Interpretable Knowledge Tracing

Synthesis: Interpretable Knowledge Tracing — 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 Large Language Models (LLMs) 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) — arXiv:2605.01097, May 2026.

Key Findings

Existing dialogue-based Knowledge Tracing (KT) systems suffer from two fundamental limitations. First, they ignore question difficulty — 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 — 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 — Extracts next-token logits for vocabulary tokens "GOOD" and "BAD" from the LLM's output layer after processing the dialogue history. Student ability θ = z^GOOD − z^BAD, producing a scalar that represents the model's latent assessment of the student's current knowledge state.
2. Difficulty Estimator — Similarly extracts logits for "HARD" and "EASY" to compute tutor-turn difficulty d = z^HARD − z^EASY. This captures how challenging the current tutor utterance is, independent of the student's ability.
3. IRT-based Predictor — Combines θ and d using a 1PL (Rasch) model with a learnable scalar α: p(correct) = 1 / (1 + exp(−α(θ − 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 — 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 θ and d at any dialogue turn to understand the model's assessment: "the student seems confused (low θ) and this question is particularly hard (high d)." This transparency is essential for building AI Tutoring systems that tutors can trust and act upon.

What this means for practice

• Instructors. Read the two reported quantities rather than the bare prediction: student ability θ = z^GOOD − z^BAD and tutor-turn difficulty d = z^HARD − z^EASY are exposed at every dialogue turn, so a low θ is not mistaken for weak knowledge when the tutor has simply asked a harder question.
• Instructors. Use the difficulty estimate to choose the next move: because d is estimated independently of student ability, scaffolding can be raised or lowered on evidence rather than on how the last answer felt.
• Designers. Build threshold alerts on θ and longitudinal θ trajectories into Intelligent Tutoring dashboards and use d to author tutor turns, since a human tutor can inspect both quantities at any turn and act on them.
• Designers. Reuse the model's own next-token probabilities for cognitive signals instead of training separate classifiers, and budget compute accordingly: the framework adds only a 1PL (Rasch) predictor p(correct) = 1 / (1 + exp(−α(θ − d))) on top of Llama-3.1-8B-Instruct.
• Researchers. Validate interpretability with real tutors before deployment — accuracy was established on QATD2k (64.29% accuracy, 65.25 AUC) and MathDial (68.82% accuracy, 76.59 AUC), but no study has yet asked tutors whether θ and d change their decisions, and single-ability Learner Modeling and Adaptive Instruction may not hold in multi-skill settings.

Limitations

• There are no existing difficulty-prediction baselines to compare d against, because this is the first work to explicitly estimate the difficulty of tutor turns in dialogues; the comparison is limited to five deep-learning KT baselines (DKT, DKVMN, SAINT, AKT, simpleKT) and LLMKT at the prediction level.
• Evaluation covers math dialogues only — QATD2k (Eedi platform; 1,573 train / 393 test) and MathDial (2,235 train / 588 test) — with no investigation of other domains such as language learning or computer science.
• MathDial's student turns come from GPT-3.5-simulated students and crowd-sourced tutors rather than real learners, and the framework was not deployed in a live tutoring setting with real tutor feedback on interpretability.
• LLM-based knowledge tracing is much more computationally expensive than the traditional models it slightly outperforms, and the authors note the standing bias risk that students from populations less represented in the training data may receive less accurate assessments.

Citation

Huang, S., Scarlatos, A., Lee, J., & Lan, A. (2026). Interpretable Knowledge Tracing.