Interpretable Knowledge Tracing

Created: 2026-07-29 | Tags: knowledge-tracinginterpretabilitystudent-modeling
Interpretable Knowledge Tracing โ€” A novel framework for dialogue-based Knowledge Tracing that explicitly models both student ability and tutor-turn difficulty using Item Response Theory, producing interpretable cognitive quantities from LLM output logits. Addresses two critical gaps in prior work: ignored question difficulty and opaque latent representations that undermine tutor trust.

Shuyan Huang, Alexander Scarlatos, Jaewook Lee, Andrew Lan (UMass Amherst) โ€” arXiv:2605.01097, May 2026.

Key Findings

Existing dialogue-based Knowledge Tracing (KT) systems suffer from two fundamental limitations. First, they ignore question difficulty โ€” prior work relies solely on student knowledge states, leading to inaccurate predictions when tutors scaffold or increase task difficulty across turns. Second, they produce opaque latent representations โ€” high-dimensional LLM embeddings lack interpretability, making it impossible for human tutors to understand why the system made a particular prediction, which erodes trust.

Huang et al. propose an interpretable difficulty-aware conversational KT framework built on Llama-3.1-8B-Instruct. The architecture has three modules:

1. Knowledge Estimator โ€” Extracts next-token logits for vocabulary tokens "GOOD" and "BAD" from the LLM's output layer after processing the dialogue history. Student ability ฮธ = z^GOOD โˆ’ z^BAD, producing a scalar that represents the model's latent assessment of the student's current knowledge state.

2. Difficulty Estimator โ€” Similarly extracts logits for "HARD" and "EASY" to compute tutor-turn difficulty d = z^HARD โˆ’ z^EASY. This captures how challenging the current tutor utterance is, independent of the student's ability.

3. IRT-based Predictor โ€” Combines ฮธ and d using a 1PL (Rasch) model with a learnable scalar ฮฑ: p(correct) = 1 / (1 + exp(โˆ’ฮฑ(ฮธ โˆ’ d))). This mirrors classical item-response-theory formulations where the probability of a correct response depends on the difference between ability and difficulty.

The framework was evaluated on two datasets: QATD2k (real-world dialogues from the Eedi math tutoring platform, 1,573 train / 393 test) and MathDial (simulated dialogues with GPT-3.5 students and crowd-sourced tutors, 2,235 train / 588 test). It was compared against five deep learning baselines (DKT, DKVMN, SAINT, AKT, simpleKT) and one LLM-based baseline (LLMKT).

Results show that LLM-based methods substantially outperform deep learning models, and the proposed framework achieves state-of-the-art performance: 64.29% accuracy and 65.25 AUC on QATD2k, 68.82% accuracy and 76.59 AUC on MathDial โ€” modestly but consistently outperforming LLMKT across both datasets and metrics.

The key insight is that both student knowledge and task difficulty are explicitly represented as meaningful cognitive quantities rather than uninterpretable latent vectors. A human tutor can inspect ฮธ and d at any dialogue turn to understand the model's assessment: "the student seems confused (low ฮธ) and this question is particularly hard (high d)." This transparency is essential for building ai-tutoring systems that tutors can trust and act upon.

Implications for AI in Education

This work bridges a critical gap between the predictive power of LLMs and the interpretability demands of real educational settings. While knowledge-tracing has traditionally used deep learning models like DKT and SAINT that trade interpretability for accuracy, Huang et al. demonstrate that LLMs can produce both โ€” competitive or superior accuracy while yielding transparent, inspectable intermediate representations.

The difficulty-aware component addresses a well-known blind spot in student-modeling: student performance is a joint function of knowledge and task demands. By disentangling these, the framework avoids falsely attributing poor performance to low knowledge when a tutor has simply asked a harder question, or conversely, mistaking scaffolded success for mastery. This connects to broader work on knowledge-tracing-irt that uses IRT to add structure to neural KT models.

The logit-extraction approach is notable for its simplicity โ€” rather than training separate classifiers or using complex prompting strategies, it repurposes the LLM's own vocabulary probabilities as cognitive signals. This technique relates to emerging work on using LLM internal representations for educational assessment, including cold-start-knowledge-tracing-safeinsights and neural-symbolic-knowledge-tracing.

For practical deployment, the framework's interpretability enables several downstream applications: tutors can receive real-time alerts when student ability drops below a threshold; difficulty estimates can guide adaptive scaffolding decisions; and longitudinal ฮธ trajectories can track learning over multiple sessions. The explicit difficulty model also supports better ai-tutor-authoring-promptdecipher by helping content creators understand which tutor utterances are most effective at different ability levels.

Future work should extend the framework to multi-skill settings (where students may have different abilities across different knowledge components), explore fine-tuning strategies that further improve the knowledge and difficulty estimators, and validate the approach in live tutoring deployments with real tutor feedback on interpretability quality.

Related Pages