INSIDE the Student's Mind: Jointly Modeling Latent Reasoning and Action in LLM Student Simulators

Synthesis: Niousha, Kang, & Norouzi (2026) introduce INTERNAL STUDENT DIALOGUE (INSIDE), a student modeling framework that fine-tunes LLMs to both act like students and think like them. Two students may submit identical work for entirely different reasons, so INSIDE generates internal dialogue grounded in Bloom's Taxonomy across cognitive, affective, and action dimensions, fine-tuning on paired think-traces and actions. Evaluated against prompting baselines, INSIDE improves action fidelity (matching real students' code generation) and reasoning alignment (up to 57.9% across models). The work advances Learner Modeling and Adaptive Instruction and Simulating Students research toward cognitively faithful student simulation.

Key Findings

1. LLM-based student simulators reproduce observable actions but fail to model the latent reasoning behind them; two students can submit identical code for entirely different reasons, so action alone is a partial view of student cognition.
2. INSIDE fine-tunes models to generate internal dialogue grounded in Bloom's Taxonomy across cognitive, affective, and action dimensions, trained on paired think traces and actions that are reconstructed via retrospective inference with a teacher model.
3. On action fidelity, fine-tuned INSIDE models produce code distributions closest to real students (lowest Wasserstein distance) on familiar problems and match plain fine-tuning on unseen problems, while prompting baselines maintain artificially high pass rates.
4. On reasoning quality, INSIDE achieves the highest alignment between generated internal dialogue and ground-truth code edits (up to 57.9%), and—unlike prompting baselines—does so without sacrificing action fidelity, bridging explanation and behavior.

The Reasoning Gap in Student Simulation

LLM-based simulators increasingly reproduce human behavior across domains, and in education they promise improved fidelity in predicting student actions for building tutoring systems. Yet such simulations are often limited to replicating surface-level patterns rather than modeling the latent processes underpinning observable outcomes. In education, understanding the reasoning behind an action is crucial for diagnosing Misconceptions about AI, generating targeted Feedback, and evaluating tutoring systems—applications that go back to classic diagnostic models of procedural bugs.

The problem is that observable actions provide only a partial view of cognition: two students may arrive at the same solution through entirely different reasoning, or produce the same incorrect answer for fundamentally different reasons. Psychology research highlights the value of accessing internal reasoning processes via think-aloud protocols, but such methods are expensive and rarely scale. Moreover, existing work on Large Language Models (LLMs) reasoning has focused on improving correctness, pushing models toward logically consistent and factually accurate outputs—whereas human actions are often not rational or correct at all. People make errors, hold misconceptions, and apply incomplete strategies, and in education modeling this incorrect or partial reasoning is essential because mistakes are central to the learning process.

Related Work

Prior work on Simulating Students has conditioned language models on virtual personas or fine-tuned them on real student code trajectories to capture realistic error patterns, stylistic variation, and incremental revision behavior. However, this line of work tends to treat student behavior as a black box, predicting outcomes without modeling the internal reasoning that produces them. Closest to INSIDE, related diagnostic approaches model incorrect student reasoning by inferring misconceptions from erroneous answers, but do so reconstructively—explaining errors post-hoc rather than anticipating the intent that precedes each attempt.

INSIDE instead introduces internal dialogue as an intermediate cognitive layer that precedes student actions during learning interactions. Its generated reasoning traces are a form of verbalized chain-of-thought, but trained for fundamentally different goals than the standard CoT used to improve correctness: here the aim is to replicate the potentially erroneous, uncertain reasoning real students produce. Structuring this simulation data draws on Bloom's Taxonomy, whose hierarchical classification of cognitive objectives motivates explicitly organizing simulation data around cognitive levels rather than only lower-order skills.

The INSIDE Framework

INSIDE generates reasoning followed by action. Given a student's prior submission history and the corresponding AI tutor Feedback, a language model generates an internal dialogue representing the reasoning process leading to the next code submission, and conditions the next action on that inferred dialogue. This key affordance—explicitly modeling the unobservable latent reasoning preceding each action—distinguishes INSIDE from outcome-only student models.

Because ground-truth reasoning does not exist in real student logs, training data is produced by retrospective inference: a teacher model conditions on the interaction context and the observed student-generated next submission to infer the student's internal state in third person, producing structured summaries of cognitive, affective, and action states inspired by the three domains of Bloom's Taxonomy. Given these inferred states, the model then generates a first-person internal dialogue (a think trace) reflecting the student's reasoning prior to the observed submission. This two-stage procedure approximates latent cognition even though the true reasoning was never recorded.

Data and Problem Formulation

The authors study student code generation in an introductory programming course at UC Berkeley (~900 students per semester). Students complete ~10 homework assignments of 3–6 problems each, an autograder gives instant test-case feedback, and students may resubmit freely, producing a multi-attempt submission stream per student–problem pair. Submissions that fail a test may also receive natural-language feedback from an LLM-based AI tutor when the student consents. Training uses Spring 2025 data and testing uses Spring 2024 data, with two subsets: test_OP (unseen students on familiar problems) and test_NP (unseen students on unseen problems). The model generates the next code submission conditioned on up to k prior submissions with their tutor feedback.

Evaluation

INSIDE is baselined against different prompting frameworks and evaluated on two axes: fidelity of simulated actions and quality of generated internal dialogue.

Action fidelity compares model-generated code to real student submissions using functionality and stylistic metrics—pass rate against the autograder suite, code length, abstract syntax tree depth and width, and PEP 8 violations—reporting Wasserstein distance for distributional similarity. Reasoning quality is measured via alignment: an LLM judge decomposes each generated internal dialogue into atomic claims representing intended actions and checks whether each claim is reflected in the ground-truth code diff. The judge was validated against teacher-generated traces (95.2% coverage) and against human annotations (88.0% agreement, κ = 0.754), indicating it reliably attributes stated reasoning to observed code changes.

Results

Fine-tuned models substantially outperform prompting-based methods on action fidelity, and within prompting approaches, more structured strategies improve alignment. On test_OP, INSIDE consistently achieves the lowest Wasserstein distances across all metrics; on test_NP, INSIDE and plain fine-tuning perform comparably, largely because that split contains more successful submissions where fine-tuning already reduces the over-competence bias of prompted models. Pass-rate trajectory analysis shows real students start low and rise sharply near the final step; fine-tuned models track this incremental progress, while prompting models maintain artificially high pass rates (~80%) throughout.

On reasoning quality, INSIDE achieves the highest alignment across both settings: 51.8% on test_OP and 57.9% on test_NP, outperforming the strongest BloomCoT prompting baseline. Notably, larger and more capable models such as GPT-5 tend to score lower, suggesting stronger reasoning ability does not automatically translate into reasoning that matches student-like code edits. INSIDE achieves both high alignment and strong action fidelity—producing reasoning consistent with observed edits while generating code that follows realistic student progression—rather than optimizing alignment at the expense of realistic behavior.

What this means for practice

• Designers. Surface the internal dialogue, not just the predicted action: INSIDE reached 51.8% (test_OP) and 57.9% (test_NP) alignment between generated reasoning and real code edits without losing action fidelity, so the traces can expose likely misconceptions.
• Researchers. Evaluate tutors against reconstructed student reasoning before deployment, using the think traces to test whether feedback resolves a misconception rather than only whether the next submission passes.
• Designers. Score simulators on the pass-rate trajectory, not raw accuracy: prompting baselines held ~80% pass rates from the first attempt while real students started low and rose late, and INSIDE's best mean absolute error was 0.094.
• Designers. Prefer fine-tuned open models over larger prompted ones for student-like reasoning — GPT-5-BloomCoT reached only 45.5% alignment versus 51.8% for Qwen2.5-7B-INSIDE — and reuse the traces for reasoning-based clustering instead of surface behavior.
• Researchers. Validate the LLM judge before trusting alignment scores: it was checked against teacher traces (95.2% coverage) and human annotations (88.0% agreement, κ = 0.754).

Limitations

• Internal dialogue is reconstructed, not observed: the think traces come from teacher-LLM retrospective inference over 6,911 training submissions from 445 students, so no ground-truth reasoning exists to check them against; the authors propose validating them through think-aloud or retrospective verbalization studies.
• Because LLMs are trained toward expert-like reasoning, the generated traces may read as more coherent than real novice reasoning, and even INSIDE leaves about 42% of generated claims unexplained by the student's code edits.
• The two test splits are not clean: test_OP (5,262 submissions) and test_NP (1,054 submissions) differ in their student pass-rate distributions, and test_NP's standard errors (≈2.0–2.3 vs. ≈0.8–0.9) limit direct comparison between settings.
• Results come from one introductory programming course at a single university (UC Berkeley, ~900 students per semester, two semesters), and reasoning quality is scored by an LLM judge whose agreement with human annotations was 88.0% (κ = 0.754).

Citation

Niousha, R., Kang, M., & Norouzi, N. (2026). INSIDE the student's mind: Jointly modeling latent reasoning and action in LLM student simulators.