







































R. Dang & NA. Samad /Future Technology                                                                    November 2025| Volume 04 | Issue 04 | Pages 
159-172 

159 

 

 

 

Article 

Research on a virtual teacher personalized 

interaction model integrating affective computing 

and multi-agent systems 
Rili Dang1,2, Noorazman Abd Samad1* 

1Universiti Tun Hussein Onn Malaysia (UTHM), Panchor 84600, Johor, Malaysia 
2Zhuhai City Polytechnic, Zhuhai 519000, Guangdong, China 

A R T I C L E   I N F O 
 

Article history: 
Received 22 June 2025  
Received in revised form 
29 July 2025 
Accepted 18 August 2025 
 
Keywords:  
Virtual teacher, Affective computing,  
Multi-agent systems, Personalized learning, 
Intelligent teaching systems 
 
*Corresponding author 
Email address: 
noorazman@uthm.edu.my 
 
 
DOI: 10.55670/fpll.futech.4.4.14 

A B S T R A C T 
 

This research develops a novel virtual teacher personalized interaction model 
integrating multimodal affective computing with multi-agent coordination 
mechanisms to address fundamental limitations in emotional intelligence and 
adaptive capabilities within contemporary educational technology systems. A 
three-layer distributed architecture was implemented, incorporating 
synchronized multimodal emotion recognition through confidence-weighted 
fusion of facial, vocal, and textual data streams, Byzantine Fault Tolerant 
consensus algorithms for coordinated multi-agent decision-making, and 
dynamic personality adaptation mechanisms based on Big Five psychological 
modeling. Experimental validation employed 500 participants across diverse 
educational contexts using established emotion recognition benchmarks 
supplemented with domain-specific educational interaction datasets. The 
multimodal emotion fusion component achieved 91.2% recognition accuracy, 
with overall system performance reaching 89.7% under realistic educational 
conditions while demonstrating substantial educational effectiveness 
improvements, including 43% higher learner engagement scores, 37% 
emotional satisfaction enhancement, 30% learning effectiveness increase, and 
40% knowledge retention improvement compared to traditional virtual 
teaching approaches. Multi-agent coordination exhibited superior decision 
quality with 31% improvement over single-agent baselines, though personality 
adaptation effectiveness varied significantly across learner populations with 
88% success rates for extraverted individuals compared to 65% for high-
neuroticism learners. The integrated approach successfully bridges the 
emotional intelligence gap in virtual educational systems through sophisticated 
technological convergence, establishing theoretical foundations for distributed 
educational intelligence while revealing important implementation challenges. 
This research enables the development of emotionally responsive virtual 
teachers capable of sustained personalized instruction across diverse 
educational contexts, though deployment requires careful consideration of 
privacy protection and institutional adaptation requirements for broader 
educational technology transformation. 

1. Introduction 

Contemporary educational technologies exhibit 
substantial limitations in identifying and responding to 
learner emotional states, creating a critical gap for 
personalized learning interventions in post-pandemic 
adaptive educational systems [1]. Affective computing 
technologies offer opportunities to address this gap, though 
current solutions remain fragmented and lack comprehensive 
emotional intelligence integration [2]. Social-emotional 

learning technologies show promise but lack integrated 
intelligent operations for holistic responses to diverse learner 
demands [3]. Virtual human technologies demonstrate 
potential for human-like educational interactions while 
revealing adaptation challenges for avatar-based learning 
systems [4]. Intelligent educational systems have progressed 
through deep learning algorithms, multi-agent coordination, 
and advanced human-computer interaction paradigms. Deep 
learning systems have enhanced multimodal pattern 

 

 

Future Technology 

Open Access Journal 

https://doi.org/10.55670/fpll.futech.4.4.14 

 

 

November 2025| Volume 04 | Issue 04 | Pages 159-172 

Journal homepage: https://fupubco.com/futech 

 

ISSN 2832-0379 

mailto:noorazman@uthm.edu.my
https://doi.org/10.55670/fpll.futech.4.4.14
https://fupubco.com/futech


R. Dang & NA. Samad /Future Technology                                                                    November 2025| Volume 04 | Issue 04 | Pages 159-172 

160 

 

recognition for interpreting learner emotional states through 
facial, vocal, and textual analysis [5]. Neural network 
developments in human-computer interaction systems 
improve recognition accuracy and response appropriateness, 
particularly where emotional subtlety affects learning 
efficiency [6]. Multi-agent architectural models provide 
powerful paradigms for orchestrating complex educational 
interactions through specialized agents addressing different 
learning facets [7]. Reviews reveal gaps in achieving 
personalized emotional reactivity across educational 
contexts despite distributed intelligence architectures [8], 
with K-12 systems requiring advanced emotion integration 
for individual profiles [9]. The sustainability considerations of 
AI deployment in education emphasize the necessity for 
systems to adapt and evolve to meet different pedagogical 
demands, yet to maintain a stable ability to offer emotional 
support across diverse learning situations and cultural 
contexts [10]. Machine learning methods for predicting 
individual learning styles require integration of cognitive and 
affective models [11]. Empathic conversational agents 
demonstrate effectiveness but face limitations in 
personalized emotional responsiveness [12]. Current 
chatbots lack multi-dimensional personality modeling [13] 
and advanced emotional intelligence [14], while embodied 
agents rely on rigid personality models without 
individualized adaptation [15]. 

Recent studies reveal theoretical and practical 
limitations in integrating affective computing with multi-
agent coordination mechanisms. Immersive learning 
environments demonstrate the necessity of affective 
computing integration, though significant challenges exist in 
simultaneous intelligent function implementation [16]. 
Affective intelligent teaching systems show promise in 
detecting and responding to learner emotions, though 
pedagogical efficacy challenges persist across domain-
specific agents [17]. AI-based fast development frameworks 
for intelligent teaching systems mark some progress toward 
emotion-aware educational technology, yet even these 
systems are confronted with the problem of coordinating 
with multiple intelligent agents to appropriately respond to 
complex emotional states in diverse educational contexts 
[18]. Avatar-based systems demonstrate the importance of 
visual representation for emotional involvement [19] , while 
scaffolding agents show gains when enriched with emotional 
intelligence and adaptation mechanisms [20]. AI educational 
models require validated psychological constructs for 
interventions targeting both cognitive and affective learning 
dimensions [21]. Cognitive neuropsychology perspectives 
indicate that robust integration of cognitive models and 
affective computing enhances educational effectiveness and 
learner satisfaction [22]. 

Current virtual educational systems face critical 
limitations, including insufficient multimodal emotion 
recognition accuracy under real-world conditions, a lack of 
coherent multi-agent coordination mechanisms, and 
inadequate personality adaptation frameworks balancing 
consistency with flexibility. This investigation aims to 
develop comprehensive multimodal emotion recognition, 
design distributed multi-agent coordination mechanisms, 
implement adaptive personality modeling, and validate 
educational effectiveness across diverse learner populations. 
The key innovations include confidence-weighted 
multimodal fusion with real-time quality assessment, 
modified Byzantine Fault Tolerant consensus for educational 
contexts, and regularized personality adaptation balancing 
character consistency with behavioral flexibility. 

2. Methodology 

2.1 System architecture design 
The virtual teacher prototype proposed avoids problems 

of emotionally intelligent pedagogical systems with a three-
layer distributed structure allowing real-time multimodal 
emotion recognition, coordinated multi-agent decision-
making, and personality modeling adaptivity. Scalability is 
managed with layering problems of data acquisition, 
processing, and presentation in a manner where individual 
layer optimizations can be conducted with system 
consistency being preserved, along with complexity-
maintainability trade-offs eliminated for real-time reactivity 
in pedagogy. The architecture overcomes isolated emotion 
recognition limitations through integrated processing 
pipelines, maintaining temporal coherence across multiple 
data streams. The data acquisition layer incorporates 
synchronized RGB-D cameras (30fps), omnidirectional 
microphone arrays (48kHz), and natural language processing 
modules for real-time multimodal emotion analysis. The 
distributed processing employs an edge-cloud hybrid 
configuration with local processing handling time-sensitive 
emotion recognition, while cloud services manage 
personality adaptation algorithms and learning analytics. 
This approach addresses cloud-only latency issues and edge-
only computational constraints for complex personality 
modeling. Figure 1 illustrates the system architecture 
displaying linkages among data acquisition, processing 
engines, and decision coordination mechanisms. Figure 1 
depicts a hierarchical processing architecture demonstrating 
data flow among acquisition modules, processing engines, 
and decision coordination mechanisms. This three-tier 
structure preserves the real-time responsiveness of the 
system through parallel processing channels and guarantees 
data integrity through synchronized communication 
protocols. Due to its modularization, parts of the system can 
be optimized individually, and coherence between system 
components can be maintained by means of standardized 
interface protocols, supporting synchronous and 
asynchronous communication patterns, according to 
computational needs and time-dependent constraints. 

2.2 Multimodal affective computing model 
The multimodal emotion recognition framework 

addresses real-time emotional state interpretation in 
educational interactions through heterogeneous data stream 
processing. Conventional unimodal systems demonstrate 
limited reliability due to environmental degradation affecting 
pedagogical effectiveness, while multimodal fusion leverages 
complementary information to ensure recognition 
robustness across different scenarios. The system addresses 
emotion recognition ambiguity through sophisticated fusion 
strategies exploiting complementary facial expressions, vocal 
patterns, and linguistic content to achieve robust emotion 
estimation under adverse conditions, including partial 
occlusion, background noise, and communication problems. 
The confidence-weighted fusion automatically adapts 
modality contributions in real-time according to input signal 
quality, preventing unreliable modalities from corrupting 
final emotion estimation. The facial emotion recognition 
module implements a modified EfficientNet-B4 architecture 
enhanced with spatial attention mechanisms and temporal 
convolutional networks, processing 224×224 pixel facial 
regions through real-time detection and landmark 
localization.  

 
 



R. Dang & NA. Samad /Future Technology                                                                    November 2025| Volume 04 | Issue 04 | Pages 159-172 

161 

 

 
Figure 1. Hierarchical three-layer system architecture 

The vocal emotion analysis employs hybrid Wav2Vec 2.0 
feature extraction with bidirectional LSTM networks to 
process audio waveforms and extract emotional 
characteristics from prosodic features, spectral content, and 
temporal dynamics. The mathematical foundation for 
emotion fusion addresses the critical challenge of optimal 
information integration across modalities with varying 
reliability and temporal characteristics. The emotion state 
estimation at time t employs a confidence-weighted fusion 
mechanism where the final emotion vector et is computed 
through: 

𝑒𝑡 =
∑ 𝜔𝑖(𝑡)𝑖∈{𝑓,𝑣,𝑙} ⋅𝑐𝑖(𝑡)⋅𝑒𝑖(𝑡)

∑ 𝜔𝑖(𝑡)𝑖∈{𝑓,𝑣,𝑙} ⋅𝑐𝑖(𝑡)
               (1) 

Where 𝑒𝑖(𝑡)  represents the emotion vector from modality i 
(facial, vocal, linguistic), 𝑐𝑖(𝑡)  denotes the confidence score 
computed as: 

𝑐𝑖 = 1 −
𝐻(𝑒𝑖(𝑡))

𝑙𝑜𝑔𝐾
⋅

1

1+𝑒𝑥𝑝⁡(−𝛼⋅𝐶𝑀𝐶𝑖(𝑡))
               (2) 

where H(ei(t)) = −∑ pi,k(t)
K
k=1 log pi,k (t) represents 

prediction entropy, 𝐾 = 7  emotion classes, and 𝐶𝑀𝐶𝑖(𝑡) =
1

𝑁−1
∑ cos (𝑒𝑖(𝑡), 𝑒𝑗(𝑡))𝑗≠𝑖 measures cross-modal consistency 

with 𝛼 = ⁡2.5  empirically determined. The adaptive weight 
𝜔𝑖(𝑡) is calculated as: 

ω𝑖(𝑡) = β𝑖 ⋅ exp(−λ||𝑒𝑖(𝑡) − 𝑒(𝑡)||2
2)                (3) 

 

 

 

Where 𝑒(𝑡) =
1

3
∑ 𝑒𝑖(𝑡)𝑖  represents the mean emotion vector, 

𝛽𝑖  are modality-specific weights ( β𝑓 = 0.45, β𝑣 = 0.35, β𝑙 =

0.20 ), and λ = ⁡1.2  controls sensitivity to cross-modal 
deviation. Experimental validation shows confidence scores 
correlate strongly with recognition accuracy (r = 0.847, p < 
0.001) while adaptive weights improve fusion robustness by 
12.3% under noisy conditions. 

The technical implementation specifications of the three 
core processing modalities and their integration mechanism 
require systematic documentation to demonstrate the 
architectural coherence and processing capabilities of the 
proposed multimodal emotion recognition framework, as 
detailed in Table 1. 

Table 1 indicates that the proposed multimodal 
architecture integrates three specialized processing 
components through a confidence-weighted fusion 
mechanism. The facial expression module processes visual 
input through an enhanced EfficientNet-B4 architecture with 
spatial attention mechanisms, while vocal analysis combines 
Wav2Vec 2.0 feature extraction with bidirectional LSTM 
processing. The textual analysis employs transformer-based 
natural language processing, and a fusion mechanism 
dynamically integrates heterogeneous emotion vectors using 
confidence estimation based on prediction entropy and cross-
modal consistency measures. 

 
 
 

Presentation Layer

User Interface

Management

Module

Interaction

Protocols

Controller

Virtual Teacher

Visual

Representation

Standardized

Interface

Protocols

Intermediate Processing Layer

Edge Processing (Local)

Real-time 

Emotion

Recognition Engin

Multimodal 

Fusion

Processor

Cloud Processing (Remote)

Decision 

Coordination

Hub

Cross-session 

Learning

Analytics

Pedagogical

Agent

Affective

Agent

Personalty

Agent

Dialogue

Agent

Data Acquisition Layer

RGB-D Cameras

(30fps)

Facial Analysis

Omnidirectional

Microphones 

(48kHz)

Vocal Analysis

Natural Language

Processing

Modules

Temporal Sync

Controller

(Synchronized 

Arrays)

Synchronized Communication Protocols

Parallel Processing Pathways with Temporal Coherence

S
ta

n
d
ar

d
iz

ed
 I

n
te

rf
ac

e
s

D
e
c
is

io
n
 C

o
o
rd

in
a
ti

o
n

Multi-Agent

Coordination

Mechanisms



R. Dang & NA. Samad /Future Technology                                                                    November 2025| Volume 04 | Issue 04 | Pages 159-172 

162 

 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

2.3 Multi-agent coordination mechanism 
The proposed multi-agent coordination framework 

focuses on the basic problem of maintaining consistent 
educational interactions, whilst affording specialisation-
oriented agent autonomy, through the use of an innovative 
consensus decision architecture that reconciles domain-
specific expertise of single agents with system-level 
pedagogical goals. The distributed architecture described 
above tackles the expertise dilution in monolithic systems, in 
which a single decision-maker attempts to deal single-
handedly with a wide variety of educational issues, resulting 
in sub-optimal coverage of several domains. The coordination 
system goes beyond conventional hierarchical systems and 
the supremacy of a central authority to facilitate distributed 
consensus algorithms to optimize collaboration by giving 
power to agents to negotiate solutions while preserving the 
organic properties of genuine educational participation. This 
is done in a way that bypasses the limitations of centralized 
systems, where single points of control get overwhelmed in 
dealing with nested educational environments with multiple 
concurrent goals.  The agent's architecture is divided into four 
modules: the Pedagogical Agent, with the responsibility for 
course management and learning goals optimization; the 
Affective Agent, for emotional states tracking and triggering 
appropriate interventions; the Personality Agent, for student 
dynamic modeling and adaptation of interaction patterns; the 
Dialogue Agent, for generation of natural speech and control 
of conversational flow. These modules have a specific 
knowledge base and carry out coordinated decision-making 
activities informed by defined protocols for negotiation. 

The coordination mechanism employs a modified 
Byzantine Fault Tolerant consensus protocol designed for 
educational decision-making scenarios where agents' 
decisions should accommodate diverse goals like learning 
effectiveness, emotional appropriateness, and personality 
consistency. Its conflict resolution for agents' conflicting 
recommendations utilizes utility-based voting, where the 
extent of each agent's contribution towards the ultimate 
decisions hinges on both their knowledge about the domain, 
along with context appropriateness. 

The mathematical formulation for distributed decision 
consensus addresses the challenge of optimal action selection 
when agents have potentially conflicting recommendations. 
The system utility maximization employs a multi-objective 
optimization approach where the global action a∗  is 
determined through: 

𝑎∗ = argmax
𝑎∈𝐴

[∑ α𝑗(𝑠𝑡)
4
𝑗=1 ⋅ 𝑈𝑗(𝑎, 𝑠𝑡) − λ ⋅ ϕ(𝑎, ℎ𝑡)]           (4) 

 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Where 𝑈𝑗(𝑎, 𝑠𝑡) represents the utility function for agent j 

given action 𝑎  and the current state 𝑠𝑡 , 𝛼𝑗(𝑠𝑡)  denotes the 

context-dependent weighting for agent j, 𝜙(𝑎, ℎ𝑡) represents 
the coordination cost function based on interaction history 
ℎ𝑡 , and λ balances individual utility against coordination 
overhead. 

The complex interaction patterns and decision flow 
within the multi-agent coordination system require detailed 
visualization to understand agent communication protocols 
and consensus formation processes during typical 
educational interaction scenarios, as illustrated in Figure 2. 
Figure 2 shows how agents exchange information about 
learner state, propose intervention strategies, and negotiate 
final decisions through structured message passing protocols 
that ensure both efficiency and transparency in the decision-
making process. The coordination mechanism maintains 
decision traceability to support system explainability and 
continuous improvement through interaction outcome 
analysis. 

2.4 Personalized interaction strategy 
The personalized interaction strategy framework 

addresses the challenge of creating adaptive virtual teacher 
personalities that dynamically adjust interaction styles based 
on comprehensive learner profiling and real-time contextual 
assessment. The dynamic personality adaptation design 
addresses engagement plateau problems in static virtual 
teacher systems where learners lose interest due to 
predictable patterns, while solving personality inconsistency 
issues arising from arbitrary behavioral changes without 
character coherence.  

The framework addresses personality consistency 
versus adaptivity through regulated adaptation processes, 
maintaining basic personality components while allowing 
fine-grained behavioral variations according to learner 
preferences and interaction efficacy. This regularized 
adaptation approach reconciles the trade-off between 
responsiveness to learner feedback and adherence to credible 
character consistency in adaptive virtual teacher systems. 
The personality modeling employs hierarchical Bayesian 
approaches to update dynamic personality profiles with 
observed behavior, incorporating explicit feedback and 
learning outcome correlations. The model mitigates small 
interaction data limitations through transfer learning 
techniques, exploiting population-level personality 
tendencies while enabling custom-fit adaptation to individual 
participants.  

 
 

Table 1. Technical specifications of multimodal emotion recognition architecture 

Component Architecture Input Specifications Processing Method Output Format 

Facial Expression EfficientNet-B4 + 
Attention 

224×224 pixels, 30fps Spatial attention + 
TCN 

512-dim emotion 
vector 

Vocal Analysis Wav2Vec 2.0 + BiLSTM 48kHz audio sampling Prosodic + spectral 
features 

Temporal emotion 
sequence 

Textual Analysis Transformer-based NLP Real-time text input Contextual sentiment 
processing 

Emotion probability 
vector 

Multimodal Fusion Confidence-weighted All modality vectors Adaptive weight 
integration 

Unified emotion state 

 



R. Dang & NA. Samad /Future Technology                                                                    November 2025| Volume 04 | Issue 04 | Pages 159-172 

163 

 

 
 
Figure 2. Multi-agent coordination and decision flow architecture 

The personality adaptation mechanism implements 
constrained optimization where virtual teacher personality 
parameters evolve within predefined bounds to maintain 
believable character consistency while optimizing interaction 
effectiveness. The adaptation process addresses multi-
dimensional personality optimization through Markov 
Decision Process formulation, treating personality 
adjustments as sequential decision problems. 

The personality state evolution employs a regularized 
adaptation mechanism that balances responsiveness to 
learner feedback with personality stability requirements. The 
personality parameter update follows: 

𝑝𝑡+1 = 𝑝𝑡 + η∇𝑝𝐽(𝑝𝑡, 𝑟𝑡) ⋅ exp(−γ||𝑝𝑡 − 𝑝0||2
2)           (5) 

where 𝑝
𝑡

 represents the personality parameter vector, 

𝐽(𝑝𝑡, 𝑟𝑡) denotes the interaction effectiveness function based 
on personality configuration using Five-Factor Model (OCEAN) 
dimensions, η controls adaptation rate, and γ regulates 
personality consistency constraints, with computational 
assessment revealing 88% adaptation success for extraverted 
learners versus 65% for high-neuroticism individuals. 
The comprehensive personality modeling and adaptation 
capabilities require detailed specification of personality 
dimensions, adaptation ranges, and behavioral manifestation 
patterns to demonstrate the framework's sophisticated 
interaction customization capabilities, as presented in Table 
2. Table 2 shows how the framework maintains personality 
coherence across multiple interaction dimensions while 
enabling sufficient flexibility to accommodate diverse learner 
preferences and educational contexts.  

 
 
 
 
The experimental validation methodology focuses on 
demonstrating measurable improvements in learner 
engagement, emotional satisfaction, learning effectiveness, 
and retention rates through controlled comparative studies 
involving diverse educational scenarios and learner 
populations. 

3. Results 

3.1 Experimental design and dataset construction 
The experimental validation demonstrates affective 

computing and multi-agent coordination integration 
effectiveness in educational environments, evaluating the 
system's capability to recognize learner emotional states, 
coordinate intelligent agents, and adapt personality 
characteristics for optimized learning outcomes. The 
systematic experimental design utilizes established emotion 
recognition benchmarks for reliable comparison with existing 
methodologies while incorporating controlled educational 
scenarios for domain-specific validation. The study 
encompasses 500 participants aged 12-65 years (350 cross-
sectional, 150 longitudinal), randomly assigned through 
stratified sampling based on age, educational level, and 
cultural background. Demographic subgroup analysis reveals 
emotion recognition accuracy variance within 2.1% across 
age cohorts (12-25: 90.8%, 26-45: 91.2%, 46-65: 89.7%) with 
no statistically significant differences (F(2,497) = 1.34, p = 
0.264). Cultural background assessment across Western 
(n=187), East Asian (n=156), and other populations (n=157) 
shows consistent system performance, though personality 
adaptation effectiveness differs significantly across cultural 
contexts (χ² = 12.47, p < 0.01).  

Current Learner State

Emotion   Personality   Learning Context   Educational Needs

Pedagogical Agent

Analyzes Leaming Context 

Educational Objectives

Affective Agent

Processes Emotional Data& 

Intervention Strategies

Personality Agent

Evaluates individual Traits& 

Adaptation Needs.

Dialogue Agent

Assesses Communication

& Response Generation

Curriculum Adjustment

Proposal

Learning Path Optimization

Emotional Support

Proposal

Intervention Strategy

Personality Adaptation

Proposal

Interaction Style Tuning

Dialogue Strategy

Proposal

Response Planning

Message Exchange

Protocol

Structured Communication

Utility-Based

Voting

Multi-Objective Assessment

Byzantine Fault Tolerant Consensus Algorithm

Educational Action

Implemented Response

Emotional Intervention

Support Delivery

Personality Response

Adaptive Behavior

Dialogue Output

Generated Communication

Step 1: Learner State Information

Step 2:Agent Information Exchange

Step 3: Intervention Strategy Proposals

Step 4: Structured Message Passing& Negotiation

Step 5: Consensus Formation

Step 6: Coordinated Fihal Decision

Legend

              Process Flow

              Information Exchange

              Feedback Loop



R. Dang & NA. Samad /Future Technology                                                                    November 2025| Volume 04 | Issue 04 | Pages 159-172 

164 

 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
The experimental framework integrates established 

emotion recognition benchmarks, including FER2013, 
RAVDESS, and AffectNet, comprising over 500,000 annotated 
samples with controlled educational validation. The 
validation protocol employs randomized controlled trials 
with 5-fold stratified sampling and 6-month longitudinal 
studies tracking personality adaptation effectiveness. Table 3 
presents the systematic integration of public datasets with 
educational-specific data collection for the establishment of a 
comprehensive benchmarking framework. Table 3 
demonstrates integration of established emotion recognition 
benchmarks with custom educational datasets, enabling 
performance comparison against state-of-the-art systems. 
FER2013 and AffectNet provide over 485,000 annotated 
facial images, while RAVDESS and IEMOCAP offer multimodal 
validation with high inter-annotator agreement. The 
Educational-EAC dataset introduces learning-specific 
emotional states, including engagement, frustration, and 
confusion, crucial for educational applications. Multimodal 
fusion achieves 91.2% accuracy under controlled conditions, 
while system-level performance averages 89.7% when 
integrated with real-time coordination and personality 
adaptation mechanisms. 

3.2 Emotion recognition performance evaluation 
The validation employs systematic comparison 

methodologies evaluating proposed fusion mechanisms 
against established single-modality and multimodal 
approaches using standardized protocols. Experimental 
design involves setting up controlled test conditions wherein 
subjects take part in pedagogic interactions. At the same time, 
multimodal systems record facial expressions, audio cues, 
and text messages. It tests recognition accuracy for separate 
modalities using confidence-weighted fusion, along with the 
latency for real-time verifiability of performance. 

 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
The architectural assessment includes deep learning 

methods tailored for learning emotion recognition optimized 
for education, using transfer learning from pre-trained 
models with fine-tuning on educational datasets. 
Performance assessment under different environmental 
conditions, such as lighting conditions, background noise, and 
multiple speakers, confirms robustness under real classroom 
conditions, showing drastic performance degradation under 
harsh constraints. Table 4 presents detailed performance 
comparison results across evaluation metrics and operational 
constraints. Table 4 shows that EfficientNet-B4 achieves 
89.3% facial recognition accuracy (145ms latency), and 
Wav2Vec 2.0 demonstrates 82.4% vocal accuracy. 
Multimodal fusion achieves 91.2% accuracy (185ms latency) 
under controlled conditions, while system-level performance 
averages 89.7% when integrated with coordination and 
adaptation mechanisms, reflecting computational overhead 
from multi-agent architecture. Comparative evaluation 
against established multimodal emotion recognition 
architectures validates system superiority. Table 5 presents 
benchmark performance analysis under identical 
experimental conditions. Table 5 demonstrates that the 
confidence-weighted fusion mechanism achieves superior 
performance while maintaining computational efficiency 
compared to existing state-of-the-art approaches. The 
systematic evaluation reveals significant challenges in 
maintaining consistent performance across diverse 
educational environments, with accuracy dropping 12-15% 
in real classroom settings compared to controlled laboratory 
conditions. Performance optimization addresses 
computational constraints through edge-cloud hybrid 
processing, though network latency variations introduce 
complexity with response times ranging 150ms-300ms 
depending on connection quality. Figure 3 demonstrates the 
relationship between processing latency and recognition 
accuracy across various implementation approaches. 

Table 2. Personality dimension specifications and adaptation framework 

Big Five Dimension Adaptation 
Range 

Behavioral Manifestations Contextual Factors Assessment Metrics 

Extraversion 0.2 - 0.8 High: Frequent encouragement, 
group activities, enthusiastic 
tone 
Low: Calm guidance, individual 
focus, reflective questioning 

Learner social comfort, 
class size, interaction 
history 

Engagement level, 
interaction frequency 

Agreeableness 0.3 - 0.9 High: Supportive feedback, 
collaborative approach, gentle 
correction 
Low: Direct criticism, 
competitive elements, 
challenging questions 

Learner confidence, skill 
mastery, learning 
objectives 

Emotional 
satisfaction, stress 
levels 

Conscientiousness 0.4 - 0.9 High: Structured approach, 
detailed planning, systematic 
feedback 
Low: Flexible pacing, adaptive 
scheduling, creative freedom 

Learning timeline, 
assessment deadlines, 
task complexity 

Learning 
effectiveness, goal 
completion 

Openness 0.3 - 0.8 High: Creative exercises, novel 
approaches, experimental 
methods 
Low: Traditional methods, 
proven techniques, structured 
content 

Subject matter, learner 
background, innovation 
comfort 

Knowledge retention, 
creative output 

Neuroticism 0.1 - 0.6 Low: Calm demeanor, stress 
reduction, emotional stability 
High: Cautious approach, 
detailed explanations, anxiety 
awareness 

Learner emotional state, 
exam pressure, difficulty 
level 

Emotional well-being, 
anxiety reduction 

 



R. Dang & NA. Samad /Future Technology                                                                    November 2025| Volume 04 | Issue 04 | Pages 159-172 

165 

 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Table 5. Benchmark performance comparison 

Architecture Accuracy 
(%) 

F1-
Score 

Latency 
(ms) 

Dataset 

Proposed Fusion 91.2 0.897 185 Integrated 
EmotiNet 86.8 0.831 245 FER2013 
AffectNet Fusion 84.7 0.819 267 AffectNet 
Transformer-
based 

87.9 0.854 298 RAVDESS 

 
 

 
Figure 3. Multimodal emotion recognition accuracy vs. real-time 
processing trade-off 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Figure 3 shows the performance optimization space with 
confidence-weighted fusion achieving 91.2% accuracy at 
185ms latency through adaptive edge-cloud processing. Pure 
cloud processing achieves higher accuracy (93.8%) but 
problematic latency (350-450ms), while pure edge 
processing maintains acceptable latency (130ms) but lower 
accuracy (86.7%). The hybrid approach dynamically routes 
operations, though network interruptions cause a 15-20% 
accuracy reduction during connectivity issues. System latency 
comprises sequential processing stages, including data 
acquisition, multimodal emotion recognition, multi-agent 
consensus formation, and response generation, with 
performance degradation observed under concurrent multi-
user scenarios where communication overhead and 
computational resource contention introduce additional 
delays beyond single-user baseline measurements.  

The validation employs systematic testing protocols 
evaluating consensus formation efficiency, decision quality, 
and system scalability. The experiment involves controlled 
scenarios where multiple agents agree on educational 
decisions while balancing conflicting goals. The benchmark 
assesses convergence time, communication costs, and 
decision quality against expert standards. 

The coordination efficiency study examines agent 
performance across varying complexities, from simple 
content selection to complex integrated systems. The 
experimental protocol applies stress testing, including high-
frequency decision-making, partial communication loss, and 
agent behavior alterations. Table 6 presents efficiency 
metrics across operational scenarios. Table 6 shows that the 
Byzantine Fault Tolerant consensus algorithm achieves 
coordination with convergence times ranging from 1.8 to 5.7 
seconds, maintaining decision quality scores of 0.692 to 
0.847.  

Table 3. Public dataset integration and benchmarking framework 

Dataset Modality Size Emotion Categories Usage Purpose Performance 
Baseline 

FER2013 Facial 35,887 images 7 basic emotions Facial expression 
training 

71.2% accuracy 

RAVDESS Audio-Visual 7,356 clips 8 emotions + neutral Speech emotion 
validation 

78.4% accuracy 

AffectNet Facial 450,000 
images 

8 expressions + 
valence/arousal 

Large-scale facial 
training 

65.2% accuracy 

IEMOCAP Multimodal 12 hours 4 emotions + dimensions Multimodal fusion 
testing 

73.8% accuracy 

EmoDB Audio 535 
utterances 

7 emotions German speech 
validation 

84.3% accuracy 

Educational-EAC Custom 
Multimodal 

15,000 
sessions 

11 learning states Domain-specific 
training 

New benchmark 

 

Table 4. Deep learning model performance comparison for emotion recognition 

Architecture Modality Accuracy 
(%) 

F1-
Score 

Processing Latency 
(ms) 

Memory Usage 
(MB) 

Robustness 
Score 

EfficientNet-B4 Facial 89.3 0.876 145 78 0.812 

ResNet-50 Facial 85.7 0.843 180 102 0.787 

Wav2Vec 2.0 Audio 82.4 0.798 160 124 0.723 

BERT-base Text 78.9 0.761 95 89 0.695 

Multimodal Fusion Combined 91.2 0.897 185 156 0.834 

Baseline CNN Facial 76.8 0.734 230 145 0.642 

 



R. Dang & NA. Samad /Future Technology                                                                    November 2025| Volume 04 | Issue 04 | Pages 159-172 

166 

 

 

 

 

 

 

 

 

 

 

 

 

 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Table 6. Multi-agent coordination efficiency and consistency metrics 

Scenario Type Agents 
Involved 

Convergence Time 
(s) 

Decision Quality 
Score 

Communication 
Overhead (%) 

Consistency Rate 
(%) 

Simple Content 
Selection 

2-3 agents 1.8 ± 0.7 0.847 12.4 91.3 

Complex Multi-
objective 

4 agents 3.2 ± 1.1 0.763 24.6 84.7 

High-frequency 
Decisions 

4 agents 2.9 ± 1.3 0.721 31.8 79.2 

Partial 
Communication 
Loss 

3-4 agents 5.7 ± 2.1 0.692 18.9 73.4 

Single-Agent 
Baseline 

1 agent 0.6 ± 0.2 0.698 0 87.1 

 

 

 

Figure 4. Multi-agent system convergence analysis and performance comparison 



R. Dang & NA. Samad /Future Technology                                                                    November 2025| Volume 04 | Issue 04 | Pages 159-172 

167 

 

The distributed approach demonstrates superior 
decision quality (0.763) compared to the single-agent 
baseline (0.698), though communication overhead reaches 
31.8%. Consistency rates degrade from 91.3% to 73.4% 
under communication failures, highlighting network 
vulnerability. The coordination demonstrates superior 
decision quality (0.763) compared to the single-agent 
baseline (0.698), with convergence times ranging from 1.8 to 
5.7 seconds and consistency rates maintaining 91.3% under 
normal conditions, degrading to 73.4% during 
communication failures. The convergence analysis reveals 
important limitations of the coordination mechanism under 
challenging operational conditions, particularly showing 
increased variability in response times and occasional failure 
to reach consensus within acceptable time limits for real-time 
educational interactions. To visualize the coordination 
dynamics and performance characteristics, including failure 
modes, Figure 4 illustrates the convergence patterns and 
comparative performance analysis. Figure 4 shows a 
comprehensive multi-agent coordination performance 
analysis across six key dimensions. Convergence analysis 
Figure 4(a) demonstrates consensus achievement within 2-4 
iterations with quality scores 0.692-0.847. Performance 
comparison Figure 4(b) reveals 31% superior decision 
quality over single-agent baselines. Response time evaluation 
Figure 4(c) indicates a 28% reduction through parallel 
processing. Load condition testing, Figure 4(d), exhibits 
graceful degradation, maintaining 0.63 performance at peak 
loads versus single-agent collapse at 0.25. Communication 
overhead analysis. Figure 4(e) shows acceptable coordination 
costs (8.2-35.8%). Recovery performance Figure 4(f) 
demonstrates 1.5-3.2 second fault tolerance, outperforming 
single-agent systems requiring 6.8-15.2 seconds. 

3.3 Personalized interaction effectiveness assessment 
The evaluation employs longitudinal experimental 

designs assessing personality adaptation effectiveness across 
diverse learner populations, revealing significant benefits and 
notable limitations. The protocol implements randomized 
controlled trials where participants interact with adaptive or 
static systems while measuring engagement levels, learning 
satisfaction, and educational outcomes. The framework 
incorporates Big Five personality profiling and learning style 
assessment for baseline characteristics, guiding adaptation 
algorithms. Personality adaptation evaluation implements 
systematic parameter optimization using Bayesian 
techniques and multi-armed bandit approaches, though 
convergence requires 8-12 interaction sessions.  

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Table 7 presents detailed outcome measurements across 
evaluation criteria. Table 7 shows meaningful performance 
improvements with the adaptive personality system, 
achieving 14.7% enhancement in learning effectiveness 
scores and 19.6% improvement in user satisfaction ratings 
compared to static configurations. The system demonstrates 
96.8% uptime during a 6-month deployment. Statistical 
analysis reveals significant improvements across most 
measures, though effect sizes remain moderate with high 
variance, suggesting system effectiveness depends heavily on 
implementation environment and user characteristics. 
Component contribution analysis through a systematic 
ablation study quantifies individual module impacts on 
overall system performance. Table 8 details the experimental 
results. Table 8 reveals personality adaptation as the most 
critical component for engagement enhancement, while 
multimodal fusion provides substantial accuracy 
improvements. The differential effectiveness analysis across 
personality types reveals significant variation in adaptation 
benefits, with some personality combinations showing 
minimal improvement while others demonstrate substantial 
gains, indicating the need for more sophisticated adaptation 
strategies. Figure 5 presents a comprehensive analysis of 
personality-specific adaptation effectiveness and limitations 
across different learner types. Figure 5 shows a 
comprehensive personality-based interaction effectiveness 
analysis. Extraversion analysis Figure 5(a) demonstrates 
34% higher engagement for extraverted learners. 
Introversion assessment Figure 5(b) reveals a 12% learning 
satisfaction improvement. Neuroticism evaluation Figure 5(c) 
indicates a 15% anxiety reduction. Personality combination 
analysis. Figure 5(d) shows minimal benefits for high 
conscientiousness with low openness (3-5% improvement), 
highlighting algorithm limitations.  

Table 8. Component ablation analysis 

Removed 
Component 

Accuracy 
Drop (%) 

Engagement 
Impact (%) 

Learning 
Effectiveness 
Impact (%) 

Multimodal Fusion -8.4 -18.2 -12.1 

Multi-Agent 
Coordination 

-6.7 -15.8 -9.3 

Personality 
Adaptation 

-12.3 -26.4 -8.7 

Spatial Attention -4.1 -7.9 -4.2 

Full System 91.2 43.0 14.7 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Table 7. Comprehensive system performance and user acceptance results 

Performance Metric Adaptive 
System 

Static Control Improvement (%) Statistical Significance 

Emotion Recognition 
Accuracy 

91.2% 82.6% +10.4% p < 0.001 

System Uptime (6 
months) 

96.8% 94.2% +2.8% p < 0.05 

Average Response 
Time 

203ms 278ms -27.0% p < 0.001 

Learning 
Effectiveness Score 

0.724 0.631 +14.7% p < 0.01 

User Satisfaction 
Rating 

3.84/5.0 3.21/5.0 +19.6% p < 0.01 

Knowledge Retention 
(30 days) 

68.7% 61.4% +11.9% p < 0.05 

Educator Acceptance 
Rate 

74% 58% +27.6% p < 0.05 

 



R. Dang & NA. Samad /Future Technology                                                                    November 2025| Volume 04 | Issue 04 | Pages 159-172 

168 

 

Big Five effectiveness comparison Figure 5(e) 
demonstrates variable adaptation success. Success rate 
analysis Figure 5(f) reveals differential outcomes, with 
extraverted learners achieving 88% success rates compared 
to 65% for high-neuroticism learners. 

3.4 Comprehensive system performance testing 
The comprehensive evaluation implements large-scale 

deployment testing to validate system performance under 
realistic operational conditions while measuring educational 
effectiveness through controlled longitudinal studies, 
revealing both promising results and significant 
implementation challenges. The testing framework 
encompasses systematic assessment of technical reliability, 
educational outcome improvements, and user acceptance 
across diverse educational contexts, including individual 
tutoring, small group instruction, and classroom integration 
scenarios. The evaluation protocol implements pre-post 
assessment designs with 6-month follow-up periods to  

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

measure sustained educational improvements, though 
participant attrition of 23% complicated longitudinal analysis 
and required imputation methods for missing data. Scalability 
assessment validates system performance under realistic 
deployment conditions across multiple educational 
environments. Table 9 presents empirical analysis results. 
Table 9 demonstrates graceful performance degradation 
under increased load while maintaining educational 
effectiveness above 87% across all deployment scenarios, 
validating practical scalability for institutional adoption. The 
large-scale testing methodology incorporates deployment 
across 12 educational institutions with systematic 
measurement of system stability, performance consistency, 
and educational outcome improvements compared to 
traditional virtual teaching approaches, encountering 
substantial implementation challenges, including hardware 
compatibility issues, network infrastructure limitations, and 
varying institutional support levels. 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 
 

Figure 5. Personality-based interaction effectiveness across different learner types 

 



R. Dang & NA. Samad /Future Technology                                                                    November 2025| Volume 04 | Issue 04 | Pages 159-172 

169 

 

Table 9. Scalability performance analysis 

Deployment 
Scenario 

Concurrent 
Users 

System 
Accuracy 

(%) 

Response 
Time 
(ms) 

Bandwidth 
Usage 

(Mbps) 

Single 
Classroom 

8-12 91.2 203 2.1 

Multiple 
Classrooms 

23-35 89.1 278 7.8 

Institution-
wide 

45-67 86.8 356 14.3 

Cross-
platform 
Mixed 

28-41 88.3 312 9.7 

 
 The evaluation framework implements comprehensive 

statistical analysis, including effect size calculations and 
power analysis, though several planned comparisons proved 
underpowered due to smaller-than-anticipated effect sizes 
and higher-than-expected variance in educational outcomes. 
To provide a realistic comparison of educational effectiveness 
improvements achieved by the proposed system, Figure 6 

  

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

presents a detailed analysis of teaching effectiveness 
enhancements across key educational metrics. Figure 6 
demonstrates substantial teaching effectiveness 
improvements across six dimensions. Main comparison 
Figure 6(a) shows 43% engagement enhancement, 37% 
emotional satisfaction improvement, 30% learning 
effectiveness increase, and 40% knowledge retention over 
traditional systems. Cross-age analysis Figure 6(b) reveals 
consistent gains across 12-65 years. Cultural assessment 
Figure 6(c) indicates 39-54% improvements across 
populations. Learning style evaluation. Figure 6(d) shows 
consistent VARK gains. Component analysis Figure 6(e) 
confirms personality adaptation as the primary driver (28%), 
with coordination (22%) and emotion recognition (18%) 
contributions. Temporal study Figure 6(f) demonstrates 
sustained six-month performance. 

4. Discussion 

The experimental validation demonstrates substantial 
advancement in educational emotion recognition capabilities, 
achieving 91.2% multimodal accuracy that significantly 
exceeds conventional approaches.  

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

 

 

 

Figure 6. Teaching effectiveness improvement: proposed system vs. traditional methods 



R. Dang & NA. Samad /Future Technology                                                                    November 2025| Volume 04 | Issue 04 | Pages 159-172 

170 

 

Contemporary research emphasizing AI-driven emotion 
detection for adaptive teaching optimization [23] provides 
theoretical validation for observed 43% engagement 
enhancement and 37% emotional satisfaction improvement, 
though findings expose critical gaps between laboratory and 
classroom implementation. The multi-agent coordination 
mechanism achieves superior pedagogical decision-making 
through distributed consensus formation, demonstrating 
31% higher decision quality compared to centralized 
approaches. Recent advances in LLM-powered multi-agent 
frameworks for goal-oriented learning [24] support 
theoretical foundations for distributed educational 
intelligence, yet performance degradation under 
communication failures highlights network vulnerability. The 
personality adaptation framework exhibits variable 
effectiveness, with extraverted individuals achieving 88% 
adaptation success rates compared to 65% for high-
neuroticism learners. Cognitive assessment studies utilizing 
multi-agent deep learning architectures [25] demonstrate 
potential for distributed intelligence approaches. 

The proposed multimodal emotion fusion algorithm 
realizes substantial theoretical progress with confidence-
weighted integration mechanisms dynamically adjusting 
modality contributions based on real-time quality 
assessment, overcoming current systems' single-point-of-
failure bottlenecks. The Byzantine Fault Tolerant consensus 
algorithm tailored for education represents a theoretical 
contribution to distributed decision-making mechanisms, 
enabling coherent sub-agent coordination despite operating 
constraint changes. Studies addressing the integration of 
large language models in educational agent design [26] point 
out the great transformative potential from powerful 
language abilities, but existing realizations suggest advanced 
natural language processing technologies cannot meet the 
subtle psychological adaptation needs prerequisite for 
efficient personalized education. Studies regarding virtual 
simulations focusing on avatars for educating relational 
competencies [27] show possibilities but restrictions 
inherent in real educational relations, thereby substantiating 
findings addressing visual representation, with behavioral 
consistency being essential, but at the same time shedding 
light on difficulties in sustaining personality consistency in 
the face of adaptive interactions. 

The comprehensive system evaluation establishes 
practical viability for sustained educational deployment 
through demonstrated 40% knowledge retention 
enhancement over six-month periods, though 
implementation barriers, including hardware compatibility 
issues, constrain broader adoption potential. Social presence 
research in virtual reality environments [28] demonstrates 
significant influence on learning engagement, providing 
empirical support for observed personality adaptation effects 
while emphasizing the critical importance of maintaining 
believable character consistency. The sustainability 
perspective on AI-driven educational transformation [29] 
emphasizes long-term adaptation capabilities, suggesting 
that demonstrated system resilience aligns with educational 
technology evolution trends toward adaptive learning 
solutions. Virtual environment psychological mechanism 
studies [30] reveal significant influence on learner mental 
states, providing theoretical support for observed 
personality-dependent effectiveness variations while 
highlighting the complex interplay between technological 
capabilities and psychological factors. The deployment of 
sophisticated emotion monitoring systems raises critical 
ethical considerations regarding learner privacy, data 

security, and psychological manipulation concerns, requiring 
robust data protection frameworks and careful examination 
of personalization benefits versus potential risks to learner 
autonomy. The system addresses these concerns through 
GDPR/FERPA-compliant protocols, including AES-256 real-
time encryption, federated learning architecture preventing 
raw data transmission, and k-anonymity preservation (k≥5). 
Informed consent procedures ensure transparency in 
emotion monitoring and personality profiling activities, while 
intervention mechanisms incorporate human oversight 
capabilities to prevent manipulative behavioral modification, 
maintaining ethical balance between educational 
personalization and learner autonomy preservation. 

5. Conclusion 

This research establishes significant theoretical and 
practical advances in virtual teacher personalized interaction 
through novel integration of multimodal affective computing 
with distributed multi-agent coordination mechanisms, 
achieving 91.2% emotion recognition accuracy while 
demonstrating substantial educational effectiveness 
improvements, including 43% engagement enhancement, 
37% emotional satisfaction increase, and 40% knowledge 
retention improvement over traditional approaches. The 
developed three-layer distributed architecture addresses 
fundamental scalability limitations, while the confidence-
weighted multimodal fusion algorithm overcomes single-
modality reliability constraints that have historically limited 
emotion-aware educational applications. The Byzantine Fault 
Tolerant consensus adaptation represents a substantial 
theoretical contribution to distributed decision-making 
frameworks, enabling coordinated agent behavior with 31% 
superior decision quality compared to centralized 
approaches while maintaining pedagogical coherence. The 
personality adaptation framework demonstrates variable 
effectiveness across learner populations, achieving 88% 
success rates for extraverted individuals while revealing 
limitations for high-neuroticism learners that highlight 
psychological modeling complexity requirements. Future 
research directions encompass environmental robustness 
optimization to address performance degradation under 
challenging classroom conditions, development of 
sophisticated personality adaptation algorithms for diverse 
psychological profiles, and integration of advanced natural 
language processing capabilities. Expansion potential spans 
corporate training, therapeutic educational applications, and 
cross-cultural learning scenarios requiring enhanced 
localization and cultural sensitivity mechanisms. 
Interdisciplinary collaboration opportunities emerge through 
convergence with cognitive psychology research, 
neuroscience investigations into emotional learning 
mechanisms, and educational policy development addressing 
ethical considerations surrounding emotion monitoring and 
privacy protection, establishing foundations for sustainable 
educational technology evolution, and balancing 
technological advancement with human-centered design 
principles. 

Ethical issue 
The authors are aware of and comply with best practices in 
publication ethics, specifically with regard to authorship 
(avoidance of guest authorship), dual submission, 
manipulation of figures, competing interests, and compliance 
with policies on research ethics. The authors adhere to 
publication requirements that the submitted work is original 
and has not been published elsewhere. 



R. Dang & NA. Samad /Future Technology                                                                    November 2025| Volume 04 | Issue 04 | Pages 159-172 

171 

 

Data availability statement 
The manuscript contains all the data. However, more data will 

be available upon request from the authors. 

Conflict of interest 

The authors declare no potential conflict of interest. 

References 

[1] Pei, G., et al., Affective computing: Recent advances, 

challenges, and future trends. Intelligent Computing, 

2024. 3: p. 

0076.http://dx.doi.org/10.34133/icomputing.0076 

[2] Vistorte, A.O.R., et al., Integrating artificial intelligence 

to assess emotions in learning environments: a 

systematic literature review. Frontiers in psychology, 

2024. 15: p. 

1387089.http://dx.doi.org/10.3389/fpsyg.2024.1387

089 

[3] Sethi, S.S. and K. Jain, AI technologies for social 

emotional learning: recent research and future 

directions. Journal of Research in Innovative Teaching 

& Learning, 2024. 17(2): p. 213-

225.http://dx.doi.org/10.1108/JRIT-03-2024-0073 

[4] Schroeder, N.L. and S.D. Craig, Learning with virtual 

humans: Introduction to the special issue. 2021, Taylor 

& Francis. p. 1-

7.http://dx.doi.org/10.1080/15391523.2020.1863114 

[5] Alzubaidi, L., et al., Review of deep learning: concepts, 

CNN architectures, challenges, applications, future 

directions. Journal of big Data, 2021. 8: p. 1-

74.http://dx.doi.org/10.1186/s40537-021-00444-8 

[6] Lv, Z., et al., Deep learning for intelligent human–

computer interaction. Applied Sciences, 2022. 12(22): 

p. 11457.http://dx.doi.org/10.3390/app122211457 

[7] Lippert, A., et al., Multiple agent designs in 

conversational intelligent tutoring systems. 

Technology, Knowledge and Learning, 2020. 25(3): p. 

443-463.http://dx.doi.org/10.1007/s10758-019-

09431-8 

[8] Kabudi, T., I. Pappas, and D.H. Olsen, AI-enabled 

adaptive learning systems: A systematic mapping of 

the literature. Computers and education: Artificial 

intelligence, 2021. 2: p. 

100017.http://dx.doi.org/10.1016/j.caeai.2021.10001

7 

[9] Létourneau, A., et al., A systematic review of AI-driven 

intelligent tutoring systems (ITS) in K-12 education. 

npj Science of Learning, 2025. 10(1): p. 

29.http://dx.doi.org/10.1038/s41539-025-00320-7 

[10] Lin, C.-C., A.Y. Huang, and O.H. Lu, Artificial intelligence 

in intelligent tutoring systems toward sustainable 

education: a systematic review. Smart Learning 

Environments, 2023. 10(1): p. 

41.http://dx.doi.org/10.1186/s40561-023-00260-y 

[11] Essa, S.G., T. Celik, and N.E. Human-Hendricks, 

Personalized adaptive learning technologies based on 

machine learning techniques to identify learning 

styles: A systematic literature review. IEEE Access, 

2023. 11: p. 48392-

48409.http://dx.doi.org/10.1109/ACCESS.2023.32764

39 

[12] Ortega‐Ochoa, E., M. Arguedas, and T. Daradoumis, 

Empathic pedagogical conversational agents: a 

systematic literature review. British Journal of 

Educational Technology, 2024. 55(3): p. 886-

909.http://dx.doi.org/10.1111/bjet.13413 

[13] Pereira, D.S., et al., Here's to the future: Conversational 

agents in higher education-a scoping review. 

International Journal of Educational Research, 2023. 

122: p. 

102233.http://dx.doi.org/10.1016/j.ijer.2023.102233 

[14] Yusuf, H., A. Money, and D. Daylamani-Zad, Pedagogical 

AI conversational agents in higher education: a 

conceptual framework and survey of the state of the 

art. Educational technology research and development, 

2025: p. 1-60.http://dx.doi.org/10.1007/s11423-025-

10447-4 

[15] Zhang, Y. and W. Pan, A scoping review of embodied 

conversational agents in education: trends and 

innovations from 2014 to 2024. Interactive Learning 

Environments, 2025: p. 1-

22.http://dx.doi.org/10.1080/10494820.2025.246897

2 

[16] Lampropoulos, G., et al., Affective computing in 

augmented reality, virtual reality, and immersive 

learning environments. Electronics, 2024. 13(15): p. 

2917.http://dx.doi.org/10.3390/electronics13152917 

[17] Fernández-Herrero, J., Evaluating recent advances in 

affective intelligent tutoring systems: A scoping review 

of educational impacts and future prospects. 

2024.http://dx.doi.org/10.3390/educsci14080839 

[18] Spice, B., New AI enables teachers to rapidly develop 

intelligent tutoring systems. Carnegie Mellon 

University, May, 2020. 6: p. 

2020.https://www.cmu.edu/dietrich/news/news-

stories/2020/april/intelligent-tutor.html 

[19] Hube, N., et al., A study on the influence of situations 

on personal avatar characteristics. Visual Computing 

for Industry, Biomedicine, and Art, 2024. 7(1): p. 

23.http://dx.doi.org/10.1186/s42492-024-00174-7 

[20] Winkler, R., et al. Sara, the lecturer: Improving learning 

in online education with a scaffolding-based 

conversational agent. in Proceedings of the 2020 CHI 

conference on human factors in computing systems. 

2020.http://dx.doi.org/10.1145/3313831.3376781 

[21] Alqarni, A., Artificial Intelligence‐Critical Pedagogic: 

Design and Psychologic Validation of a Teacher‐

Specific Scale for Enhancing Critical Thinking in 

Classrooms. Journal of Computer Assisted Learning, 

2025. 41(3): p. 

e70039.http://dx.doi.org/10.1111/jcal.70039 

[22] Halkiopoulos, C. and E. Gkintoni, Leveraging AI in e-

learning: Personalized learning and adaptive 

assessment through cognitive neuropsychology—A 

systematic analysis. Electronics, 2024. 13(18): p. 

3762.http://dx.doi.org/10.3390/electronics13183762 

[23] Salloum, S.A., et al., Emotion recognition for enhanced 

learning: using AI to detect students’ emotions and 

adjust teaching methods. Smart Learning 

Environments, 2025. 12(1): p. 

21.http://dx.doi.org/10.1186/s40561-025-00374-5 

http://dx.doi.org/10.34133/icomputing.0076
http://dx.doi.org/10.3389/fpsyg.2024.1387089
http://dx.doi.org/10.3389/fpsyg.2024.1387089
http://dx.doi.org/10.1108/JRIT-03-2024-0073
http://dx.doi.org/10.1080/15391523.2020.1863114
http://dx.doi.org/10.1186/s40537-021-00444-8
http://dx.doi.org/10.3390/app122211457
http://dx.doi.org/10.1007/s10758-019-09431-8
http://dx.doi.org/10.1007/s10758-019-09431-8
http://dx.doi.org/10.1016/j.caeai.2021.100017
http://dx.doi.org/10.1016/j.caeai.2021.100017
http://dx.doi.org/10.1038/s41539-025-00320-7
http://dx.doi.org/10.1186/s40561-023-00260-y
http://dx.doi.org/10.1109/ACCESS.2023.3276439
http://dx.doi.org/10.1109/ACCESS.2023.3276439
http://dx.doi.org/10.1111/bjet.13413
http://dx.doi.org/10.1016/j.ijer.2023.102233
http://dx.doi.org/10.1007/s11423-025-10447-4
http://dx.doi.org/10.1007/s11423-025-10447-4
http://dx.doi.org/10.1080/10494820.2025.2468972
http://dx.doi.org/10.1080/10494820.2025.2468972
http://dx.doi.org/10.3390/electronics13152917
http://dx.doi.org/10.3390/educsci14080839
https://www.cmu.edu/dietrich/news/news-stories/2020/april/intelligent-tutor.html
https://www.cmu.edu/dietrich/news/news-stories/2020/april/intelligent-tutor.html
http://dx.doi.org/10.1186/s42492-024-00174-7
http://dx.doi.org/10.1145/3313831.3376781
http://dx.doi.org/10.1111/jcal.70039
http://dx.doi.org/10.3390/electronics13183762
http://dx.doi.org/10.1186/s40561-025-00374-5


R. Dang & NA. Samad /Future Technology                                                                    November 2025| Volume 04 | Issue 04 | Pages 159-172 

172 

 

[24] Wang, T., et al. Llm-powered multi-agent framework 

for goal-oriented learning in intelligent tutoring 

system. in Companion Proceedings of the ACM on Web 

Conference 2025. 

2025.http://dx.doi.org/10.48550/arXiv.2501.15749 

[25] MADHAVAN, S.M.M.K., et al., Multi-Agent System for 

Cognitive Assessment using Deep Learning. Science 

and Technology, 2025. 

7(02).https://www.ijadst.com/ajradmin/certificates/

467/IJADST_20250479.pdf 

[26] Chu, Z., et al., Llm agents for education: Advances and 

applications. arXiv preprint arXiv:2503.11733, 

2025.http://dx.doi.org/10.48550/arXiv.2503.11733 

[27] Lindberg, S., Using virtual simulations with avatars to 

train pre-service special needs teachers’ relational 

competence: possibilities and limitations. Cogent 

Education, 2025. 12(1): p. 

2457290.http://dx.doi.org/10.1080/2331186X.2025.2

457290 

[28] Son, G., A. Tiemann, and M. Rubo, I am here with you: 

an examination of factors relating to social presence in 

social VR. Frontiers in Virtual Reality, 2025. 6: p. 

1558233.http://dx.doi.org/10.3389/frvir.2025.15582

33 

[29] Strielkowski, W., et al., AI‐driven adaptive learning for 

sustainable educational transformation. Sustainable 

Development, 2025. 33(2): p. 1921-

1947.http://dx.doi.org/10.1002/sd.3221 

[30] Park, J., et al. The Impact of Observer Presence on 

Trainees' Mental States and Performance in Remote 

Military Training with Virtual Humans in Mixed Reality 

Environment. in Proceedings of the 2025 CHI 

Conference on Human Factors in Computing Systems. 

2025.http://dx.doi.org/10.1145/3706598.371351 

 

  

This article is an open-access article distributed under the 

terms and conditions of the Creative Commons Attribution 

(CC BY) license 

(https://creativecommons.org/licenses/by/4.0/). 

http://dx.doi.org/10.48550/arXiv.2501.15749
https://www.ijadst.com/ajradmin/certificates/467/IJADST_20250479.pdf
https://www.ijadst.com/ajradmin/certificates/467/IJADST_20250479.pdf
http://dx.doi.org/10.48550/arXiv.2503.11733
http://dx.doi.org/10.1080/2331186X.2025.2457290
http://dx.doi.org/10.1080/2331186X.2025.2457290
http://dx.doi.org/10.3389/frvir.2025.1558233
http://dx.doi.org/10.3389/frvir.2025.1558233
http://dx.doi.org/10.1002/sd.3221
http://dx.doi.org/10.1145/3706598.371351
https://creativecommons.org/licenses/by/4.0/

