Research map / Large language models / Alignment and preferences
Second Thoughts are Best: Learning to Re-Align With Human Values from Text Edits
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 |
|---|---|---|---|
| add RL refinement with value modeling to AEM (ablation) | Second Thoughts · Moral Stories | alignment: 3.80 -> 4.85 | Helped |
| add RL refinement with value modeling to AEM (ablation) | Second Thoughts · ETHICS-Deontology | alignment: 3.86 -> 5.57 | Helped |
| add RL refinement with value modeling to AEM (ablation) | Second Thoughts · MIC | coherence: 5.47 -> 5.88 | Mixed |
| add adversarial imitation learning to AEM (ablation) | Second Thoughts · Moral Stories | alignment: 3.80 -> 4.55 | Helped |
| use value modeling instead of adversarial imitation learning (ablation) | Second Thoughts · Moral Stories | alignment: 4.55 -> 4.85 | Helped |
| fine-tune on chain-of-edits instead of vanilla text-to-text (ablation) | Second Thoughts · TruthfulQA and HHH transfer | alignment | Helped |
| increase chain-of-edits augmentation factor (ablation) | Second Thoughts · Moral Stories | ROUGE-L and perplexity | Mixed |
| increase balancing factor alpha in AIL (ablation) | Second Thoughts · Moral Stories | ROUGE-L and perplexity | Mixed |
| use human-guided edits on chain-of-edits instead of re-prompting | Second Thoughts · Moral Stories | alignment: 4.56 -> 5.28 | Helped |
| use small GPT-2 with Second Thoughts instead of GPT-3 175B | GPT-2 medium vs GPT-3 · Moral Stories | alignment: 3.28 -> 4.85 | Helped |
| use DExperts token-level control instead of Second Thoughts | DExperts · ETHICS-Deontology | coherence: 3.91 vs 6.03 | Hurt |
Closest papers by meaning
- Exploring Intrinsic Alignments Within Text Corpus (AAAI 2025 · similarity 0.87)
- SEAL: Systematic Error Analysis for Value ALignment (AAAI 2025 · similarity 0.86)
- RLTHF: Targeted Human Feedback for LLM Alignment (ICML 2025 · similarity 0.85)
- ABC-Align: Prediction-Powered Alignment with Adaptive Bias Control (NeurIPS 2026 · similarity 0.85)
- Beyond Imitation: Leveraging Fine-grained Quality Signals for Alignment (ICLR 2024 · similarity 0.85)
- DENEVIL: TOWARDS DECIPHERING AND NAVIGATING THE ETHICAL VALUES OF LARGE LANGUAGE MODELS VIA INSTRUCTION LEARNING (ICLR 2024 · 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.