Research map / Large language models / Language, safety and interpretability
Post Hoc Explanations of Language Models Can Improve Language Models
Satyapriya Krishna, Jiaqi Ma, Dylan Z Slack, Asma Ghandeharioun, Sameer Singh, Himabindu Lakkaraju
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 |
|---|---|---|---|
| replace human CoT rationales with post hoc explanation keyword rationales | GPT-3.5 · 7-task average | accuracy: 62.9 -> 72.7 | Improved |
| replace human CoT rationales with post hoc explanation rationales | GPT-3 · 7-task average | accuracy: 58.0 -> 67.2 | Improved |
| replace CoT with AMPLIFY on a math multi-step task | GPT-3.5 · GSM8k | accuracy: 43.5 -> 27.4 | Hurt |
| replace CoT with AMPLIFY on word sorting | GPT-3.5 · Word Sorting | accuracy: 43.1 -> 43.6 | No clear effect |
| fine-tune GPT-2 proxy for 200 epochs instead of none (ablation) | AMPLIFY · 7-task average (GPT-3.5) | accuracy: 65.4 -> 72.7 | Improved |
| use highest-MCS samples instead of random samples (ablation) | AMPLIFY · 7-task average (GPT-3.5) | accuracy: 62.0 -> 72.7 | Improved |
| use highest-MCS instead of lowest-MCS samples (ablation) | AMPLIFY · 7-task average (GPT-3.5) | accuracy: 69.8 -> 72.7 | Improved |
| use highest-MCS instead of faithful-explanation sample selection (ablation) | AMPLIFY · 7-task average (GPT-3.5) | accuracy: 69.7 -> 72.7 | Improved |
| Gradient x Input instead of vanilla gradient explanations (ablation) | AMPLIFY · 7-task average (GPT-3.5) | accuracy: 68.6 -> 72.5 | Improved |
| use contrastive variants of gradient explanations (ablation) | AMPLIFY · 7-task average (GPT-3.5) | accuracy: 70.1 vs 68.6/72.5 | No clear effect |
| increase keywords k and shots s from (2,5) to (5,10) (ablation) | AMPLIFY · 7-task average (GPT-3.5) | accuracy: 66.0 -> 72.6 | Improved |
| tailor rationale template to the question on Causal Judgment (ablation) | AMPLIFY · Causal Judgment | accuracy: 60.5 -> 63.1 | Improved |
Closest papers by meaning
- Supervised Knowledge Makes Large Language Models Better In-context Learners (ICLR 2024 · similarity 0.88)
- Hypothesis Search: Inductive Reasoning with Language Models (ICLR 2024 · similarity 0.87)
- Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.87)
- SCOUT: Teaching Pre-trained Language Models to Enhance Reasoning via Flow Chain-of-Thought (NeurIPS 2025 · similarity 0.87)
- Automated Assessment of Fidelity and Interpretability: An Evaluation Framework for Large Language Models’ Explanations (Student Abstract) (AAAI 2024 · similarity 0.87)
- In-Context Unlearning: Language Models as Few-Shot Unlearners (ICML 2024 · similarity 0.87)
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.
