CogEvolution: A Human-like Generative Educational Agent to Simulate Student's Cognitive Evolution

Synthesis: Zhang et al. (2026) propose CogEvolution, a generative educational agent that simulates a student's cognitive evolution rather than merely cloning observable behavior. It couples three components — an ICAP-based cognitive depth perceptron that quantifies cognitive engagement, an IRT-driven memory retrieval module that simulates the assimilation of new and prior knowledge, and an evolutionary-algorithm state update mechanism constrained by the Zone of Proximal Development — to model the dynamic transition from "static persona" to "dynamic cognitive flow." On a new CogMath-948 dataset of 1,245 real eighth-grade students, CogEvolution outperforms static persona and Knowledge Tracing baselines on behavioral fidelity and learning-curve fitting (R²LC = 0.92), reproduces plausible cognitive evolutionary paths consistent with educational psychology, and offers a more interpretable paradigm for student simulation.

The paper argues that existing educational agents rely on static personas that parametrically fit a snapshot of student ability, neglecting the dynamic fluidity of knowledge internalization, transfer, and cognitive state transitions that characterize real learning.

Key Findings

1. A paradigm shift from static persona to dynamic cognitive flow. CogEvolution is presented as the first computational agent grounded in educational psychology that focuses on the dynamic evolution of cognitive states and cognitive paths during agent learning activity, moving beyond the behavioral cloning that dominates prior student Simulation work.
2. ICAP-based cognitive depth perceptron quantifies learner engagement. The agent's cognitive adapter maps interaction features to a probability distribution across the four ICAP levels (passive/active/constructive/interactive) and computes a weighted continuous Cognitive Evolution Rate that determines the step size of subsequent cognitive state updates.
3. IRT-driven memory retrieval models assimilation. A hybrid IRT-based structural similarity (the integral distance between Item Characteristic Curves near the learner's latent ability) combined with semantic similarity lets the agent retrieve prior knowledge to solve structurally similar problems or generate "confusion"/"exploration" states when prior knowledge is insufficient.
4. Evolutionary update mechanism simulates cognitive change. Following a Mutation-Selection-Update paradigm, the LLM acts as a mutation operator generating cognitive hypotheses, a fitness function penalizes leaps or stagnation deviating from the Zone of Proximal Development, and Differential Evolution integrates the winning hypothesis into the knowledge structure at the ICAP rate.
5. CogMath-948 and strong empirical results. The new dataset (1,245 eighth-grade students over six months with cognitive state annotations) shows CogEvolution matching the KT-augmented PEERS baseline on AUC (0.80 vs 0.82) while far exceeding it and static agents on Mistake Precision (76.8%) and learning-curve fit (R²LC = 0.92).

The Cognitive Dynamics Disconnect

CogEvolution responds to what the authors call the "Cognitive Dynamics Disconnect" in generative agents. Mainstream architectures (e.g., the SmallVille/Stanford Town paradigm of generative agents) model social behavior breadth but neglect the internal cognitive architectures driving it, and in education this surfaces as two failures: Missing Cognitive Features and Cognitive Path Stagnation. Current educational agents — whether base LLMs used directly or LLMs combined with deep neural networks — fit a static parametric snapshot of ability, missing the essence of learning as a continuous reconstruction of cognitive schemas through assimilation and accommodation (Piaget). This connects to the wider validity concerns in the student-simulation literature about models that reproduce surface behavior without epistemic fidelity.

Architecture

CogEvolution comprises three tightly coupled components: the Cognitive Adapter (a depth perceptron serving as the agent's perception entry point), the IRT-Driven Memory Retrieval Module, and the Evolutionary Algorithm-Based Cognitive State Update Mechanism. The agent maintains both a long-term stable Knowledge Structure and a transient Cognitive State, and outputs mastery-confidence probabilities to support downstream Knowledge Tracing-style assessments such as AUC/RMSE.

Cognitive Adapter

The adapter extracts Multimodal AI features from interaction logs — cognitive-semantic features (via Sentence-BERT embeddings of reflections and explanations) and behavioral-operation features (interaction type, response latency). An attention-based hierarchical mapping projects these into a probability distribution over the four ICAP engagement levels, from which a weighted, continuous Cognitive Evolution Rate (with constructive engagement weighted higher, e.g., 1.5) is computed via a SoftPlus transform to allow smooth state transitions.

IRT-Driven Memory Retrieval

The retrieval module simulates the psychological mechanism of "connecting new and old knowledge." It computes a hybrid similarity between the current question and each memory, combining semantic similarity with an IRT-based structural similarity defined as the integral distance between Item Characteristic Curves near the learner's current ability. High structural similarity triggers retrieval of the old schema (stable cognitive state); low structural similarity with high perceived engagement generates a "confusion" or "exploration" state that triggers evolutionary updates — an explicit mechanism for cognitive state and misconception reproduction.

Evolutionary Cognitive State Update

Treating the current knowledge structure as the parent individual, the system triggers mutation when cognitive conflict is perceived, using the Large Language Models (LLMs) as a mutation operator to generate candidate hypotheses (e.g., "Is the formula wrong?" or "Is this a new question type?"). A fitness function, incorporating Vygotsky's Zone of Proximal Development as an effective search radius, penalizes both excessive leaps and stagnation; tournament selection picks the best hypothesis, and a Differential Evolution step integrates it into the knowledge structure scaled by the ICAP cognitive evolution rate. A Behavior-Cognition Alignment Check then measures consistency between the generated internal state and external behavior, flagging "cognitive dissonance" (e.g., low confidence but forced answers).

CogMath-948 Dataset and Evaluation

Because suitable open datasets were lacking, the authors constructed CogMath-948 from real educational settings: learning data from 1,245 eighth-grade students over six months, with four dimensions per interaction — question/response information, reflective explanations, ICAP-based cognitive engagement labels, and fine-grained misconception classification. This makes it well suited for research on fine-grained cognitive evolution and Cognitive Diagnosis.

Three research questions drive the evaluation: RQ1 (ability prediction accuracy and mistake fidelity vs. static persona and KT models), RQ2 (reproduction of non-linear learning trajectories conforming to the Power Law of Practice), and RQ3 (validity of the simulated cognitive evolution and module contributions). Baselines include general LLMs (Llama 3 70B, GPT-4o), a reasoning-enhanced model (Gemini 3 Pro Think), and a reproduced PEERS (BKT+LLM) Knowledge Tracing-augmented agent.

Findings in Detail

On task performance (RQ1), CogEvolution's AUC (0.80) is comparable to PEERS (0.82) and both far exceed static agents (0.65–0.73), while its Mistake Precision (76.8%) is the highest — static agents produce "hallucinatory" or "logical leap" errors, whereas ICAP perception and memory retrieval let CogEvolution reproduce the typical misconceptions of real students. On learning dynamics (RQ2), static agents show a nearly flat trajectory (R²LC = 0.45), the KT-augmented PEERS model a downward but overly linear one (0.78), and CogEvolution a curve tightly aligned with ground truth (R²LC = 0.92) that exhibits an early steep drop followed by stabilization — replicating the novice-to-proficient transition predicted by the Power Law of Practice through early "cognitive spikes" and later fine-tuning.

The ablation study (RQ3) isolates each module's contribution: removing the ICAP depth perceptron collapses R²LC to 0.58 (the agent can no longer distinguish shallow from deep learning); removing structured retrieval cuts Mistake Precision by 12.3% (knowledge assimilation is essential for reproducing Misconceptions about AI); and removing the evolutionary update falls back to static persona modeling (R²LC = 0.51, Align 0.76).

What this means for practice

• Software developers. Model the learner's cognitive state as dynamic rather than storing a static persona profile: the ablation shows learning-curve fit collapsing from R²LC = 0.92 to 0.58 once the ICAP depth perceptron is removed.
• Software developers. Retrieve prior knowledge by structural similarity — the integral distance between IRT characteristic curves near the learner's current ability — as well as by semantics, which offers learning analytics an interpretable alternative to purely data-driven knowledge tracing; removing structured retrieval cut Mistake Precision by 12.3 percentage points against 76.8% for the full system.
• Software developers. Emit a behavior-cognition alignment check with every prediction and flag dissonance such as low confidence paired with forced answers; the agent scored 0.91 on alignment against 0.75 for the knowledge-tracing baseline.
• Researchers. Use the simulator as a testbed for instructional strategies and for training novice teachers in low-risk settings rather than as a substitute for measured outcomes, and report simulation validity separately, since its cognitive paths are checked against one cohort's responses rather than against learning gains.

Limitations

• Evaluation rests on a single self-constructed dataset: 1,245 eighth-grade students over six months in one educational setting, with ICAP engagement labels and misconception categories supplied as annotations whose quality bounds what the fidelity metrics can show.
• Mastery-prediction accuracy does not exceed the knowledge-tracing baseline (AUC 0.80 vs. 0.82; RMSE 0.37 vs. 0.35), so the advantage is confined to mistake fidelity and learning-curve fit rather than ability estimation.
• There is no external benchmark or independent replication: R²LC = 0.92 and Mistake Precision = 76.8% are measured against one cohort, not against other grades, subjects, or datasets.
• Several mechanisms are design choices rather than fitted quantities — for example the constructive-level gain coefficient given as 1.5 and the ZPD penalty in the fitness function — and generated hypotheses are bounded by the LLM used as the mutation operator.

Citation

Zhang, W., Cheng, Y., Ye, Z., & Huang, K. (2026). CogEvolution: A human-like generative educational agent to simulate student's cognitive evolution. arXiv preprint.