fcis-25702	364	12	,	,	PUNCT
fcis-25702	364	13	zhengxiao	zhengxiao	PROPN
fcis-25702	364	14	liu	liu	PROPN
fcis-25702	364	15	,	,	PUNCT
fcis-25702	364	16	lei	lei	PROPN
fcis-25702	364	17	wang	wang	PROPN
fcis-25702	364	18	,	,	PUNCT
fcis-25702	364	19	and	and	CCONJ
fcis-25702	364	20	weiping	weipe	VERB
fcis-25702	364	21	wang	wang	PROPN
fcis-25702	364	22	.	.	PUNCT
fcis-25702	364	23	2022	2022	NUM
fcis-25702	364	24	.	.	PUNCT
fcis-25702	365	1	cost	cost	NOUN
fcis-25702	365	2	-	-	PUNCT
fcis-25702	365	3	eff	eff	PROPN
fcis-25702	365	4	:	:	PUNCT
fcis-25702	365	5	col	col	PROPN
fcis-25702	365	6	laborative	laborative	ADJ
fcis-25702	365	7	9	9	NUM
fcis-25702	365	8	optimization	optimization	NOUN
fcis-25702	365	9	of	of	ADP
fcis-25702	365	10	spatial	spatial	ADJ
fcis-25702	365	11	and	and	CCONJ
fcis-25702	365	12	temporal	temporal	ADJ
fcis-25702	365	13	effi	effi	PROPN
fcis-25702	365	14	ciency	ciency	NOUN
fcis-25702	365	15	with	with	ADP
fcis-25702	365	16	slenderized	slenderize	VERB
fcis-25702	365	17	multi	multi	ADJ
fcis-25702	365	18	-	-	ADJ
fcis-25702	365	19	exit	exit	ADJ
fcis-25702	365	20	language	language	NOUN
fcis-25702	365	21	models	model	NOUN
fcis-25702	365	22	.	.	PUNCT
fcis-25702	366	1	arxiv	arxiv	PROPN
fcis-25702	366	2	preprint	preprint	PROPN
fcis-25702	366	3	arxiv	arxiv	PROPN
fcis-25702	366	4	:	:	PUNCT
fcis-25702	366	5	2210.15523	2210.15523	NOUN
fcis-25702	366	6	.	.	PUNCT
fcis-25702	367	1	[	[	X
fcis-25702	367	2	29	29	NUM
fcis-25702	367	3	]	]	X
fcis-25702	367	4	xin	xin	PROPN
fcis-25702	367	5	men	man	NOUN
fcis-25702	367	6	,	,	PUNCT
fcis-25702	367	7	mingyu	mingyu	PROPN
fcis-25702	367	8	xu	xu	PROPN
fcis-25702	367	9	,	,	PUNCT
fcis-25702	367	10	qingyu	qingyu	PROPN
fcis-25702	367	11	zhang	zhang	PROPN
fcis-25702	367	12	,	,	PUNCT
fcis-25702	367	13	bingning	bingning	PROPN
fcis-25702	367	14	wang	wang	PROPN
fcis-25702	367	15	,	,	PUNCT
fcis-25702	367	16	hongyu	hongyu	VERB
fcis-25702	367	17	lin	lin	PROPN
fcis-25702	367	18	,	,	PUNCT
fcis-25702	367	19	yaojie	yaojie	PROPN
fcis-25702	367	20	lu	lu	PROPN
fcis-25702	367	21	,	,	PUNCT
fcis-25702	367	22	xianpei	xianpei	PROPN
fcis-25702	367	23	han	han	PROPN
fcis-25702	367	24	,	,	PUNCT
fcis-25702	367	25	and	and	CCONJ
fcis-25702	367	26	weipeng	weipeng	VERB
fcis-25702	367	27	chen	chen	PROPN
fcis-25702	367	28	.	.	PUNCT
fcis-25702	368	1	2024	2024	NUM
fcis-25702	368	2	.	.	PUNCT
fcis-25702	369	1	shortgpt	shortgpt	NOUN
fcis-25702	369	2	:	:	PUNCT
fcis-25702	369	3	layers	layer	NOUN
fcis-25702	369	4	in	in	ADP
fcis-25702	369	5	large	large	ADJ
fcis-25702	369	6	language	language	NOUN
fcis-25702	369	7	models	model	NOUN
fcis-25702	369	8	are	be	AUX
fcis-25702	369	9	more	more	ADV
fcis-25702	369	10	redundant	redundant	ADJ
fcis-25702	369	11	than	than	SCONJ
fcis-25702	369	12	you	you	PRON
fcis-25702	369	13	expect	expect	VERB
fcis-25702	369	14	.	.	PUNCT
fcis-25702	370	1	arxiv	arxiv	PROPN
fcis-25702	370	2	preprint	preprint	VERB
fcis-25702	370	3	arxiv:2403.03853	arxiv:2403.03853	NOUN
fcis-25702	370	4	.	.	PUNCT
fcis-25702	371	1	[	[	X
fcis-25702	371	2	30	30	NUM
fcis-25702	371	3	]	]	X
fcis-25702	371	4	tal	tal	PROPN
fcis-25702	371	5	schuster	schuster	PROPN
fcis-25702	371	6	,	,	PUNCT
fcis-25702	371	7	adam	adam	PROPN
fcis-25702	371	8	fisch	fisch	PROPN
fcis-25702	371	9	,	,	PUNCT
fcis-25702	371	10	jai	jai	PROPN
fcis-25702	371	11	gupta	gupta	PROPN
fcis-25702	371	12	,	,	PUNCT
fcis-25702	371	13	mostafa	mostafa	PROPN
fcis-25702	371	14	dehghani	dehghani	PROPN
fcis-25702	371	15	,	,	PUNCT
fcis-25702	371	16	dara	dara	PROPN
fcis-25702	371	17	bahri	bahri	PROPN
fcis-25702	371	18	,	,	PUNCT
fcis-25702	371	19	vinh	vinh	PROPN
fcis-25702	371	20	tran	tran	PROPN
fcis-25702	371	21	,	,	PUNCT
fcis-25702	371	22	yi	yi	PROPN
fcis-25702	371	23	tay	tay	PROPN
fcis-25702	371	24	,	,	PUNCT
fcis-25702	371	25	and	and	CCONJ
fcis-25702	371	26	donald	donald	PROPN
fcis-25702	371	27	metzler.2022	metzler.2022	PROPN
fcis-25702	371	28	.	.	PUNCT
fcis-25702	371	29	confident	confident	ADJ
fcis-25702	371	30	adaptive	adaptive	ADJ
fcis-25702	371	31	language	language	NOUN
fcis-25702	371	32	modeling	modeling	NOUN
fcis-25702	371	33	.	.	PUNCT
fcis-25702	372	1	ad	ad	NOUN
fcis-25702	372	2	vances	vance	NOUN
fcis-25702	372	3	in	in	ADP
fcis-25702	372	4	neural	neural	ADJ
fcis-25702	372	5	information	information	NOUN
fcis-25702	372	6	processing	processing	NOUN
fcis-25702	372	7	systems	system	NOUN
fcis-25702	372	8	.	.	PUNCT
fcis-25702	373	1	arxiv:2207.07061	arxiv:2207.07061	PROPN
fcis-25702	373	2	.	.	PUNCT
fcis-25702	374	1	[	[	X
fcis-25702	374	2	31	31	NUM
fcis-25702	374	3	]	]	X
fcis-25702	374	4	colin	colin	PROPN
fcis-25702	374	5	raffel	raffel	PROPN
fcis-25702	374	6	,	,	PUNCT
fcis-25702	374	7	noam	noam	PROPN
fcis-25702	374	8	shazeer	shazeer	PROPN
fcis-25702	374	9	,	,	PUNCT
fcis-25702	374	10	adam	adam	PROPN
fcis-25702	374	11	roberts	roberts	PROPN
fcis-25702	374	12	,	,	PUNCT
fcis-25702	374	13	katherine	katherine	PROPN
fcis-25702	374	14	lee	lee	PROPN
fcis-25702	374	15	,	,	PUNCT
fcis-25702	374	16	sharan	sharan	PROPN
fcis-25702	374	17	narang	narang	PROPN
fcis-25702	374	18	,	,	PUNCT
fcis-25702	374	19	michael	michael	PROPN
fcis-25702	374	20	matena	matena	PROPN
fcis-25702	374	21	,	,	PUNCT
fcis-25702	374	22	yanqi	yanqi	PROPN
fcis-25702	374	23	zhou	zhou	PROPN
fcis-25702	374	24	,	,	PUNCT
fcis-25702	374	25	wei	wei	PROPN
fcis-25702	374	26	li	li	PROPN
fcis-25702	374	27	,	,	PUNCT
fcis-25702	374	28	and	and	CCONJ
fcis-25702	374	29	peter	peter	PROPN
fcis-25702	374	30	j	j	PROPN
fcis-25702	374	31	liu	liu	PROPN
fcis-25702	374	32	.	.	PROPN
fcis-25702	375	1	2020	2020	NUM
fcis-25702	375	2	.	.	PUNCT
fcis-25702	376	1	exploring	explore	VERB
fcis-25702	376	2	the	the	DET
fcis-25702	376	3	lim	lim	NOUN
fcis-25702	376	4	its	its	PRON
fcis-25702	376	5	of	of	ADP
fcis-25702	376	6	transfer	transfer	NOUN
fcis-25702	376	7	learning	learn	VERB
fcis-25702	376	8	with	with	ADP
fcis-25702	376	9	a	a	DET
fcis-25702	376	10	unified	unified	ADJ
fcis-25702	376	11	text	text	NOUN
fcis-25702	376	12	-	-	PUNCT
fcis-25702	376	13	to	to	ADP
fcis-25702	376	14	-	-	PUNCT
fcis-25702	376	15	text	text	NOUN
fcis-25702	376	16	transformer	transformer	NOUN
fcis-25702	376	17	.	.	PUNCT
fcis-25702	377	1	journal	journal	NOUN
fcis-25702	377	2	of	of	ADP
fcis-25702	377	3	machine	machine	NOUN
fcis-25702	377	4	learning	learn	VERB
fcis-25702	377	5	research	research	NOUN
fcis-25702	377	6	,	,	PUNCT
fcis-25702	377	7	21(140):1	21(140):1	NUM
fcis-25702	377	8	-	-	SYM
fcis-25702	377	9	67	67	NUM
fcis-25702	377	10	.	.	PUNCT
fcis-25702	378	1	[	[	X
fcis-25702	378	2	32	32	NUM
fcis-25702	378	3	]	]	PUNCT
fcis-25702	378	4	hugo	hugo	PROPN
fcis-25702	378	5	touvron	touvron	PROPN
fcis-25702	378	6	,	,	PUNCT
fcis-25702	378	7	thibaut	thibaut	PROPN
fcis-25702	378	8	lavril	lavril	NOUN
fcis-25702	378	9	,	,	PUNCT
fcis-25702	378	10	gautier	gautier	PROPN
fcis-25702	378	11	izacard	izacard	PROPN
fcis-25702	378	12	,	,	PUNCT
fcis-25702	378	13	xavier	xavier	PROPN
fcis-25702	378	14	martinet	martinet	PROPN
fcis-25702	378	15	,	,	PUNCT
fcis-25702	378	16	marie	marie	PROPN
fcis-25702	378	17	-	-	PUNCT
fcis-25702	378	18	anne	anne	PROPN
fcis-25702	378	19	lachaux	lachaux	PROPN
fcis-25702	378	20	,	,	PUNCT
fcis-25702	378	21	timothée	timothée	NOUN
fcis-25702	378	22	lacroix	lacroix	NOUN
fcis-25702	378	23	,	,	PUNCT
fcis-25702	378	24	baptiste	baptiste	NOUN
fcis-25702	378	25	rozière	rozière	NOUN
fcis-25702	378	26	,	,	PUNCT
fcis-25702	378	27	naman	naman	PROPN
fcis-25702	378	28	goyal	goyal	PROPN
fcis-25702	378	29	,	,	PUNCT
fcis-25702	378	30	eric	eric	PROPN
fcis-25702	378	31	hambro	hambro	PROPN
fcis-25702	378	32	,	,	PUNCT
fcis-25702	378	33	faisal	faisal	PROPN
fcis-25702	378	34	azhar	azhar	PROPN
fcis-25702	378	35	,	,	PUNCT
fcis-25702	378	36	et	et	PROPN
fcis-25702	378	37	al	al	PROPN
fcis-25702	378	38	.	.	PROPN
fcis-25702	378	39	2023	2023	NUM
fcis-25702	378	40	.	.	PUNCT
fcis-25702	379	1	llama	llama	NOUN
fcis-25702	379	2	:	:	PUNCT
fcis-25702	379	3	open	open	ADJ
fcis-25702	379	4	and	and	CCONJ
fcis-25702	379	5	efficient	efficient	ADJ
fcis-25702	379	6	foundation	foundation	NOUN
fcis-25702	379	7	language	language	NOUN
fcis-25702	379	8	models	model	NOUN
fcis-25702	379	9	.	.	PUNCT
fcis-25702	380	1	arxiv	arxiv	PROPN
fcis-25702	380	2	preprint	preprint	VERB
fcis-25702	380	3	arxiv:2302.13971	arxiv:2302.13971	PROPN
fcis-25702	380	4	.	.	PUNCT
fcis-25702	381	1	[	[	X
fcis-25702	381	2	33	33	NUM
fcis-25702	381	3	]	]	X
fcis-25702	381	4	luciano	luciano	PROPN
fcis-25702	381	5	del	del	PROPN
fcis-25702	381	6	corro	corro	PROPN
fcis-25702	381	7	,	,	PUNCT
fcis-25702	381	8	allie	allie	PROPN
fcis-25702	381	9	del	del	PROPN
fcis-25702	381	10	giorno	giorno	PROPN
fcis-25702	381	11	,	,	PUNCT
fcis-25702	381	12	sahaj	sahaj	PROPN
fcis-25702	381	13	agarwal	agarwal	PROPN
fcis-25702	381	14	,	,	PUNCT
fcis-25702	381	15	bin	bin	PROPN
fcis-25702	381	16	yu	yu	PROPN
fcis-25702	381	17	,	,	PUNCT
fcis-25702	381	18	ahmed	ahmed	PROPN
fcis-25702	381	19	awadallah	awadallah	PROPN
fcis-25702	381	20	,	,	PUNCT
fcis-25702	381	21	and	and	CCONJ
fcis-25702	381	22	subhabrata	subhabrata	PROPN
fcis-25702	381	23	mukher	mukher	PROPN
fcis-25702	381	24	jee	jee	PROPN
fcis-25702	381	25	.	.	PUNCT
fcis-25702	381	26	2023	2023	NUM
fcis-25702	381	27	.	.	PUNCT
fcis-25702	382	1	skipdecode	skipdecode	NOUN
fcis-25702	382	2	:	:	PUNCT
fcis-25702	382	3	autoregressive	autoregressive	ADJ
fcis-25702	382	4	skip	skip	NOUN
fcis-25702	382	5	decoding	decode	VERB
fcis-25702	382	6	with	with	ADP
fcis-25702	382	7	batching	batching	NOUN
fcis-25702	382	8	and	and	CCONJ
fcis-25702	382	9	caching	cache	VERB
fcis-25702	382	10	for	for	ADP
fcis-25702	382	11	efficient	efficient	ADJ
fcis-25702	382	12	llm	llm	PROPN
fcis-25702	382	13	inference.arxiv	inference.arxiv	PROPN
fcis-25702	382	14	preprint	preprint	NOUN
fcis-25702	382	15	arxiv	arxiv	NOUN
fcis-25702	382	16	:	:	PUNCT
fcis-25702	382	17	2307	2307	NUM
fcis-25702	382	18	.	.	PUNCT
fcis-25702	382	19	02628	02628	NUM
fcis-25702	382	20	.	.	PUNCT
fcis-25702	383	1	[	[	X
fcis-25702	383	2	34	34	NUM
fcis-25702	383	3	]	]	SYM
fcis-25702	383	4	shuzhou	shuzhou	PROPN
fcis-25702	383	5	yuan	yuan	PROPN
fcis-25702	383	6	,	,	PUNCT
fcis-25702	383	7	ercong	ercong	PROPN
fcis-25702	383	8	nie	nie	PROPN
fcis-25702	383	9	,	,	PUNCT
fcis-25702	383	10	bolei	bolei	PROPN
fcis-25702	383	11	ma	ma	PROPN
fcis-25702	383	12	,	,	PUNCT
fcis-25702	383	13	and	and	CCONJ
fcis-25702	383	14	michael	michael	PROPN
fcis-25702	383	15	färber	färber	PROPN
fcis-25702	383	16	.	.	PUNCT
fcis-25702	383	17	2024	2024	NUM
fcis-25702	383	18	.	.	PUNCT
fcis-25702	384	1	why	why	SCONJ
fcis-25702	384	2	lift	lift	VERB
fcis-25702	384	3	so	so	ADV
fcis-25702	384	4	heavy	heavy	ADJ
fcis-25702	384	5	?	?	PUNCT
fcis-25702	385	1	slimming	slimme	VERB
fcis-25702	385	2	large	large	ADJ
fcis-25702	385	3	language	language	NOUN
fcis-25702	385	4	models	model	NOUN
fcis-25702	385	5	by	by	ADP
fcis-25702	385	6	cutting	cut	VERB
fcis-25702	385	7	off	off	ADP
fcis-25702	385	8	the	the	DET
fcis-25702	385	9	layers	layer	NOUN
fcis-25702	385	10	.	.	PUNCT
fcis-25702	386	1	arxiv	arxiv	PROPN
fcis-25702	386	2	preprint	preprint	VERB
fcis-25702	386	3	arxiv:2402.11700	arxiv:2402.11700	NOUN
fcis-25702	386	4	.	.	PUNCT
fcis-25702	387	1	[	[	X
fcis-25702	387	2	35	35	NUM
fcis-25702	387	3	]	]	X
fcis-25702	387	4	wei	wei	PROPN
fcis-25702	387	5	zhu	zhu	PROPN
fcis-25702	387	6	.	.	PROPN
fcis-25702	387	7	2021	2021	NUM
fcis-25702	387	8	.	.	PUNCT
fcis-25702	388	1	leebert	leebert	PROPN
fcis-25702	388	2	:	:	PUNCT
fcis-25702	388	3	learned	learn	VERB
fcis-25702	388	4	early	early	ADJ
fcis-25702	388	5	exit	exit	NOUN
fcis-25702	388	6	for	for	ADP
fcis-25702	388	7	bert	bert	PROPN
fcis-25702	388	8	with	with	ADP
fcis-25702	388	9	crosslevel	crosslevel	NOUN
fcis-25702	388	10	optimization	optimization	NOUN
fcis-25702	388	11	.	.	PUNCT
fcis-25702	389	1	in	in	ADP
fcis-25702	389	2	proceedings	proceeding	NOUN
fcis-25702	389	3	of	of	ADP
fcis-25702	389	4	the	the	DET
fcis-25702	389	5	59th	59th	ADJ
fcis-25702	389	6	annual	annual	ADJ
fcis-25702	389	7	meeting	meeting	NOUN
fcis-25702	389	8	of	of	ADP
fcis-25702	389	9	the	the	DET
fcis-25702	389	10	association	association	NOUN
fcis-25702	389	11	for	for	ADP
fcis-25702	389	12	computational	computational	ADJ
fcis-25702	389	13	linguistics	linguistic	NOUN
fcis-25702	389	14	and	and	CCONJ
fcis-25702	389	15	the	the	DET
fcis-25702	389	16	11th	11th	NOUN
fcis-25702	389	17	internationaly	internationaly	VERB
fcis-25702	389	18	joint	joint	ADJ
fcis-25702	389	19	conference	conference	NOUN
fcis-25702	389	20	on	on	ADP
fcis-25702	389	21	natural	natural	ADJ
fcis-25702	389	22	language	language	NOUN
fcis-25702	389	23	processing	processing	NOUN
fcis-25702	389	24	(	(	PUNCT
fcis-25702	389	25	volume	volume	NOUN
fcis-25702	389	26	1	1	NUM
fcis-25702	389	27	:	:	PUNCT
fcis-25702	389	28	long	long	ADJ
fcis-25702	389	29	papers	paper	NOUN
fcis-25702	389	30	)	)	PUNCT
fcis-25702	389	31	,	,	PUNCT
fcis-25702	389	32	pages	page	NOUN
fcis-25702	389	33	2968	2968	NUM
fcis-25702	389	34	-	-	SYM
fcis-25702	389	35	2980	2980	NUM
fcis-25702	389	36	.	.	PUNCT
fcis-25702	390	1	[	[	X
fcis-25702	390	2	36	36	NUM
fcis-25702	390	3	]	]	X
fcis-25702	390	4	ari	ari	PROPN
fcis-25702	390	5	holtzman	holtzman	PROPN
fcis-25702	390	6	,	,	PUNCT
fcis-25702	390	7	jan	jan	PROPN
fcis-25702	390	8	buys	buys	PROPN
fcis-25702	390	9	,	,	PUNCT
fcis-25702	390	10	li	li	PROPN
fcis-25702	390	11	du	du	PROPN
fcis-25702	390	12	,	,	PUNCT
fcis-25702	390	13	maxwell	maxwell	PROPN
fcis-25702	390	14	forbes	forbes	PROPN
fcis-25702	390	15	,	,	PUNCT
fcis-25702	390	16	and	and	CCONJ
fcis-25702	390	17	yejin	yejin	NOUN
fcis-25702	390	18	choi	choi	NOUN
fcis-25702	390	19	.	.	PUNCT
fcis-25702	391	1	2019	2019	NUM
fcis-25702	391	2	.	.	PUNCT
fcis-25702	392	1	the	the	DET
fcis-25702	392	2	curious	curious	ADJ
fcis-25702	392	3	case	case	NOUN
fcis-25702	392	4	of	of	ADP
fcis-25702	392	5	neural	neural	ADJ
fcis-25702	392	6	text	text	NOUN
fcis-25702	392	7	degeneration	degeneration	NOUN
fcis-25702	392	8	.	.	PUNCT
fcis-25702	393	1	arxiv	arxiv	PROPN
fcis-25702	393	2	preprint	preprint	VERB
fcis-25702	393	3	arxiv:1904.09751	arxiv:1904.09751	NOUN
fcis-25702	393	4	.	.	PUNCT
fcis-25702	394	1	[	[	X
fcis-25702	394	2	37	37	NUM
fcis-25702	394	3	]	]	AUX
fcis-25702	394	4	edward	edward	PROPN
fcis-25702	394	5	j	j	PROPN
fcis-25702	394	6	hu	hu	PROPN
fcis-25702	394	7	,	,	PUNCT
fcis-25702	394	8	yelong	yelong	PROPN
fcis-25702	394	9	shen	shen	PROPN
fcis-25702	394	10	,	,	PUNCT
fcis-25702	394	11	phillip	phillip	PROPN
fcis-25702	394	12	wallis	wallis	PROPN
fcis-25702	394	13	,	,	PUNCT
fcis-25702	394	14	zeyuan	zeyuan	PROPN
fcis-25702	394	15	allen	allen	PROPN
fcis-25702	394	16	-	-	PROPN
fcis-25702	394	17	zhu	zhu	PROPN
fcis-25702	394	18	,	,	PUNCT
fcis-25702	394	19	yuanzhi	yuanzhi	PROPN
fcis-25702	394	20	li	li	PROPN
fcis-25702	394	21	,	,	PUNCT
fcis-25702	394	22	shean	shean	PROPN
fcis-25702	394	23	wang	wang	PROPN
fcis-25702	394	24	,	,	PUNCT
fcis-25702	394	25	lu	lu	PROPN
fcis-25702	394	26	wang	wang	PROPN
fcis-25702	394	27	,	,	PUNCT
fcis-25702	394	28	and	and	CCONJ
fcis-25702	394	29	weizhu	weizhu	PROPN
fcis-25702	394	30	chen	chen	PROPN
fcis-25702	394	31	.	.	PUNCT
fcis-25702	394	32	2021	2021	NUM
fcis-25702	394	33	.	.	PUNCT
fcis-25702	395	1	lora	lora	NOUN
fcis-25702	395	2	:	:	PUNCT
fcis-25702	395	3	low	low	ADJ
fcis-25702	395	4	-	-	PUNCT
fcis-25702	395	5	rank	rank	NOUN
fcis-25702	395	6	adap	adap	PROPN
fcis-25702	395	7	-	-	PUNCT
fcis-25702	395	8	tation	tation	NOUN
fcis-25702	395	9	of	of	ADP
fcis-25702	395	10	large	large	ADJ
fcis-25702	395	11	language	language	NOUN
fcis-25702	395	12	models	model	NOUN
fcis-25702	395	13	.	.	PUNCT
fcis-25702	396	1	arxiv	arxiv	PROPN
fcis-25702	396	2	preprint	preprint	PROPN
fcis-25702	396	3	arxiv:2106.09685	arxiv:2106.09685	ADJ
fcis-25702	396	4	.	.	PUNCT
fcis-25702	397	1	[	[	X
fcis-25702	397	2	38	38	NUM
fcis-25702	397	3	]	]	X
fcis-25702	397	4	yi	yi	PROPN
fcis-25702	397	5	-	-	PUNCT
fcis-25702	397	6	lin	lin	PROPN
fcis-25702	397	7	sung	sung	PROPN
fcis-25702	397	8	,	,	PUNCT
fcis-25702	397	9	jaemin	jaemin	PROPN
fcis-25702	397	10	cho	cho	PROPN
fcis-25702	397	11	,	,	PUNCT
fcis-25702	397	12	and	and	CCONJ
fcis-25702	397	13	mohit	mohit	PROPN
fcis-25702	397	14	bansal	bansal	NOUN
fcis-25702	397	15	.	.	PUNCT
fcis-25702	397	16	2022	2022	NUM
fcis-25702	397	17	.	.	PUNCT
fcis-25702	398	1	lst	lst	NOUN
fcis-25702	398	2	:	:	PUNCT
fcis-25702	398	3	ladder	ladder	VERB
fcis-25702	398	4	side	side	NOUN
fcis-25702	398	5	-	-	PUNCT
fcis-25702	398	6	tuning	tuning	NOUN
fcis-25702	398	7	for	for	ADP
fcis-25702	398	8	parameter	parameter	NOUN
fcis-25702	398	9	and	and	CCONJ
fcis-25702	398	10	memoryefficient	memoryefficient	NOUN
fcis-25702	398	11	transfer	transfer	NOUN
fcis-25702	398	12	learning	learning	NOUN
fcis-25702	398	13	.	.	PUNCT
fcis-25702	399	1	advances	advance	NOUN
fcis-25702	399	2	in	in	ADP
fcis-25702	399	3	neural	neural	ADJ
fcis-25702	399	4	information	information	NOUN
fcis-25702	399	5	processing	processing	NOUN
fcis-25702	399	6	systems	system	NOUN
fcis-25702	399	7	,	,	PUNCT
fcis-25702	399	8	35:12991	35:12991	NUM
fcis-25702	399	9	-	-	SYM
fcis-25702	399	10	13005	13005	NUM
fcis-25702	399	11	.	.	PUNCT
fcis-25702	400	1	[	[	X
fcis-25702	400	2	39	39	NUM
fcis-25702	400	3	]	]	PUNCT
fcis-25702	400	4	ronald	ronald	PROPN
fcis-25702	400	5	j	j	PROPN
fcis-25702	400	6	williams	williams	PROPN
fcis-25702	400	7	.	.	PROPN
fcis-25702	400	8	1992	1992	NUM
fcis-25702	400	9	.	.	PUNCT
fcis-25702	401	1	simple	simple	ADJ
fcis-25702	401	2	statistical	statistical	ADJ
fcis-25702	401	3	gradientfollowing	gradientfollowing	NOUN
fcis-25702	401	4	algorithms	algorithm	NOUN
fcis-25702	401	5	for	for	ADP
fcis-25702	401	6	connectionist	connectionist	NOUN
fcis-25702	401	7	reinforcement	reinforcement	NOUN
fcis-25702	401	8	learning	learning	NOUN
fcis-25702	401	9	.	.	PUNCT
fcis-25702	402	1	machine	machine	NOUN
fcis-25702	402	2	learning	learning	NOUN
fcis-25702	402	3	,	,	PUNCT
fcis-25702	402	4	8:229	8:229	NUM
fcis-25702	402	5	-	-	SYM
fcis-25702	402	6	256	256	NUM
fcis-25702	402	7	.	.	PUNCT
fcis-25702	403	1	[	[	X
fcis-25702	403	2	40	40	NUM
fcis-25702	403	3	]	]	X
fcis-25702	403	4	richard	richard	PROPN
fcis-25702	403	5	s	s	PROPN
fcis-25702	403	6	sutton	sutton	PROPN
fcis-25702	403	7	,	,	PUNCT
fcis-25702	403	8	david	david	PROPN
fcis-25702	403	9	mcallester	mcallester	PROPN
fcis-25702	403	10	,	,	PUNCT
fcis-25702	403	11	satinder	satinder	PROPN
fcis-25702	403	12	singh	singh	PROPN
fcis-25702	403	13	,	,	PUNCT
fcis-25702	403	14	and	and	CCONJ
fcis-25702	403	15	yishay	yishay	PROPN
fcis-25702	403	16	mansour	mansour	PROPN
fcis-25702	403	17	.	.	PROPN
fcis-25702	403	18	1999	1999	NUM
fcis-25702	403	19	.	.	PUNCT
fcis-25702	404	1	policy	policy	NOUN
fcis-25702	404	2	gradient	gradient	NOUN
fcis-25702	404	3	methods	method	NOUN
fcis-25702	404	4	for	for	ADP
fcis-25702	404	5	reinforcement	reinforcement	NOUN
fcis-25702	404	6	learning	learning	NOUN
fcis-25702	404	7	with	with	ADP
fcis-25702	404	8	function	function	NOUN
fcis-25702	404	9	approximation	approximation	NOUN
fcis-25702	404	10	.	.	PUNCT
fcis-25702	405	1	advances	advance	NOUN
fcis-25702	405	2	in	in	ADP
fcis-25702	405	3	neural	neural	ADJ
fcis-25702	405	4	information	information	NOUN
fcis-25702	405	5	processing	processing	NOUN
fcis-25702	405	6	systems	system	NOUN
fcis-25702	405	7	,	,	PUNCT
fcis-25702	405	8	12	12	NUM
fcis-25702	405	9	.	.	PUNCT
fcis-25702	406	1	[	[	X
fcis-25702	406	2	41	41	NUM
fcis-25702	406	3	]	]	X
fcis-25702	406	4	tim	tim	PROPN
fcis-25702	406	5	dettmers	dettmers	PROPN
fcis-25702	406	6	,	,	PUNCT
fcis-25702	406	7	artidoro	artidoro	PROPN
fcis-25702	406	8	pagnoni	pagnoni	PROPN
fcis-25702	406	9	,	,	PUNCT
fcis-25702	406	10	ari	ari	PROPN
fcis-25702	406	11	holtzman	holtzman	NOUN
fcis-25702	406	12	,	,	PUNCT
fcis-25702	406	13	and	and	CCONJ
fcis-25702	406	14	luke	luke	PROPN
fcis-25702	406	15	zettlemoyer	zettlemoyer	NOUN
fcis-25702	406	16	.	.	PUNCT
fcis-25702	407	1	2024	2024	NUM
fcis-25702	407	2	.	.	PUNCT
fcis-25702	408	1	qlora	qlora	ADJ
fcis-25702	408	2	:	:	PUNCT
fcis-25702	408	3	efficient	efficient	ADJ
fcis-25702	408	4	finetuning	finetuning	NOUN
fcis-25702	408	5	of	of	ADP
fcis-25702	408	6	quantized	quantize	VERB
fcis-25702	408	7	llms	llm	NOUN
fcis-25702	408	8	.	.	PUNCT
fcis-25702	409	1	advances	advance	NOUN
fcis-25702	409	2	in	in	ADP
fcis-25702	409	3	neural	neural	ADJ
fcis-25702	409	4	information	information	NOUN
fcis-25702	409	5	processing	processing	NOUN
fcis-25702	409	6	systems	system	NOUN
fcis-25702	409	7	,	,	PUNCT
fcis-25702	409	8	36	36	NUM
fcis-25702	409	9	.	.	PUNCT
fcis-25702	410	1	[	[	X
fcis-25702	410	2	42	42	NUM
fcis-25702	410	3	]	]	PUNCT
fcis-25702	410	4	marc’aurelio	marc’aurelio	NOUN
fcis-25702	410	5	ranzato	ranzato	NOUN
fcis-25702	410	6	,	,	PUNCT
fcis-25702	410	7	sumit	sumit	PROPN
fcis-25702	410	8	chopra	chopra	PROPN
fcis-25702	410	9	,	,	PUNCT
fcis-25702	410	10	michael	michael	PROPN
fcis-25702	410	11	auli	auli	PROPN
fcis-25702	410	12	,	,	PUNCT
fcis-25702	410	13	and	and	CCONJ
fcis-25702	410	14	wojciech	wojciech	PROPN
fcis-25702	410	15	zaremba	zaremba	NOUN
fcis-25702	410	16	.	.	PUNCT
fcis-25702	411	1	2015	2015	NUM
fcis-25702	411	2	.	.	PUNCT
fcis-25702	412	1	sequence	sequence	NOUN
fcis-25702	412	2	level	level	NOUN
fcis-25702	412	3	train	train	NOUN
fcis-25702	412	4	ing	e	VERB
fcis-25702	412	5	with	with	ADP
fcis-25702	412	6	recurrent	recurrent	ADJ
fcis-25702	412	7	neural	neural	ADJ
fcis-25702	412	8	networks	network	NOUN
fcis-25702	412	9	.	.	PUNCT
fcis-25702	413	1	arxiv	arxiv	PROPN
fcis-25702	413	2	preprint	preprint	VERB
fcis-25702	413	3	arxiv:1511.06732	arxiv:1511.06732	PROPN
fcis-25702	413	4	.	.	PUNCT
fcis-25702	414	1	[	[	X
fcis-25702	414	2	43	43	NUM
fcis-25702	414	3	]	]	X
fcis-25702	414	4	alex	alex	PROPN
fcis-25702	414	5	wang	wang	PROPN
fcis-25702	414	6	,	,	PUNCT
fcis-25702	414	7	amanpreet	amanpreet	VERB
fcis-25702	414	8	singh	singh	PROPN
fcis-25702	414	9	,	,	PUNCT
fcis-25702	414	10	julian	julian	PROPN
fcis-25702	414	11	michael	michael	PROPN
fcis-25702	414	12	,	,	PUNCT
fcis-25702	414	13	felix	felix	PROPN
fcis-25702	414	14	hill	hill	PROPN
fcis-25702	414	15	,	,	PUNCT
fcis-25702	414	16	omer	omer	PROPN
fcis-25702	414	17	levy	levy	PROPN
fcis-25702	414	18	,	,	PUNCT
fcis-25702	414	19	and	and	CCONJ
fcis-25702	414	20	samuel	samuel	PROPN
fcis-25702	414	21	r	r	PROPN
fcis-25702	414	22	bowman	bowman	PROPN
fcis-25702	414	23	.	.	PROPN
fcis-25702	414	24	2018	2018	NUM
fcis-25702	414	25	.	.	PUNCT
fcis-25702	415	1	glue	glue	NOUN
fcis-25702	415	2	:	:	PUNCT
fcis-25702	415	3	a	a	DET
fcis-25702	415	4	multi	multi	ADJ
fcis-25702	415	5	-	-	ADJ
fcis-25702	415	6	task	task	ADJ
fcis-25702	415	7	benchmark	benchmark	NOUN
fcis-25702	415	8	and	and	CCONJ
fcis-25702	415	9	analysis	analysis	NOUN
fcis-25702	415	10	platformfor	platformfor	ADP
fcis-25702	415	11	natural	natural	ADJ
fcis-25702	415	12	language	language	NOUN
fcis-25702	415	13	understanding	understanding	NOUN
fcis-25702	415	14	.	.	PUNCT
fcis-25702	416	1	arxiv	arxiv	PROPN
fcis-25702	416	2	preprint	preprint	PROPN
fcis-25702	416	3	arxiv:1804.07461	arxiv:1804.07461	NOUN
fcis-25702	416	4	.	.	PUNCT
fcis-25702	417	1	[	[	X
fcis-25702	417	2	44	44	NUM
fcis-25702	417	3	]	]	PUNCT
fcis-25702	417	4	thomas	thomas	PROPN
fcis-25702	417	5	wolf	wolf	PROPN
fcis-25702	417	6	,	,	PUNCT
fcis-25702	417	7	lysandre	lysandre	PROPN
fcis-25702	417	8	debut	debut	PROPN
fcis-25702	417	9	,	,	PUNCT
fcis-25702	417	10	victor	victor	PROPN
fcis-25702	417	11	sanh	sanh	PROPN
fcis-25702	417	12	,	,	PUNCT
fcis-25702	417	13	julien	julien	PROPN
fcis-25702	417	14	chaumond	chaumond	PROPN
fcis-25702	417	15	,	,	PUNCT
fcis-25702	417	16	clement	clement	PROPN
fcis-25702	417	17	delangue	delangue	PROPN
fcis-25702	417	18	,	,	PUNCT
fcis-25702	417	19	anthony	anthony	PROPN
fcis-25702	417	20	moi	moi	PROPN
fcis-25702	417	21	,	,	PUNCT
fcis-25702	417	22	pierric	pierric	ADJ
fcis-25702	417	23	cistac	cistac	NOUN
fcis-25702	417	24	,	,	PUNCT
fcis-25702	417	25	tim	tim	PROPN
fcis-25702	417	26	rault	rault	NOUN
fcis-25702	417	27	,	,	PUNCT
fcis-25702	417	28	rémi	rémi	NOUN
fcis-25702	417	29	louf	louf	PROPN
fcis-25702	417	30	,	,	PUNCT
fcis-25702	417	31	morgan	morgan	PROPN
fcis-25702	417	32	funtowicz	funtowicz	PROPN
fcis-25702	417	33	,	,	PUNCT
fcis-25702	417	34	et	et	PROPN
fcis-25702	417	35	al	al	PROPN
fcis-25702	417	36	.	.	PROPN
fcis-25702	417	37	2020	2020	NUM
fcis-25702	417	38	.	.	PUNCT
fcis-25702	418	1	transformers	transformer	NOUN
fcis-25702	418	2	:	:	PUNCT
fcis-25702	418	3	state	state	NOUN
fcis-25702	418	4	-	-	PUNCT
fcis-25702	418	5	of	of	ADP
fcis-25702	418	6	-	-	PUNCT
fcis-25702	418	7	the	the	DET
fcis-25702	418	8	-	-	PUNCT
fcis-25702	418	9	art	art	NOUN
fcis-25702	418	10	natural	natural	ADJ
fcis-25702	418	11	language	language	NOUN
fcis-25702	418	12	processing	processing	NOUN
fcis-25702	418	13	.	.	PUNCT
fcis-25702	419	1	in	in	ADP
fcis-25702	419	2	proceedings	proceeding	NOUN
fcis-25702	419	3	of	of	ADP
fcis-25702	419	4	the	the	DET
fcis-25702	419	5	2020	2020	NUM
fcis-25702	419	6	con	con	NOUN
fcis-25702	419	7	ference	ference	NOUN
fcis-25702	419	8	on	on	ADP
fcis-25702	419	9	empirical	empirical	ADJ
fcis-25702	419	10	methods	method	NOUN
fcis-25702	419	11	in	in	ADP
fcis-25702	419	12	natural	natural	ADJ
fcis-25702	419	13	language	language	NOUN
fcis-25702	419	14	processing	processing	NOUN
fcis-25702	419	15	:	:	PUNCT
fcis-25702	419	16	system	system	NOUN
fcis-25702	419	17	demonstrations	demonstration	NOUN
fcis-25702	419	18	,	,	PUNCT
fcis-25702	419	19	pages	page	NOUN
fcis-25702	419	20	38	38	NUM
fcis-25702	419	21	-	-	SYM
fcis-25702	419	22	45	45	NUM
fcis-25702	419	23	.	.	PUNCT
fcis-25702	420	1	[	[	X
fcis-25702	420	2	45	45	NUM
fcis-25702	420	3	]	]	PUNCT
fcis-25702	420	4	alec	alec	PROPN
fcis-25702	420	5	radford	radford	PROPN
fcis-25702	420	6	,	,	PUNCT
fcis-25702	420	7	jeffrey	jeffrey	PROPN
fcis-25702	420	8	wu	wu	PROPN
fcis-25702	420	9	,	,	PUNCT
fcis-25702	420	10	rewon	rewon	ADJ
fcis-25702	420	11	child	child	NOUN
fcis-25702	420	12	,	,	PUNCT
fcis-25702	420	13	david	david	PROPN
fcis-25702	420	14	luan	luan	PROPN
fcis-25702	420	15	,	,	PUNCT
fcis-25702	420	16	dario	dario	PROPN
fcis-25702	420	17	amodei	amodei	PROPN
fcis-25702	420	18	,	,	PUNCT
fcis-25702	420	19	ilya	ilya	PROPN
fcis-25702	420	20	sutskever	sutskever	PROPN
fcis-25702	420	21	,	,	PUNCT
fcis-25702	420	22	et	et	PROPN
fcis-25702	420	23	al	al	PROPN
fcis-25702	420	24	.	.	PROPN
fcis-25702	420	25	2019	2019	NUM
fcis-25702	420	26	.	.	PUNCT
fcis-25702	421	1	language	language	NOUN
fcis-25702	421	2	models	model	NOUN
fcis-25702	421	3	are	be	AUX
fcis-25702	421	4	unsupervised	unsupervised	ADJ
fcis-25702	421	5	multitask	multitask	ADJ
fcis-25702	421	6	learners	learner	NOUN
fcis-25702	421	7	.	.	PUNCT
fcis-25702	422	1	openai	openai	PROPN
fcis-25702	422	2	blog	blog	PROPN
fcis-25702	422	3	,	,	PUNCT
fcis-25702	422	4	1(8):9	1(8):9	PROPN
fcis-25702	422	5	.	.	PUNCT
fcis-25702	423	1	[	[	X
fcis-25702	423	2	46	46	NUM
fcis-25702	423	3	]	]	X
fcis-25702	423	4	susan	susan	PROPN
fcis-25702	423	5	zhang	zhang	PROPN
fcis-25702	423	6	,	,	PUNCT
fcis-25702	423	7	stephen	stephen	PROPN
fcis-25702	423	8	roller	roller	PROPN
fcis-25702	423	9	,	,	PUNCT
fcis-25702	423	10	naman	naman	NOUN
fcis-25702	423	11	goyal	goyal	PROPN
fcis-25702	423	12	,	,	PUNCT
fcis-25702	423	13	mikel	mikel	PROPN
fcis-25702	423	14	artetxe	artetxe	PROPN
fcis-25702	423	15	,	,	PUNCT
fcis-25702	423	16	moya	moya	PROPN
fcis-25702	423	17	chen	chen	PROPN
fcis-25702	423	18	,	,	PUNCT
fcis-25702	423	19	shuohui	shuohui	PROPN
fcis-25702	423	20	chen	chen	PROPN
fcis-25702	423	21	,	,	PUNCT
fcis-25702	423	22	christopher	christopher	PROPN
fcis-25702	423	23	dewan	dewan	PROPN
fcis-25702	423	24	,	,	PUNCT
fcis-25702	423	25	mona	mona	PROPN
fcis-25702	423	26	diab	diab	PROPN
fcis-25702	423	27	,	,	PUNCT
fcis-25702	423	28	xian	xian	PROPN
fcis-25702	423	29	li	li	PROPN
fcis-25702	423	30	,	,	PUNCT
fcis-25702	423	31	xi	xi	PROPN
fcis-25702	423	32	victoria	victoria	PROPN
fcis-25702	423	33	lin	lin	PROPN
fcis-25702	423	34	,	,	PUNCT
fcis-25702	423	35	et	et	PROPN
fcis-25702	423	36	al	al	PROPN
fcis-25702	423	37	.	.	PROPN
fcis-25702	423	38	2022	2022	NUM
fcis-25702	423	39	.	.	PUNCT
fcis-25702	423	40	opt	opt	PROPN
fcis-25702	423	41	:	:	PUNCT
fcis-25702	423	42	open	open	ADJ
fcis-25702	423	43	pre	pre	ADJ
fcis-25702	423	44	-	-	ADJ
fcis-25702	423	45	trained	train	VERB
fcis-25702	423	46	transformer	transformer	NOUN
fcis-25702	423	47	language	language	NOUN
fcis-25702	423	48	models	model	NOUN
fcis-25702	423	49	.	.	PUNCT
fcis-25702	424	1	arxiv	arxiv	PROPN
fcis-25702	424	2	preprint	preprint	NOUN
fcis-25702	424	3	arxiv:2205	arxiv:2205	PROPN
fcis-25702	424	4	.	.	PROPN
fcis-25702	424	5	01068	01068	NUM
fcis-25702	424	6	.	.	PUNCT
fcis-25702	425	1	[	[	X
fcis-25702	425	2	47	47	NUM
fcis-25702	425	3	]	]	X
fcis-25702	425	4	ebtesam	ebtesam	PROPN
fcis-25702	425	5	almazrouei	almazrouei	PROPN
fcis-25702	425	6	,	,	PUNCT
fcis-25702	425	7	hamza	hamza	PROPN
fcis-25702	425	8	alobeidli	alobeidli	PROPN
fcis-25702	425	9	,	,	PUNCT
fcis-25702	425	10	abdulaziz	abdulaziz	PROPN
fcis-25702	425	11	al	al	PROPN
fcis-25702	425	12	shamsi	shamsi	PROPN
fcis-25702	425	13	,	,	PUNCT
fcis-25702	425	14	alessandro	alessandro	PROPN
fcis-25702	425	15	cappelli	cappelli	PROPN
fcis-25702	425	16	,	,	PUNCT
fcis-25702	425	17	ruxandra	ruxandra	PROPN
fcis-25702	425	18	cojocaru	cojocaru	PROPN
fcis-25702	425	19	,	,	PUNCT
fcis-25702	425	20	mérouane	mérouane	NOUN
fcis-25702	425	21	debbah	debbah	PROPN
fcis-25702	425	22	,	,	PUNCT
fcis-25702	425	23	étienne	étienne	PROPN
fcis-25702	425	24	goffinet	goffinet	PROPN
fcis-25702	425	25	,	,	PUNCT
fcis-25702	425	26	daniel	daniel	PROPN
fcis-25702	425	27	hesslow	hesslow	PROPN
fcis-25702	425	28	,	,	PUNCT
fcis-25702	425	29	julien	julien	PROPN
fcis-25702	425	30	launay	launay	PROPN
fcis-25702	425	31	,	,	PUNCT
fcis-25702	425	32	quentin	quentin	PROPN
fcis-25702	425	33	malartic	malartic	PROPN
fcis-25702	425	34	,	,	PUNCT
fcis-25702	425	35	et	et	PROPN
fcis-25702	425	36	al	al	PROPN
fcis-25702	425	37	.	.	PROPN
fcis-25702	425	38	2023	2023	NUM
fcis-25702	425	39	.	.	PUNCT
fcis-25702	426	1	the	the	DET
fcis-25702	426	2	falcon	falcon	PROPN
fcis-25702	426	3	series	series	NOUN
fcis-25702	426	4	of	of	ADP
fcis-25702	426	5	open	open	ADJ
fcis-25702	426	6	language	language	NOUN
fcis-25702	426	7	models	model	NOUN
fcis-25702	426	8	.	.	PUNCT
fcis-25702	427	1	arxiv	arxiv	PROPN
fcis-25702	427	2	preprint	preprint	VERB
fcis-25702	427	3	arxiv:2311.16867	arxiv:2311.16867	PRON
fcis-25702	427	4	.	.	PUNCT
fcis-25702	428	1	[	[	X
fcis-25702	428	2	48	48	NUM
fcis-25702	428	3	]	]	X
fcis-25702	428	4	mario	mario	PROPN
fcis-25702	428	5	michael	michael	PROPN
fcis-25702	428	6	krell	krell	PROPN
fcis-25702	428	7	,	,	PUNCT
fcis-25702	428	8	matej	matej	PROPN
fcis-25702	428	9	kosec	kosec	PROPN
fcis-25702	428	10	,	,	PUNCT
fcis-25702	428	11	sergio	sergio	PROPN
fcis-25702	428	12	p	p	PROPN
fcis-25702	428	13	perez	perez	PROPN
fcis-25702	428	14	,	,	PUNCT
fcis-25702	428	15	and	and	CCONJ
fcis-25702	428	16	andrew	andrew	PROPN
fcis-25702	428	17	fitzgibbon	fitzgibbon	PROPN
fcis-25702	428	18	.	.	PUNCT
fcis-25702	429	1	2021	2021	NUM
fcis-25702	429	2	.	.	PUNCT
fcis-25702	430	1	efficient	efficient	ADJ
fcis-25702	430	2	sequence	sequence	NOUN
fcis-25702	430	3	packing	pack	VERB
fcis-25702	430	4	without	without	ADP
fcis-25702	430	5	cross	cross	NOUN
fcis-25702	430	6	-	-	NOUN
fcis-25702	430	7	contamination	contamination	NOUN
fcis-25702	430	8	:	:	PUNCT
fcis-25702	430	9	accelerating	accelerate	VERB
fcis-25702	430	10	large	large	ADJ
fcis-25702	430	11	language	language	NOUN
fcis-25702	430	12	models	model	NOUN
fcis-25702	430	13	without	without	ADP
fcis-25702	430	14	impacting	impact	VERB
fcis-25702	430	15	performance	performance	NOUN
fcis-25702	430	16	.	.	PUNCT
fcis-25702	431	1	arxiv	arxiv	PROPN
fcis-25702	431	2	preprint	preprint	PROPN
fcis-25702	431	3	arxiv:2107.02027	arxiv:2107.02027	NOUN
fcis-25702	431	4	.	.	PUNCT
fcis-25702	432	1	[	[	X
fcis-25702	432	2	49	49	NUM
fcis-25702	432	3	]	]	PUNCT
fcis-25702	432	4	ilya	ilya	PROPN
fcis-25702	432	5	loshchilov	loshchilov	PROPN
fcis-25702	432	6	and	and	CCONJ
fcis-25702	432	7	frank	frank	PROPN
fcis-25702	432	8	hutter	hutter	NOUN
fcis-25702	432	9	.	.	PUNCT
fcis-25702	433	1	2017	2017	NUM
fcis-25702	433	2	.	.	PUNCT
fcis-25702	434	1	decou	decou	NOUN
fcis-25702	434	2	pled	plead	VERB
fcis-25702	434	3	weight	weight	NOUN
fcis-25702	434	4	decay	decay	NOUN
fcis-25702	434	5	regularization	regularization	NOUN
fcis-25702	434	6	.	.	PUNCT
fcis-25702	435	1	arxiv	arxiv	PROPN
fcis-25702	435	2	preprint	preprint	PROPN
fcis-25702	435	3	arxiv:1711.05101	arxiv:1711.05101	PROPN
fcis-25702	435	4	.	.	PUNCT
fcis-25702	436	1	[	[	X
fcis-25702	436	2	50	50	NUM
fcis-25702	436	3	]	]	X
fcis-25702	436	4	yifei	yifei	PROPN
fcis-25702	436	5	yang	yang	PROPN
fcis-25702	436	6	,	,	PUNCT
fcis-25702	436	7	zouying	zouye	VERB
fcis-25702	436	8	cao	cao	PROPN
fcis-25702	436	9	,	,	PUNCT
fcis-25702	436	10	and	and	CCONJ
fcis-25702	436	11	hai	hai	PROPN
fcis-25702	436	12	zhao	zhao	PROPN
fcis-25702	436	13	.	.	PUNCT
fcis-25702	437	1	2024	2024	NUM
fcis-25702	437	2	.	.	PUNCT
fcis-25702	438	1	laco	laco	PROPN
fcis-25702	438	2	:	:	PUNCT
fcis-25702	438	3	large	large	ADJ
fcis-25702	438	4	language	language	NOUN
fcis-25702	438	5	model	model	NOUN
fcis-25702	438	6	pruning	prune	VERB
fcis-25702	438	7	via	via	ADP
fcis-25702	438	8	layer	layer	NOUN
fcis-25702	438	9	collapse	collapse	NOUN
fcis-25702	438	10	.	.	PUNCT
fcis-25702	439	1	arxiv	arxiv	PROPN
fcis-25702	439	2	preprint	preprint	VERB
fcis-25702	439	3	arxiv:2402.11187	arxiv:2402.11187	NOUN
fcis-25702	439	4	.	.	PUNCT
fcis-25702	440	1	[	[	X
fcis-25702	440	2	51	51	NUM
fcis-25702	440	3	]	]	X
fcis-25702	440	4	koyel	koyel	PROPN
fcis-25702	440	5	chakraborty	chakraborty	PROPN
fcis-25702	440	6	,	,	PUNCT
fcis-25702	440	7	siddhartha	siddhartha	PROPN
fcis-25702	440	8	bhattacharyya	bhattacharyya	PROPN
fcis-25702	440	9	,	,	PUNCT
fcis-25702	440	10	and	and	CCONJ
fcis-25702	440	11	rajib	rajib	NOUN
fcis-25702	440	12	bag	bag	NOUN
fcis-25702	440	13	.	.	PUNCT
fcis-25702	441	1	2020	2020	NUM
fcis-25702	441	2	.	.	PUNCT
fcis-25702	442	1	a	a	DET
fcis-25702	442	2	survey	survey	NOUN
fcis-25702	442	3	of	of	ADP
fcis-25702	442	4	sentiment	sentiment	NOUN
fcis-25702	442	5	analysis	analysis	NOUN
fcis-25702	442	6	from	from	ADP
fcis-25702	442	7	social	social	ADJ
fcis-25702	442	8	media	medium	NOUN
fcis-25702	442	9	data	datum	NOUN
fcis-25702	442	10	.	.	PUNCT
fcis-25702	443	1	ieee	ieee	NOUN
fcis-25702	443	2	transactions	transaction	NOUN
fcis-25702	443	3	on	on	ADP
fcis-25702	443	4	computational	computational	ADJ
fcis-25702	443	5	social	social	ADJ
fcis-25702	443	6	systems	system	NOUN
fcis-25702	443	7	,	,	PUNCT
fcis-25702	443	8	7(2	7(2	NUM
fcis-25702	443	9	):	):	PUNCT
fcis-25702	443	10	450	450	NUM
fcis-25702	443	11	-	-	SYM
fcis-25702	443	12	464	464	NUM
fcis-25702	443	13	.	.	PUNCT
fcis-25702	444	1	[	[	X
fcis-25702	444	2	52	52	NUM
fcis-25702	444	3	]	]	X
fcis-25702	444	4	weiwei	weiwei	PROPN
fcis-25702	444	5	sun	sun	PROPN
fcis-25702	444	6	,	,	PUNCT
fcis-25702	444	7	lingyong	lingyong	PROPN
fcis-25702	444	8	yan	yan	PROPN
fcis-25702	444	9	,	,	PUNCT
fcis-25702	444	10	xinyu	xinyu	PROPN
fcis-25702	444	11	ma	ma	PROPN
fcis-25702	444	12	,	,	PUNCT
fcis-25702	444	13	pengjie	pengjie	PROPN
fcis-25702	444	14	ren	ren	PROPN
fcis-25702	444	15	,	,	PUNCT
fcis-25702	444	16	dawei	dawei	PROPN
fcis-25702	444	17	yin	yin	PROPN
fcis-25702	444	18	,	,	PUNCT
fcis-25702	444	19	and	and	CCONJ
fcis-25702	444	20	zhaochun	zhaochun	PROPN
fcis-25702	444	21	ren	ren	PROPN
fcis-25702	444	22	.	.	PUNCT
fcis-25702	444	23	2023	2023	NUM
fcis-25702	444	24	.	.	PUNCT
fcis-25702	444	25	is	be	AUX
fcis-25702	444	26	chatgpt	chatgpt	NOUN
fcis-25702	444	27	good	good	ADJ
fcis-25702	444	28	at	at	ADP
fcis-25702	444	29	search	search	NOUN
fcis-25702	444	30	?	?	PUNCT
fcis-25702	445	1	investigating	investigate	VERB
fcis-25702	445	2	large	large	ADJ
fcis-25702	445	3	lan	lan	NOUN
fcis-25702	445	4	guage	guage	NOUN
fcis-25702	445	5	models	model	NOUN
fcis-25702	445	6	as	as	ADP
fcis-25702	445	7	re	re	VERB
fcis-25702	445	8	-	-	ADJ
fcis-25702	445	9	ranking	ranking	ADJ
fcis-25702	445	10	agent	agent	NOUN
fcis-25702	445	11	.	.	PUNCT
fcis-25702	446	1	arxiv	arxiv	PROPN
fcis-25702	446	2	preprint	preprint	VERB
fcis-25702	446	3	arxiv:2304.09542	arxiv:2304.09542	ADV
fcis-25702	446	4	.	.	PUNCT
fcis-25702	447	1	[	[	X
fcis-25702	447	2	53	53	NUM
fcis-25702	447	3	]	]	X
fcis-25702	447	4	qingxiu	qingxiu	PROPN
fcis-25702	447	5	dong	dong	PROPN
fcis-25702	447	6	,	,	PUNCT
fcis-25702	447	7	lei	lei	PROPN
fcis-25702	447	8	li	li	PROPN
fcis-25702	447	9	,	,	PUNCT
fcis-25702	447	10	damai	damai	PROPN
fcis-25702	447	11	dai	dai	PROPN
fcis-25702	447	12	,	,	PUNCT
fcis-25702	447	13	ce	ce	PROPN
fcis-25702	447	14	zheng	zheng	PROPN
fcis-25702	447	15	,	,	PUNCT
fcis-25702	447	16	zhiyong	zhiyong	PROPN
fcis-25702	447	17	wu	wu	PROPN
fcis-25702	447	18	,	,	PUNCT
fcis-25702	447	19	baobao	baobao	PROPN
fcis-25702	447	20	chang	chang	PROPN
fcis-25702	447	21	,	,	PUNCT
fcis-25702	447	22	xu	xu	PROPN
fcis-25702	447	23	sun	sun	PROPN
fcis-25702	447	24	,	,	PUNCT
fcis-25702	447	25	jingjing	jingjing	PROPN
fcis-25702	447	26	xu	xu	PROPN
fcis-25702	447	27	,	,	PUNCT
fcis-25702	447	28	and	and	CCONJ
fcis-25702	447	29	zhifang	zhifang	PROPN
fcis-25702	447	30	sui	sui	PROPN
fcis-25702	447	31	.	.	PROPN
fcis-25702	447	32	2022	2022	NUM
fcis-25702	447	33	.	.	PUNCT
fcis-25702	448	1	a	a	DET
fcis-25702	448	2	survey	survey	NOUN
fcis-25702	448	3	on	on	ADP
fcis-25702	448	4	in	in	ADP
fcis-25702	448	5	-	-	PUNCT
fcis-25702	448	6	context	context	NOUN
fcis-25702	448	7	learning	learning	NOUN
fcis-25702	448	8	.	.	PUNCT
fcis-25702	449	1	arxiv	arxiv	PROPN
fcis-25702	449	2	preprint	preprint	PROPN
fcis-25702	449	3	arxiv	arxiv	PROPN
fcis-25702	449	4	:	:	PUNCT
fcis-25702	449	5	2301	2301	NUM
fcis-25702	449	6	.	.	PUNCT
fcis-25702	449	7	00234	00234	NUM
fcis-25702	449	8	.	.	PUNCT
fcis-25702	450	1	appendices	appendix	NOUN
fcis-25702	450	2	a.1	a.1	NOUN
fcis-25702	450	3	using	use	VERB
fcis-25702	450	4	llms	llm	NOUN
fcis-25702	450	5	for	for	ADP
fcis-25702	450	6	nlu	nlu	PROPN
fcis-25702	450	7	task	task	PROPN
fcis-25702	450	8	nlu	nlu	NOUN
fcis-25702	450	9	problems	problem	NOUN
fcis-25702	450	10	can	can	AUX
fcis-25702	450	11	all	all	PRON
fcis-25702	450	12	be	be	AUX
fcis-25702	450	13	viewed	view	VERB
fcis-25702	450	14	as	as	ADP
fcis-25702	450	15	classification	classification	NOUN
fcis-25702	450	16	problems	problem	NOUN
fcis-25702	450	17	.	.	PUNCT
fcis-25702	451	1	applying	apply	VERB
fcis-25702	451	2	llms	llm	NOUN
fcis-25702	451	3	to	to	ADP
fcis-25702	451	4	classification	classification	NOUN
fcis-25702	451	5	tasks	task	NOUN
fcis-25702	451	6	is	be	AUX
fcis-25702	451	7	especially	especially	ADV
fcis-25702	451	8	important	important	ADJ
fcis-25702	451	9	for	for	ADP
fcis-25702	451	10	applications	application	NOUN
fcis-25702	451	11	.	.	PUNCT
fcis-25702	452	1	first	first	ADV
fcis-25702	452	2	,	,	PUNCT
fcis-25702	452	3	classification	classification	NOUN
fcis-25702	452	4	tasks	task	NOUN
fcis-25702	452	5	are	be	AUX
fcis-25702	452	6	widely	widely	ADV
fcis-25702	452	7	used	use	VERB
fcis-25702	452	8	in	in	ADP
fcis-25702	452	9	many	many	ADJ
fcis-25702	452	10	real	real	ADJ
fcis-25702	452	11	-	-	PUNCT
fcis-25702	452	12	life	life	NOUN
fcis-25702	452	13	scenarios	scenario	NOUN
fcis-25702	452	14	.	.	PUNCT
fcis-25702	453	1	for	for	ADP
fcis-25702	453	2	example	example	NOUN
fcis-25702	453	3	,	,	PUNCT
fcis-25702	453	4	conducting	conduct	VERB
fcis-25702	453	5	sentiment	sentiment	NOUN
fcis-25702	453	6	polarity	polarity	NOUN
fcis-25702	453	7	analysis	analysis	NOUN
fcis-25702	453	8	on	on	ADP
fcis-25702	453	9	social	social	ADJ
fcis-25702	453	10	media	medium	NOUN
fcis-25702	453	11	comments	comment	NOUN
fcis-25702	453	12	,	,	PUNCT
fcis-25702	453	13	product	product	NOUN
fcis-25702	453	14	reviews	review	NOUN
fcis-25702	453	15	,	,	PUNCT
fcis-25702	453	16	etc	etc	X
fcis-25702	453	17	.	.	X
fcis-25702	453	18	,	,	PUNCT
fcis-25702	453	19	helps	helps	AUX
fcis-25702	453	20	understand	understand	VERB
fcis-25702	453	21	users	user	NOUN
fcis-25702	453	22	’	’	PART
fcis-25702	453	23	attitudes	attitude	NOUN
fcis-25702	453	24	and	and	CCONJ
fcis-25702	453	25	emotional	emotional	ADJ
fcis-25702	453	26	tendencies	tendency	NOUN
fcis-25702	453	27	towards	towards	ADP
fcis-25702	453	28	products	product	NOUN
fcis-25702	453	29	,	,	PUNCT
fcis-25702	453	30	services	service	NOUN
fcis-25702	453	31	,	,	PUNCT
fcis-25702	453	32	or	or	CCONJ
fcis-25702	453	33	events	event	NOUN
fcis-25702	453	34	[	[	X
fcis-25702	453	35	51	51	NUM
fcis-25702	453	36	]	]	PUNCT
fcis-25702	453	37	.	.	PUNCT
fcis-25702	454	1	additionally	additionally	ADV
fcis-25702	454	2	,	,	PUNCT
fcis-25702	454	3	classifying	classify	VERB
fcis-25702	454	4	the	the	DET
fcis-25702	454	5	relation	relation	NOUN
fcis-25702	454	6	or	or	CCONJ
fcis-25702	454	7	similarity	similarity	NOUN
fcis-25702	454	8	of	of	ADP
fcis-25702	454	9	two	two	NUM
fcis-25702	454	10	sentences	sentence	NOUN
fcis-25702	454	11	aids	aid	VERB
fcis-25702	454	12	information	information	NOUN
fcis-25702	454	13	retrieval	retrieval	NOUN
fcis-25702	454	14	[	[	X
fcis-25702	454	15	52	52	NUM
fcis-25702	454	16	]	]	PUNCT
fcis-25702	454	17	.	.	PUNCT
fcis-25702	455	1	second	second	ADJ
fcis-25702	455	2	,	,	PUNCT
fcis-25702	455	3	although	although	SCONJ
fcis-25702	455	4	llms	llm	NOUN
fcis-25702	455	5	are	be	AUX
fcis-25702	455	6	powerful	powerful	ADJ
fcis-25702	455	7	,	,	PUNCT
fcis-25702	455	8	they	they	PRON
fcis-25702	455	9	require	require	VERB
fcis-25702	455	10	intermediate	intermediate	ADJ
fcis-25702	455	11	steps	step	NOUN
fcis-25702	455	12	to	to	PART
fcis-25702	455	13	better	well	ADV
fcis-25702	455	14	complete	complete	ADJ
fcis-25702	455	15	tasks	task	NOUN
fcis-25702	455	16	,	,	PUNCT
fcis-25702	455	17	such	such	ADJ
fcis-25702	455	18	as	as	ADP
fcis-25702	455	19	demonstrations	demonstration	NOUN
fcis-25702	455	20	[	[	X
fcis-25702	455	21	53	53	NUM
fcis-25702	455	22	]	]	PUNCT
fcis-25702	455	23	and	and	CCONJ
fcis-25702	455	24	reasoning	reason	VERB
fcis-25702	455	25	prompts	prompt	NOUN
fcis-25702	455	26	[	[	X
fcis-25702	455	27	10	10	NUM
fcis-25702	455	28	]	]	PUNCT
fcis-25702	455	29	.	.	PUNCT
fcis-25702	456	1	chae	chae	NOUN
fcis-25702	456	2	and	and	CCONJ
fcis-25702	456	3	davidson	davidson	NOUN
fcis-25702	457	1	[	[	X
fcis-25702	457	2	14	14	NUM
fcis-25702	457	3	]	]	PUNCT
fcis-25702	457	4	uses	use	VERB
fcis-25702	457	5	instructionfollowing	instructionfollowe	VERB
fcis-25702	457	6	datasets	dataset	NOUN
fcis-25702	457	7	to	to	PART
fcis-25702	457	8	finetune	finetune	VERB
fcis-25702	457	9	llms	llm	NOUN
fcis-25702	457	10	for	for	ADP
fcis-25702	457	11	classification	classification	NOUN
fcis-25702	457	12	tasks	task	NOUN
fcis-25702	457	13	.	.	PUNCT
fcis-25702	458	1	in	in	ADP
fcis-25702	458	2	our	our	PRON
fcis-25702	458	3	work	work	NOUN
fcis-25702	458	4	,	,	PUNCT
fcis-25702	458	5	we	we	PRON
fcis-25702	458	6	fine	fine	ADJ
fcis-25702	458	7	-	-	PUNCT
fcis-25702	458	8	tune	tune	NOUN
fcis-25702	458	9	llms	llm	NOUN
fcis-25702	458	10	with	with	ADP
fcis-25702	458	11	prompts	prompt	NOUN
fcis-25702	458	12	to	to	PART
fcis-25702	458	13	enhance	enhance	VERB
fcis-25702	458	14	their	their	PRON
fcis-25702	458	15	performance	performance	NOUN
fcis-25702	458	16	.	.	PUNCT
fcis-25702	459	1	a.2	a.2	VERB
fcis-25702	459	2	prompt	prompt	ADJ
fcis-25702	459	3	design	design	NOUN
fcis-25702	459	4	for	for	ADP
fcis-25702	459	5	nlu	nlu	NOUN
fcis-25702	459	6	datasets	dataset	NOUN
fcis-25702	459	7	here	here	ADV
fcis-25702	459	8	we	we	PRON
fcis-25702	459	9	provide	provide	VERB
fcis-25702	459	10	the	the	DET
fcis-25702	459	11	prompt	prompt	NOUN
fcis-25702	459	12	we	we	PRON
fcis-25702	459	13	used	use	VERB
fcis-25702	459	14	for	for	ADP
fcis-25702	459	15	some	some	DET
fcis-25702	459	16	datasets	dataset	NOUN
fcis-25702	459	17	as	as	ADP
fcis-25702	459	18	examples	example	NOUN
fcis-25702	459	19	:	:	PUNCT
fcis-25702	459	20	10	10	NUM
fcis-25702	459	21	microsoft	microsoft	PROPN
fcis-25702	459	22	research	research	PROPN
fcis-25702	459	23	paraphrase	paraphrase	PROPN
fcis-25702	459	24	corpus	corpus	PROPN
fcis-25702	459	25	:	:	PUNCT
fcis-25702	459	26	given	give	VERB
fcis-25702	459	27	two	two	NUM
fcis-25702	459	28	sentences	sentence	NOUN
fcis-25702	459	29	,	,	PUNCT
fcis-25702	459	30	determine	determine	VERB
fcis-25702	459	31	if	if	SCONJ
fcis-25702	459	32	these	these	DET
fcis-25702	459	33	sentences	sentence	NOUN
fcis-25702	459	34	are	be	AUX
fcis-25702	459	35	semantically	semantically	ADV
fcis-25702	459	36	equivalent	equivalent	ADJ
fcis-25702	459	37	.	.	PUNCT
fcis-25702	460	1	output	output	NOUN
fcis-25702	460	2	‘	'	PUNCT
fcis-25702	460	3	equivalent	equivalent	ADJ
fcis-25702	460	4	’	'	PUNCT
fcis-25702	460	5	or	or	CCONJ
fcis-25702	460	6	‘	'	PUNCT
fcis-25702	460	7	not	not	PART
fcis-25702	460	8	equivalent	equivalent	ADJ
fcis-25702	460	9	’	'	PUNCT
fcis-25702	460	10	.	.	PUNCT
fcis-25702	461	1	sst-2	sst-2	PUNCT
fcis-25702	461	2	(	(	PUNCT
fcis-25702	461	3	stanford	stanford	PROPN
fcis-25702	461	4	sentiment	sentiment	PROPN
fcis-25702	461	5	treebank	treebank	VERB
fcis-25702	461	6	):	):	PUNCT
fcis-25702	461	7	given	give	VERB
fcis-25702	461	8	a	a	DET
fcis-25702	461	9	sentence	sentence	NOUN
fcis-25702	461	10	,	,	PUNCT
fcis-25702	461	11	evaluate	evaluate	VERB
fcis-25702	461	12	the	the	DET
fcis-25702	461	13	sentiment	sentiment	NOUN
fcis-25702	461	14	of	of	ADP
fcis-25702	461	15	this	this	DET
fcis-25702	461	16	sentence	sentence	NOUN
fcis-25702	461	17	as	as	ADP
fcis-25702	461	18	either	either	CCONJ
fcis-25702	461	19	positive	positive	ADJ
fcis-25702	461	20	or	or	CCONJ
fcis-25702	461	21	negative	negative	ADJ
fcis-25702	461	22	.	.	PUNCT
fcis-25702	462	1	output	output	NOUN
fcis-25702	462	2	‘	'	PUNCT
fcis-25702	462	3	positive	positive	ADJ
fcis-25702	462	4	’	'	PUNCT
fcis-25702	462	5	or	or	CCONJ
fcis-25702	462	6	‘	'	PUNCT
fcis-25702	462	7	negative	negative	ADJ
fcis-25702	462	8	’	'	PUNCT
fcis-25702	462	9	.	.	PUNCT
fcis-25702	463	1	qqp	qqp	PROPN
fcis-25702	463	2	(	(	PUNCT
fcis-25702	463	3	quora	quora	NOUN
fcis-25702	463	4	question	question	NOUN
fcis-25702	463	5	pairs	pair	NOUN
fcis-25702	463	6	):	):	PUNCT
fcis-25702	463	7	given	give	VERB
fcis-25702	463	8	two	two	NUM
fcis-25702	463	9	questions	question	NOUN
fcis-25702	463	10	,	,	PUNCT
fcis-25702	463	11	evaluate	evaluate	VERB
fcis-25702	463	12	whether	whether	SCONJ
fcis-25702	463	13	these	these	DET
fcis-25702	463	14	questions	question	NOUN
fcis-25702	463	15	are	be	AUX
fcis-25702	463	16	semantically	semantically	ADV
fcis-25702	463	17	similar	similar	ADJ
fcis-25702	463	18	.	.	PUNCT
fcis-25702	464	1	output	output	NOUN
fcis-25702	464	2	‘	'	PUNCT
fcis-25702	464	3	similar	similar	ADJ
fcis-25702	464	4	’	'	PUNCT
fcis-25702	464	5	or	or	CCONJ
fcis-25702	464	6	‘	'	PUNCT
fcis-25702	464	7	not	not	PART
fcis-25702	464	8	similar	similar	ADJ
fcis-25702	464	9	’	'	PUNCT
fcis-25702	464	10	.	.	PUNCT
fcis-25702	465	1	a.3	a.3	PROPN
fcis-25702	465	2	details	detail	NOUN
fcis-25702	465	3	of	of	ADP
fcis-25702	465	4	the	the	DET
fcis-25702	465	5	llms	llm	NOUN
fcis-25702	465	6	used	use	VERB
fcis-25702	465	7	for	for	ADP
fcis-25702	465	8	experiments	experiment	NOUN
fcis-25702	465	9	we	we	PRON
fcis-25702	465	10	show	show	VERB
fcis-25702	465	11	the	the	DET
fcis-25702	465	12	information	information	NOUN
fcis-25702	465	13	of	of	ADP
fcis-25702	465	14	the	the	DET
fcis-25702	465	15	large	large	ADJ
fcis-25702	465	16	language	language	NOUN
fcis-25702	465	17	models	model	NOUN
fcis-25702	465	18	we	we	PRON
fcis-25702	465	19	used	use	VERB
fcis-25702	465	20	in	in	ADP
fcis-25702	465	21	table	table	NOUN
fcis-25702	465	22	4	4	NUM
fcis-25702	465	23	.	.	PUNCT
fcis-25702	465	24	table	table	NOUN
fcis-25702	465	25	4	4	NUM
fcis-25702	465	26	.	.	PUNCT
fcis-25702	465	27	information	information	NOUN
fcis-25702	465	28	of	of	ADP
fcis-25702	465	29	llms	llm	NOUN
fcis-25702	465	30	.	.	PUNCT
fcis-25702	466	1	model	model	NOUN
fcis-25702	466	2	parameter	parameter	PROPN
fcis-25702	466	3	hidden	hide	VERB
fcis-25702	466	4	size	size	NOUN
fcis-25702	466	5	layer	layer	NOUN
fcis-25702	466	6	gpt2	gpt2	PROPN
fcis-25702	466	7	-	-	PUNCT
fcis-25702	466	8	xl	xl	PROPN
fcis-25702	467	1	[	[	X
fcis-25702	467	2	45	45	NUM
fcis-25702	467	3	]	]	SYM
fcis-25702	467	4	1.5b	1.5b	NUM
fcis-25702	467	5	1600	1600	NUM
fcis-25702	467	6	4	4	NUM
fcis-25702	467	7	opt-1.3b	opt-1.3b	NOUN
fcis-25702	467	8	[	[	X
fcis-25702	467	9	46	46	NUM
fcis-25702	467	10	]	]	SYM
fcis-25702	467	11	1.3b	1.3b	NUM
fcis-25702	467	12	2048	2048	NUM
fcis-25702	467	13	2	2	NUM
fcis-25702	467	14	falcon-7b	falcon-7b	NOUN
fcis-25702	467	15	[	[	X
fcis-25702	467	16	47	47	NUM
fcis-25702	467	17	]	]	SYM
fcis-25702	467	18	7b	7b	NOUN
fcis-25702	467	19	4544	4544	NUM
fcis-25702	467	20	3	3	NUM
fcis-25702	467	21	llama2	llama2	NOUN
fcis-25702	467	22	-	-	PUNCT
fcis-25702	467	23	7b	7b	PROPN
fcis-25702	468	1	[	[	X
fcis-25702	468	2	32	32	NUM
fcis-25702	468	3	]	]	PUNCT
fcis-25702	468	4	7b	7b	NOUN
fcis-25702	468	5	4096	4096	NUM
fcis-25702	468	6	3	3	NUM
fcis-25702	468	7	a.3.1	a.3.1	ADJ
fcis-25702	468	8	results	result	NOUN
fcis-25702	468	9	of	of	ADP
fcis-25702	468	10	other	other	ADJ
fcis-25702	468	11	llms	llm	NOUN
fcis-25702	468	12	using	use	VERB
fcis-25702	468	13	our	our	PRON
fcis-25702	468	14	method	method	NOUN
fcis-25702	468	15	to	to	PART
fcis-25702	468	16	validate	validate	VERB
fcis-25702	468	17	the	the	DET
fcis-25702	468	18	generalization	generalization	NOUN
fcis-25702	468	19	ability	ability	NOUN
fcis-25702	468	20	of	of	ADP
fcis-25702	468	21	our	our	PRON
fcis-25702	468	22	proposed	propose	VERB
fcis-25702	468	23	method	method	NOUN
fcis-25702	468	24	,	,	PUNCT
fcis-25702	468	25	we	we	PRON
fcis-25702	468	26	test	test	VERB
fcis-25702	468	27	it	it	PRON
fcis-25702	468	28	on	on	ADP
fcis-25702	468	29	other	other	ADJ
fcis-25702	468	30	llms	llm	NOUN
fcis-25702	468	31	on	on	ADP
fcis-25702	468	32	nlu	nlu	NOUN
fcis-25702	468	33	datasets	dataset	NOUN
fcis-25702	468	34	and	and	CCONJ
fcis-25702	468	35	show	show	VERB
fcis-25702	468	36	the	the	DET
fcis-25702	468	37	results	result	NOUN
fcis-25702	468	38	in	in	ADP
fcis-25702	468	39	table	table	NOUN
fcis-25702	468	40	5	5	NUM
fcis-25702	468	41	.	.	PUNCT
fcis-25702	468	42	table	table	NOUN
fcis-25702	468	43	5	5	NUM
fcis-25702	468	44	.	.	PUNCT
fcis-25702	469	1	f1	f1	NOUN
fcis-25702	469	2	-	-	PUNCT
fcis-25702	469	3	score	score	NOUN
fcis-25702	469	4	of	of	ADP
fcis-25702	469	5	method	method	NOUN
fcis-25702	469	6	with	with	ADP
fcis-25702	469	7	different	different	ADJ
fcis-25702	469	8	llms	llm	NOUN
fcis-25702	469	9	on	on	ADP
fcis-25702	469	10	nlu	nlu	NOUN
fcis-25702	469	11	tasks	task	NOUN
fcis-25702	469	12	.	.	PUNCT
fcis-25702	470	1	rte	rte	PROPN
fcis-25702	470	2	mrpc	mrpc	PROPN
fcis-25702	470	3	cola	cola	PROPN
fcis-25702	470	4	sst-2	sst-2	PUNCT
fcis-25702	470	5	gpt2	gpt2	PROPN
fcis-25702	470	6	-	-	PUNCT
fcis-25702	470	7	xl	xl	PROPN
fcis-25702	470	8	62.5	62.5	NUM
fcis-25702	470	9	68.5	68.5	NUM
fcis-25702	470	10	77.4	77.4	NUM
fcis-25702	470	11	82.5	82.5	NUM
fcis-25702	470	12	42.3	42.3	NUM
fcis-25702	470	13	46.1	46.1	NUM
fcis-25702	470	14	85.4	85.4	NUM
fcis-25702	470	15	88.2	88.2	NUM
fcis-25702	470	16	opt-1.3b	opt-1.3b	NOUN
fcis-25702	470	17	63.7	63.7	NUM
fcis-25702	470	18	70.5	70.5	NUM
fcis-25702	470	19	79.5	79.5	NUM
fcis-25702	470	20	85.3	85.3	NUM
fcis-25702	470	21	40.7	40.7	NUM
fcis-25702	470	22	45.2	45.2	NUM
fcis-25702	470	23	86.3	86.3	NUM
fcis-25702	470	24	88.4	88.4	NUM
fcis-25702	470	25	falcon-7b	falcon-7b	PROPN
fcis-25702	470	26	66.8	66.8	NUM
fcis-25702	470	27	75.9	75.9	NUM
fcis-25702	470	28	84.2	84.2	NUM
fcis-25702	470	29	89.1	89.1	NUM
fcis-25702	470	30	46.3	46.3	NUM
fcis-25702	470	31	55.1	55.1	NUM
fcis-25702	470	32	90.2	90.2	NUM
fcis-25702	470	33	92.4	92.4	NUM
fcis-25702	470	34	llama-7b	llama-7b	NOUN
fcis-25702	470	35	-	-	NOUN
fcis-25702	470	36	chat	chat	VERB
fcis-25702	470	37	68.8	68.8	NUM
fcis-25702	470	38	77.5	77.5	NUM
fcis-25702	470	39	85.6	85.6	NUM
fcis-25702	470	40	90.5	90.5	NUM
fcis-25702	470	41	49.4	49.4	NUM
fcis-25702	470	42	58.1	58.1	NUM
fcis-25702	470	43	90.6	90.6	NUM
fcis-25702	470	44	92.7	92.7	NUM