acceptodds
Under review as a conference paper at ICLR 2027

Learning How Students Update: Graph-Conditioned Meta-Optimization of Cognitive State for Knowledge Tracing

Abstract

Knowledge tracing predicts a learner's next response from a latent knowledge state and is a core component of adaptive tutoring systems. Recent models increasingly focus on richer representations of interaction histories, but the mechanism for updating the learner state is still typically fixed by the model architecture. On concept-level benchmarks, however, more expressive encoders do not consistently outperform simple recurrent models. We revisit this design choice and argue that a learner's response provides directional evidence about mastery, so the cognitive state should evolve through a learned, preconditioned gradient step on the learner's own prediction error. GRACE represents the cognitive state in concept coordinates, differentiates the prediction loss with respect to the state, and uses a graph-conditioned meta-optimizer to transform the resulting gradient into the next state. The meta-optimizer controls per-concept retention, adaptive step size, graph-based gradient diffusion, and diagonal-plus-low-rank preconditioning from the interaction context. Experiments on ASSIST2009, Junyi, and the AAAI2023 challenge data show that GRACE achieves the strongest overall predictive performance against thirteen baselines spanning recurrent, graph-based, and attention-based knowledge tracing models. Ablation studies identify the learned update mechanism as a major source of the performance gains, while sensitivity analyses consistently favor narrow embeddings.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
knowledge tracing, meta-optimization, cognitive state modeling, educational data mining

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_hyTkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduced GRACE, a knowledge tracing framework that treats cognitive-state evolution as a learned optimization process. GRACE recasts state evolution as a learned gradient step on the learner's own BCE loss: the state-gradient in concept coordinates supplies the update direction, and a graph-conditioned meta-optimizer learns the update rule from interaction context — emitting a per-concept retention gate, step size, graph-diffusion coefficient, and a diagonal-plus-low-rank preconditioner — so the same concept graph both represents concepts and propagates gradients. Experiments on multiple public datasets confirm the effectiveness of the proposed method.

Critical strengths

1.This paper turns the state transition from an architectural by-product into an explicit, learnable modeling component. 2.This paper supplies the update direction directly via the gradient, while a context-dependent meta-optimizer learns retention and magnitude. 3.This paper explicitly controls how updates propagate across concepts. 4.Experiments on three public datasets confirm the effectiveness of the proposed method.

Critical weaknesses

1.The model is over-engineered: a large stack of learned control quantities without a compelling account of why each is needed. 2.For a knowledge-tracking paper, the design is not grounded in educational or learning theory. 3.The central claim is never tested against the most obvious alternative. Every ablation retains the gradient path, so the paper never compares its gradient-guided update against a parametric transition of matched parameter count within the same framework. Without that comparison, the claim that "using the gradient" — rather than the added gating/preconditioning capacity — is what matters remains unverified, and this is exactly the comparison that would settle the paper's thesis. 4.Reported results lack variance and significance testing, and the reported protocol is not fully consistent. Results are single runs without standard deviations or significance tests, while several margins are only 0.1–0.45 AUC points and thus within plausible run-to-run variation. Separately, the preprocessing folds items into "number of items = number of concepts" (117/117, 715/715) and reports baseline AUCs far below values common in the public literature, contradicting the stated pyKT-protocol comparability and making cross-paper comparison unreliable. No parameter-count or training-cost comparison is given either, weakening the "better than wider representations" argument.

Questions for authors

1.All ablations retain the gradient path, so the paper never compares the gradient-guided update against a parametric transition of matched parameter count under the same framework. Can you provide this comparison — replacing g_t with a learned parametric map of equal capacity but keeping gating, diffusion, and preconditioning fixed? 2.In the claimed "learned optimizer" setting, an update is required at every interaction, yet no parameter-count or training-cost comparison is reported. What are the parameter counts and training/inference time per epoch relative to the strongest baseline, and how much of the gain survives when the baseline is given an equal parameter and compute budget? 3.Results appear to be single runs with no standard deviations or significance tests, while several margins are only 0.1–0.45 AUC points. Can you report variance across multiple seeds and significance tests against the strongest baseline, and clarify the inconsistency between the reported preprocessing and the claimed pyKT-protocol comparability?

Reviewer_7GqhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes GRACE, a knowledge tracing framework that models cognitive state updates through a graph-conditioned meta-optimizer. Specifically, it uses the gradient of prediction loss to guide cognitive state updates, incorporating adaptive retention, graph-based gradient diffusion, and low-rank preconditioning. Experiments on three benchmark datasets demonstrate improvements over existing knowledge tracing methods.

Critical strengths
  1. The overall framework is clearly presented, and using prediction gradients to guide cognitive state updates is an interesting idea.

  2. The paper compares with 13 representative knowledge tracing methods and achieves the best AUC on all three datasets.

Critical weaknesses
  1. The motivation is not suffiently convincing, as the paper does not clearly explain why existing state transition methds are an important limitation of knowledge tracing or why gradient-based updates are necessary.

  2. It is unclear whether updating the latent cognitive state based on prediction-loss gradients can genuinely reflect learners' knowledge acquisition and forgetting behaviors, rather than simply improving prediction accuracy.

2.The experiments lack a direct comparison with conventional state transition mechanisms under the same representation and model complexity, and no statistical significance tests or computational efficiency comparisons are provided.

Questions for authors

See above weakness

Reviewer_rkN5OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes GRACE, a knowledge tracing framework that models how students' cognitive states change after each response. Instead of relying on a conventional state transition network, GRACE uses prediction loss gradients and a graph conditioned meta optimizer to update concept level mastery states. Experiments on three datasets show the best AUC among 13 baselines, with ablations suggesting that the learned update mechanism is a major source of the gains.

Critical strengths

1.The paper focuses on an important but relatively underexplored aspect of knowledge tracing: how students' knowledge states should be updated after each interaction. Treating the response as an error signal rather than an opaque input is an appealing view of KT. As far as I know, applying a learned optimizer as a KT transition is new. 2.The framework combines gradient guided updates, concept graph propagation, and adaptive preconditioning. These components give the model explicit control over retention, update magnitude, and information transfer between related concepts. The ablation study offers evidence that these mechanisms contribute to performance. 3.The experiments compare GRACE against 13 baselines across three datasets. GRACE achieves the best AUC on all three benchmarks. The paper acknowledges that ACC on ASSIST2009 is below DKT’s, that Junyi behaves differently in the ablation, and that Junyi ACC rises when the preconditioner is removed.

Critical weaknesses

1.Although the gradient guided update is interesting, the method combines established ideas from learned optimization, graph propagation, etc. The paper needs to clarify more convincingly what is fundamentally new. The distinction from related learned optimizer approaches could be stronger. 2.I think one key ablation is missing. The ablation removes the preconditioner, graph diffusion, graph encoder, and concept attention, but never removes the gradient itself, the retention gate, or the meta-optimizer. Nothing tests “gradient-guided update vs. a learned parametric transition” with everything else held fixed. 3.Several baseline numbers look off relative to what I recall from the pyKT literature. AKT, simpleKT, and the other attention models fall well below DKT on ASSIST2009 and Junyi. 4.Phrases like “direct evidence that the learned transition is a major source of the gains” are stronger than the ablation supports. 5.The model's cognitive states are interpreted as concept level mastery, but the paper does not directly validate whether the learned updates correspond to meaningful learning or forgetting. The authors acknowledge this as future work. Without such validation, the claimed cognitive interpretation remains plausible rather than empirically established.

Questions for authors

1.Please report mean ± std over multiple seeds and splits, with significance tests. Are ASSIST2009 gains robust given the small margin? 2.Why do attention baselines say, AKT, simpleKT, etc. perform far below DKT, and SAKT far above AKT on AAAI2023? How were baselines tuned?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.