Classmate Agents in AI Classrooms

• Classmate agents are LLM-based avatars with fixed personas and defined pedagogical roles that simulate realistic student interactions in virtual classrooms.
• They operate within multi-agent systems using a session controller and function sets to orchestrate continuous lecturing and coordinated peer dialogue.
• Empirical evaluations show that these agents increase student engagement and learning outcomes, with metrics reflecting improved knowledge acquisition and classroom dynamics.

Classmate agents are LLM-driven peer student avatars embedded in multi-agent educational systems to simulate, scaffold, and enrich interactive learning environments. They are defined by fixed personas, explicit pedagogical roles, and operationalized function sets, enabling the orchestration of AI-augmented classrooms with more naturalistic student-student dynamics. Their deployment has been empirically validated in various platforms, notably MAIC, SimClass, EduAgent, and systems exploring constructively controversial dialogues.

1. Definition and Functional Taxonomy

Classmate agents are instantiated LLM-based peer agents within a "1 student + N AI agents" environment, designed to emulate the cognitive, affective, and social dynamics found in real classrooms. Each classmate agent $a_i \in \mathcal{A}$ is endowed with:

• A fixed persona specified via system and role prompts (e.g., Class Clown, Deep Thinker, Note Taker, Inquisitive Mind)
• One or more pedagogical roles $r \in \mathcal{R}$:
  • Teaching & Initiation (TI)
  • In-depth Discussion (ID)
  • Emotional Companionship (EC)
  • Classroom Management (CM)
• The ability to invoke a constrained classroom function set (e.g., ShowFile, ReadScript, AskQuestion)

The roles and agent archetypes are drawn from Schwanke’s taxonomy and operationalized in prompts to provide both diversity and coverage of peer student behaviors (Yu et al., 2024, Zhang et al., 2024).

2. System Architectures and Interaction Modeling

Classmate agents are deployed within multi-agent LLM-driven architectures featuring at minimum:

• LLM Agent Layer: Teacher, TA, and $N$ distinct classmate agents
• Session Controller: A two-part module with a Class State Receptor (tracking dialogue history $H_t$ and covered slides $P_t$) and a Manager Agent (metacontroller that selects the next agent/action pair $(a^*, \mathcal{T}^*)$)
• Student Interface: A human participant whose utterances are appended to the dialogue state

At each timestep $t$, the state receptor assembles $S_t = \{P_t, H_t, \hat{\mathcal{R}}\}$, with $H_t$ containing all utterances by agents and users up to $t$. The manager agent decision function $f_{\mathcal{L}}: S_t \rightarrow (a, \mathcal{T})$ selects an agent and teaching action, after which the system transitions on either agent/user utterance or timeout $\tau$. This structure supports both continuous lecturing and interleaved peer interaction (Yu et al., 2024, Zhang et al., 2024).

Interaction modeling leverages conversational mechanisms (e.g., Flanders Interaction Analysis), with explicit tagging of utterance roles, and session-level allocation to maintain empirically validated student-talk ratios (ST) (e.g., 13–17%, matching real classroom settings) (Zhang et al., 2024).

3. LLM Prompting and Policy Design

Classmate agent behavior is governed by prompt-engineered policy frameworks which include:

• Role Customization Operator $\rho$: $\mathcal{A} = \rho(\mathrm{LLM}, P_A)$, with $P_A$ specifying persona details, function permissions, and pedagogical goals.
• Run-Time Prompt Scheme: Each agent receives a context window (1–3 prior $H_t$ turns + current slide), a SYSTEM message encoding persona/function goals, and an explicit task instruction (e.g., "produce at most one teaching function call or supportive remark").
• Few-shot In-Context Policy: The prompt includes function-typed demonstrations rather than hard-coded rules, with LLMs learning to map context to role-appropriate utterances (Yu et al., 2024, Zhang et al., 2024).

In some frameworks, such as EduAgent, generative policies incorporate cognitive prior instructions (e.g., "students with low prior knowledge typically report high workload"), requiring the LLM to explicitly reflect on those constraints before emitting behavioral outputs (gaze, motor actions, cognitive-state vector, quiz answer) in a structured JSON format (Xu et al., 2024). This approach factors agent output as:

$$P(B_{1:T}\mid p, C_{1:T}) = \prod_{t=1}^T P(B_t\mid B_{<t}, p, C_t)$$

with further subdivisions for gaze, motor, cognition, and answer as conditionally dependent variables.

4. Emergent Behaviors and Group Dynamics

Classmate agents, when orchestrated under a session-level manager with explicit role and persona diversity, display emergent group behaviors:

• Oscillatory Role Allocation: The stochastic, relevance-based selection mechanism in the manager yields periodic switching among teaching, questioning, summarizing, and encouragement, producing multi-participant dialogue dynamics approximating human classrooms (Zhang et al., 2024).
• Self-correction Mechanisms: Assistant agents or classmates may repair or redirect discourse in response to user disengagement or negative affect, with no explicit programming for such behaviors—emergence results from the joint LLM policy landscape.
• Constructive Controversy: Peer agents equipped with regulated or unregulated argumentation protocols (per Constructive Controversy theory) generate differential engagement and learning outcomes depending on user preference for challenge or control (Tanprasert et al., 20 Jan 2026).

5. Empirical Evaluation

Multiple large-scale deployments have quantitatively validated the efficacy of classmate agents:

Platform	Metric/Outcome	Key Result
MAIC	Manager accuracy	75% w/ role desc.; 60% w/o
MAIC	Student activity (knowledge seeking)	61% of turns; positive correlation with final exam (Corr(log MsgNum, FinalExam) = 0.346***)
SimClass	Student-talk ratio (ST)	0.162 (matches real classrooms 0.17–0.23)
EduAgent	Gaze/motor/cognition/quiz pred. MAE	Gaze dist ≈0.35; Cog MAE (workload) 0.17
SimClass	CoI (cognitive/social presence, max=2)	≈1.8 (full), drops to ≈1.3 w/o classmates
EduAgent	Simulated quiz accuracy (post-test)	∼85% (full); ∼70% (w/o classmates)
CC Peer Agent	Turn count (Regulated > Unregulated)	$t(139.8)=3.682$, $p<.05$, $\eta^2=0.15$

Ablation studies confirm that student engagement and learning outcomes diminish sharply without classmate agents or with reduced agent interaction (Yu et al., 2024, Zhang et al., 2024, Xu et al., 2024). Qualitative interviews underscore benefits in guided inquiry and depth of thinking but also highlight the necessity for robust session management to avoid breakdowns.

6. Best Practices and Design Guidelines

Empirically derived recommendations for robust and effective classmate agent systems include:

• Explicit Role Definition: Including detailed roles in both manager and agent prompts increases function-routing accuracy by ∼15%.
• Persona Diversity: Deploying at least 3–4 archetypal peer agents creates richer and more adaptive discourse.
• Contextual Input: Providing both slide images and text scripts as input to agents improves the relevance and accuracy of agent-generated questions and summaries.
• Balanced Dialogue: Systems must balance continuous content delivery with structured peer interaction, nudging users toward engagement without sacrificing information flow.
• Adaptive Regulation: For contentious or debate-based peer agents, learner orientation (efficiency-driven vs. curiosity-driven) should inform whether regulated or unregulated controversy mechanisms are deployed (Tanprasert et al., 20 Jan 2026).
• Human Oversight: Retaining minimal TA involvement for script revision and factuality checking remains critical, as full automation is not yet robust against all error modes.
• Instrumentation and Data-Driven Refinement: Logging all agent-user exchanges enables continuous refinement of both manager and agent policies (Yu et al., 2024, Zhang et al., 2024).

7. Implementation and Deployment Considerations

Reference implementations (e.g., Slide2Lecture, SimClass, EduAgent) detail comprehensive integration recipes:

• Frontend: Web UI with chat, PPT viewer, selectable function triggers.
• Backend: LLM endpoints (GPT-4V for multimodal, GPT-4/Claude/MinicPM for agent LLMs) hosted on Flask/Torch; retrieval augmented generation with slide/script embedding.
• Persistence: Logging of agent dialogue, state, and session context for evaluation and fine-tuning.
• APIs: Modular endpoints for agent init, simulation on each slide, and trace retrieval.
• Prompt Engineering: Deterministic outputs (T=0), clear persona and cognitive prior encoding, context chunking.

For large-scale deployment, seamless MOOC integration is achieved by embedding agents and controllers as plugins within mainstream players; databases capture all agent and student actions for both research analytics and system adaptation (Yu et al., 2024, Xu et al., 2024).

---

Classmate agents thus represent a foundational component of LLM-driven multi-agent educational platforms, providing scalable, adaptive, and pedagogically grounded mechanisms for simulating peer interaction, sustaining motivation, and improving both engagement and learning outcomes in virtual classrooms (Yu et al., 2024, Zhang et al., 2024, Xu et al., 2024, Tanprasert et al., 20 Jan 2026).