Research map / Large language models / Language, safety and interpretability
Learning to Scaffold: Optimizing Model Explanations for Teaching
Patrick Fernandes, Marcos Vinicius Treviso, Danish Pruthi, Andre Martins, Graham Neubig
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| learn head coefficients with SMaT instead of static mean over all heads | attention explainer · IMDB, 500 samples | simulability accuracy (%): 83.00 -> 91.48 | Improved |
| learn head coefficients with SMaT instead of static all-layer attention | attention explainer · IMDB, 2,000 samples | simulability accuracy (%): 90.08 -> 92.84 | Improved |
| learn head coefficients with SMaT instead of static attention | attention explainer · CIFAR-100, 9,000 samples | simulability accuracy (%): 85.92 -> 85.96 | Improved |
| use SMaT-learned attention instead of Gradient L2 at 8,400 samples (median) | attention explainer · MLQE-PE | Pearson correlation: .8638 vs .8561 | Mixed |
| add gradient x input or integrated gradients explanations to student training | student · IMDB, 1,000 samples | simulability accuracy (%): 83.44 -> 81.15 | Hurt |
| use only last-layer heads instead of all heads (static attention) (ablation) | attention explainer · IMDB, 500 samples | simulability accuracy (%): 83.00 -> 80.91 | Hurt |
| use Softmax instead of Sparsemax for head coefficient normalization (ablation) | SMaT · MLQE-PE 4,200 | Pearson correlation: .8630 -> .7439 | Hurt |
| use 1.5-Entmax instead of Sparsemax (ablation) | SMaT · MLQE-PE 4,200 | Pearson correlation: .8630 -> .7163 | Hurt |
| remove head coefficient normalization (ablation) | SMaT · MLQE-PE 4,200 | Pearson correlation: .8630 -> .8002 | Hurt |
| train explainer with SMaT instead of static all-layer attention | attention explainer · MovieReviews plausibility | AUC: 0.68 -> 0.73 | Improved |
| train explainer with SMaT instead of static all-layer attention | attention explainer · CIFAR-100 user study | TrueSkill: 0.7±.67 -> 4.3±.70 | Improved |
Closest papers by meaning
- Can Language Models Teach? Teacher Explanations Improve Student Performance via Personalization (NeurIPS 2023 · similarity 0.89)
- Contextual Explanation Networks (JMLR 2020 · similarity 0.87)
- Few-Shot Knowledge Distillation of LLMs With Counterfactual Explanations (NeurIPS 2025 · similarity 0.86)
- ReX: A Framework for Incorporating Temporal Information in Model-Agnostic Local Explanation Techniques (AAAI 2025 · similarity 0.85)
- iLLuMinaTE: An LLM-XAI Framework Leveraging Social Science Explanation Theories Towards Actionable Student Performance Feedback (AAAI 2025 · similarity 0.85)
- How to Probe: Simple Yet Effective Techniques for Improving Post-hoc Explanations (ICLR 2025 · similarity 0.85)
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.
