OnlineMate: LLM Cognitive Companion

• OnlineMate is an LLM-based multi-agent companion system for online learning that uses Theory-of-Mind and Bloom's taxonomy to enhance cognitive scaffolding.
• It dynamically adapts interaction strategies by inferring students' cognitive and emotional states such as confusion, misunderstanding, or motivation.
• The system employs a multi-layered agent architecture and evaluation protocols to simulate authentic classroom dynamics and promote higher-order thinking.

Searching arXiv for the target paper and closely related work to ground the article in current sources. OnlineMate is an LLM-based multi-agent companion system for cognitive support in online learning. It is designed to simulate peer-like agent roles, adapt to learners’ cognitive states during collaborative discussions, and infer psychological states such as misunderstandings, confusion, or motivation. By incorporating Theory of Mind (ToM), the system dynamically adjusts interaction strategies to support higher-order thinking and cognition, addressing a limitation of earlier AI learning companions whose interactions remained largely conversational and insufficiently adapted to individualized learning and cognitive states (Gao et al., 18 Sep 2025).

1. Educational setting and conceptual orientation

OnlineMate is situated in online learning environments in which students often lack personalized peer interactions, even though such interactions play a crucial role in supporting cognitive development and learning engagement. The motivating claim is that prior LLM-based learning environments can simulate interactive discussion, but that these exchanges often lack insight into the learner’s individualized learning and cognitive states. The reported consequence is reduced student interest in discussions with AI learning companions and reduced capacity to gain inspiration from them (Gao et al., 18 Sep 2025).

The system addresses this problem by combining multi-agent interaction with explicit ToM-based inference. In this formulation, agents do not merely produce locally relevant responses; they attempt to infer latent student states and adapt subsequent turns accordingly. Bloom’s taxonomy is integrated into this process, so that the system does not treat discussion quality solely as semantic adequacy or conversational fluency, but also as movement across cognitive levels: Remember, Understand, Apply, Analyze, Evaluate, and Create (Gao et al., 18 Sep 2025).

A central implication is that OnlineMate is not framed as a replacement for ordinary chat-based tutoring, but as a classroom-like companion system. Its objective is to support collaborative discussion, cognitive scaffolding, and emotional resonance through role-diverse agents rather than through a single undifferentiated assistant.

2. System architecture and agent ecology

OnlineMate is organized as a multi-agent classroom simulation with four principal components: OnlineMate Agents, a Classroom Context Manager, a Classroom Behavior Controller, and an Evaluation Agent. Each OnlineMate Agent represents a peer or teacher persona powered by an LLM and equipped with memory and context. The architecture is intended to preserve role fidelity while enabling heterogeneous participation patterns, including peer-to-peer discussion rather than only teacher-student exchange (Gao et al., 18 Sep 2025).

The Classroom Context Manager governs the contextual information visible to each agent and maintains agent-specific memories, beliefs, inferences, and role definitions. Each context entry is represented as $(pk, ctx, range, role, type, timestamp)$. Contexts may encode dialogs, memories, or inferences, and they are visible only to relevant agents or groups. This mechanism is designed to ensure epistemic independence and consistent role-play. In effect, different agents are not assumed to share a single global conversational state; they operate under controlled informational asymmetries appropriate to their personas and roles (Gao et al., 18 Sep 2025).

The Classroom Behavior Controller determines which agent acts at each dialogue turn and how that action is realized. It preserves natural classroom dynamics and persona fidelity by prompting each agent to score its own intent to act on a 0–10 scale, selecting the next speaker from the top-$n$ candidates to introduce stochasticity, and then choosing a permissible role-specific action, possibly directed toward another agent. The chosen action type is included in the prompt used for response generation. This yields a turn-taking mechanism that is neither fully scripted nor fully unconstrained (Gao et al., 18 Sep 2025).

The Evaluation Agent is also LLM-powered. It simulates student users for automated evaluation and judges the effectiveness of OnlineMate responses together with their cognitive and emotional impact. The paper presents this as supporting robust, ethics-compliant experimentation (Gao et al., 18 Sep 2025).

3. Theory-of-Mind reasoning pipeline

At the core of OnlineMate is a three-stage, ToM-driven reasoning pipeline described as metacognition-inspired. The pipeline begins with ToM hypothesis generation. Given the current student utterance $s_r$, prior context $ctx_{<r}$, and prior memory $mem_{<r}$, an agent generates candidate hypotheses about the student’s cognitive and psychological states:

$$\text{Input: } (s_r, ctx_{<r}, mem_{<r}) \longrightarrow \{(h_i, \text{ToM}_i, \text{Bloom}_i)\}_{i=1}^k$$

Each hypothesis is tagged with a ToM label drawn from Belief, Desire, Intention, Emotion, or Thought, and each student utterance is also analyzed with respect to Bloom’s taxonomy. This pairing makes the latent-state inference explicitly both psychological and cognitive (Gao et al., 18 Sep 2025).

The second stage is hypothesis refinement and filtering. Each candidate hypothesis $h_i$ is revised into $\tilde{h}_i$ under persona and context or norm constraints:

$$\tilde{h}_i = \text{Refine}(h_i; \text{persona}, \text{context})$$

These refined hypotheses are then scored for plausibility according to contextual fit and memory consistency, and the highest-scoring one is selected:

$$\tilde{h}^* = \arg\max_{i} \text{Score}(\tilde{h}_i)$$

This stage is important because the same student utterance may require materially different interpretations depending on whether the responding agent is, for example, a teacher or a peer (Gao et al., 18 Sep 2025).

The third stage is response generation and self-validation. Conditioned on $n$0 together with memory, context, and role, the agent generates a response explicitly scaffolded to increase the student’s cognitive level according to Bloom. The response is then self-evaluated for usefulness, relevance, and tone; if the internal score falls below a threshold, the system regenerates or refines the output. This makes OnlineMate more than a standard prompt-and-response architecture: it includes an explicit internal validation loop tied to pedagogical objectives (Gao et al., 18 Sep 2025).

4. Cognitive adaptation and classroom interaction strategies

OnlineMate’s adaptive behavior is based on explicit inference and tracking of both short-term and longer-term learner states. Short-term state is estimated from the current utterance, while longer-term state is grounded in memory. The target states include confusion, misunderstanding, and motivation. These inferred states are then used to tailor responses for cognitive advancement and emotional resonance rather than merely for topic continuation (Gao et al., 18 Sep 2025).

The system’s interaction strategies are distributed across several control layers. The Behavior Controller promotes diverse and realistic turn-taking; persona control maintains role-appropriate and non-repetitive classroom conversation; and the multi-agent design supports richer peer-to-peer discussions. The paper gives illustrative classroom-role dynamics such as teachers answering in an instructional manner and introverted students potentially remaining silent. These behaviors are part of the attempt to simulate authentic classroom dynamics rather than simple multi-speaker text generation (Gao et al., 18 Sep 2025).

A recurrent misconception would be to interpret OnlineMate as a generic multi-agent chatroom. Its defining claim is narrower and more specific: it is a learning companion system whose multi-agent structure is subordinated to cognitive support. The combination of ToM labels, Bloom-level analysis, agent memory, and controlled turn-taking is intended to make responses sensitive to where a learner is cognitively and psychologically, not just to what topic is being discussed.

The design also emphasizes scaffolding “just beyond” the learner’s current level. The paper associates this with fostering deep dialog involving synthesis, evaluation, and creation rather than rote recall. This suggests a pedagogical strategy closer to structured cognitive progression than to open-ended social conversation (Gao et al., 18 Sep 2025).

5. Evaluation protocol and empirical results

Evaluation is divided into automated and human components. In the automated setting, the Evaluation Agent simulates students with specified personas, dialog histories, learning obstacles, and goals. After each turn, LLMs assess cognitive level on Bloom’s taxonomy using a 1–6 scale, where 1 corresponds to Remember and 6 to Create. An emotional score is also updated on a 0–100 scale; if an emotional floor is reached, the dialog terminates. The principal automated metrics are average cognitive and emotional progression per discussion (Gao et al., 18 Sep 2025).

Human evaluation uses standard collegiate discussion rubrics from CMU and Northwestern University. Experts independently score agent performance on participation, comment quality, critical analysis, and dialog. The combination of automated and expert-scored evaluation is methodologically notable because the system’s target construct includes both cognitive advancement and social-discursive quality (Gao et al., 18 Sep 2025).

The reported quantitative results compare three systems: Vanilla multi-LLM, OnlineMate w/o cog (ToM), and OnlineMate (full). OnlineMate (full) achieves a cognitive score of 5.20, an emotion score of 61.66, and a rubric average of 2.47. The corresponding values are 4.40, 48.33, and 2.39 for OnlineMate w/o cog (ToM), and 4.10, 27.50, and 2.05 for Vanilla multi-LLM. The paper further states that OnlineMate users advanced on average a full cognitive tier over vanilla multi-agent LLMs, and that with full ToM plus cognitive scaffolding, students’ utterances rose from “Analyze” to between “Evaluate” and “Create” (Gao et al., 18 Sep 2025).

Additional analytical findings concern interaction scale. Cognitive engagement peaked after approximately five turns per topic, which the paper interprets as an optimal discussion length for maximizing deep thinking. Peer diversity was found to facilitate richer discussions, but above 4+ simultaneous agents, cognitive overload reduced effectiveness. These observations are relevant to deployment because they indicate that more interaction is not automatically better; discussion depth depends on maintaining a balance between diversity and overload (Gao et al., 18 Sep 2025).

6. Significance, disambiguation, and related research directions

OnlineMate’s main contribution is the integration of LLM-based multi-agent interaction with a staged ToM process oriented toward cognitive support. Within the paper’s framing, this shifts AI companionship in education from generic conversation toward psychologically aware, cognitively scaffolded peer interaction. A plausible implication is that the system serves as a blueprint for online educational agents that must jointly model role, memory, learner state, and discussion trajectory rather than optimizing only next-turn relevance (Gao et al., 18 Sep 2025).

The system should also be distinguished from similarly named but substantively different research programs. “MATE: LLM-Powered Multi-Agent Translation Environment for Accessibility Applications” concerns multimodal accessibility and modality conversion, including an Interpreter agent and expert agents for tasks such as text-to-speech, image-to-text, and video-to-text; it is not an educational ToM-based companion system (Algazinov et al., 24 Jun 2025). “MATE: Meet At The Embedding -- Connecting Images with Long Texts” is a vision-language retrieval method for aligning images with long texts through a projection module and multi-stage embedding alignment, and is unrelated to classroom interaction or learning companionship (Jang et al., 2024).

A related extension direction appears in “OmniResponse: Online Multimodal Conversational Response Generation in Dyadic Interactions,” which is explicitly presented as relevant to agents such as OnlineMate because it enables synchronized verbal and non-verbal response generation through Chrono-Text and TempoVoice (Luo et al., 27 May 2025). This suggests a pathway by which an OnlineMate-like system could move from text-centric classroom simulation toward richer spoken and audiovisual interaction, although such multimodal embodiment is not part of OnlineMate as described in the target paper.

Methodologically, the paper’s use of an LLM-powered Evaluation Agent in simulated learning scenarios may invite scrutiny about evaluator-model coupling. However, the inclusion of human evaluation using collegiate discussion rubrics provides an independent assessment channel. This suggests that the reported improvements are not supported by a single evaluation protocol alone, but by both automated cognitive-affective scoring and expert judgment (Gao et al., 18 Sep 2025).