Research map / Architectures and efficiency / RNNs and normalization
Extended LSTMs for Knowledge Tracing: Peeking Inside the Black Box (Student Abstract)
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes sLSTM-KT, a knowledge tracing model built on the extended LSTM (sLSTM) from xLSTM, and explains its predictions with gradient sensitivity, gradient*input, and Deep SHAP on ASSISTment2009.
Problem
Knowledge tracing with deep models like DKT is criticized for LSTM limitations and a black-box nature, and it is unclear whether extended LSTMs work for KT or how their predictions can be interpreted.
Approach
Adapts the exponential-gated sLSTM (from xLSTM) to knowledge tracing and applies three input-attribution XAI methods (gradient sensitivity, gradient*input, Deep SHAP) to quantify how each past exercise-answer record contributes to predictions.
Key results
- sLSTM-KT reached AUC 0.7523 and accuracy 0.7202 on ASSIST2009 versus DKT's AUC 0.7481 and accuracy 0.7192 (5-fold CV).
- Deleting records ranked by Gradient Sensitivity, Gradient*Input, or Deep SHAP contribution reduces accuracy much faster than random deletion on initially correct sequences.
- Deep SHAP analysis suggests wrong answers on the same skill and position contribute more than correct ones, target-skill pairs outweigh non-target ones, and nearby pairs outweigh distant ones.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace DKT with sLSTM-KT | Knowledge tracing on ASSISTment2009 · ASSISTment2009, 5-fold cross-validation | AUC: 0.7481 -> 0.7523 | Helped |
| Replace DKT with sLSTM-KT | Knowledge tracing on ASSISTment2009 · ASSISTment2009, 5-fold cross-validation | accuracy: 0.7192 -> 0.7202 | Helped |
| Delete exercise-answer pairs in order of their contribution values instead of randomly | sLSTM-KT explanations · Correctly predicted sequences, with pairs deleted sequentially | accuracy | Hurt |
Surprising findings
- Gains over DKT are marginal (AUC 0.7523 vs 0.7481, accuracy essentially tied), which the authors describe as only slight outperformance.
Limitations
- Improvement over DKT is slight and evaluated on a single dataset.
- Explanations are validated only via a deletion test on correctly predicted sequences.
- Only the sLSTM variant is studied; mLSTM and xLSTM are not covered.
Experimental setup
Datasets: ASSISTment2009
Benchmarks: ASSISTment2009
Metrics: AUC accuracy accuracy change under contribution-based deletion
Baselines: DKT
Builds on: xLSTM DKT Deep SHAP Gradient*Input
Closest papers by meaning
- Concept-Guided Interpretability via Neural Chunking (NeurIPS 2025 · similarity 0.84)
- Rethinking and Improving Student Learning and Forgetting Processes for Attention based Knowledge Tracing Models (AAAI 2025 · similarity 0.84)
- Uncertainty-aware Knowledge Tracing (AAAI 2025 · similarity 0.84)
- Training Language Models to Explain Their Own Computations (NeurIPS 2026 · similarity 0.84)
- Cognitive Fluctuations Enhanced Attention Network for Knowledge Tracing (AAAI 2025 · similarity 0.84)
- KaVa: Latent Reasoning via Compressed KV-Cache Distillation (ICLR 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.