Research map / Large language models / Language, safety and interpretability
A Positive Case for Faithfulness: Explanations Help Predict Model Behavior
Harry Mayne, Justin Singh Kang, Dewi Sid William Gould, Kannan Ramchandran, Adam Mahdi, Noah Y. Siegel
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 |
|---|---|---|---|
| give predictors the reference model's self-explanation | 18 frontier LLMs as reference models · 7,000 counterfactual pairs | NSG: NSG 11.0-36.5%; absolute gain 3.8-10.8% | Improved |
| use self-explanations instead of explanations from external models (ablation) | NSG evaluation · top 3 models per family, cross-family | NSG: +4.3pp GPT-5, +3.0pp Qwen 3, +1.7pp Gemma 3 | Improved |
| scale up reference model size within Qwen 3 family (ablation) | Qwen 3 reference models · NSG across datasets | NSG: 11.03 (0.6B) -> 35.61 (32B) | Improved |
| scale up model size among proprietary models (ablation) | GPT-5, Claude 4.5, Gemini 3 · NSG across datasets | NSG | No clear effect |
| raise reasoning strength (none to high) (ablation) | Claude Sonnet 4.5 and GPT-5.2 · NSG across datasets | NSG | No clear effect |
| use user-facing explanation instead of raw chain-of-thought as the explanation (ablation) | Qwen 3 models · NSG with 3 predictors | NSG | Improved |
| vary reference prompt role framing (generic, professional, specialist) (ablation) | Qwen3-32B · Heart Disease, Income, Breast Cancer | NSG: 27.4, 29.9, 26.3 | No clear effect |
| provide explanation to predictor | predictor models · counterfactuals with unchanged reference answer | false change rate: 35.4 -> 18.2 (gpt-oss-20b) | Improved |
| use stronger predictors (ablation) | NSG evaluation · predictor ensemble | absolute simulatability gain: +11.9 (Qwen3-32B) vs +4.8 (gemini-3-flash) | Hurt |
| longer reference reasoning trace (ablation) | proprietary reference models · gpt-oss-20b predictor | correlation with explanation helpfulness: -0.018 to -0.061 | No clear effect |
| move to Moral Machines instead of other datasets (ablation) | NSG evaluation · Moral Machines vs Pima Diabetes | NSG: 6.0% (Moral Machines) vs 42.8% (Pima Diabetes) | Hurt |
| remove factual (original answer) from baseline: no-information baseline (ablation) | Qwen 3 family · 3 predictors | simulatability gain | Improved |
Closest papers by meaning
- Dissenting Explanations: Leveraging Disagreement to Reduce Model Overreliance (AAAI 2024 · similarity 0.88)
- Walk the Talk? Measuring the Faithfulness of Large Language Model Explanations (ICLR 2025 · similarity 0.87)
- DISCRET: Synthesizing Faithful Explanations For Treatment Effect Estimation (ICML 2024 · similarity 0.87)
- Do Models Explain Themselves? Counterfactual Simulatability of Natural Language Explanations (ICML 2024 · similarity 0.86)
- Tell me about yourself: LLMs are aware of their learned behaviors (ICLR 2025 · similarity 0.86)
- Evaluate with the Inverse: Efficient Approximation of Latent Explanation Quality Distribution (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.
