Research map / Large language models / Language, safety and interpretability
Supervising Model Attention with Human Explanations for Robust Natural Language Inference
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Supervise 3 heads of an existing self-attention layer using human explanation words | BERT NLI model · SNLI-test | accuracy: +0.40 | Helped |
| Supervise 3 heads of an existing self-attention layer using human explanation words | BERT NLI model · SNLI-hard | accuracy: +0.79 | Helped |
| Supervise 3 heads of an existing self-attention layer using human explanation words | BERT NLI model · MNLI matched and mismatched | accuracy: +0.91; +0.84 | Helped |
| Supervise 3 heads of an existing self-attention layer using human explanation words | BERT NLI model · HANS | accuracy: +1.59 | Helped |
| Supervise 3 heads of an existing self-attention layer using human explanation words | BERT NLI model · ANLI | accuracy: -0.40 | No clear effect |
| Supervise an additional attention layer | BERT NLI model · SNLI-test | accuracy: +0.32 | Helped |
| Shuffle the desired attention distribution instead of using explanation words (ablation) | BERT NLI model · SNLI-test | accuracy: 89.50% | Hurt |
| Use free-text explanations as supervision (ablation) | BERT NLI model · SNLI dev | accuracy: +0.46 | Helped |
| Use highlighted words as supervision (ablation) | BERT NLI model · SNLI dev | accuracy: +0.52 | Helped |
| Combine free-text explanations and highlighted words (ablation) | BERT NLI model · SNLI dev | accuracy: +0.57 | Helped |
| Supervise only the selected top 3 attention heads rather than all heads (ablation) | BERT NLI model · SNLI | accuracy | Helped |
| Leave the additional attention layer unsupervised (ablation) | BERT NLI model with an added attention layer · NLI | accuracy | Hurt |
Closest papers by meaning
- e-SNLI: Natural Language Inference with Natural Language Explanations (NeurIPS 2018 · similarity 0.89)
- NLX-GPT: A Model for Natural Language Explanations in Vision and Vision-Language Tasks (CVPR 2022 · similarity 0.89)
- Unveiling and Harnessing Hidden Attention Sinks: Enhancing Large Language Models without Training through Attention Calibration (ICML 2024 · similarity 0.87)
- Improving Natural Language Understanding for LLMs via Large-Scale Instruction Synthesis (AAAI 2025 · similarity 0.87)
- Supervised Knowledge Makes Large Language Models Better In-context Learners (ICLR 2024 · similarity 0.87)
- Maximizing Signal in Human-Model Preference Alignment (AAAI 2025 · similarity 0.86)
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.