Research map / Large language models / Agents, code and math
Generating Sequences by Learning to Self-Correct
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 a trained corrector | GPT-Neo 1.3B · GSM | Correct: 8.57 -> 21.26 | Helped |
| Correct only incorrect outputs instead of always correcting | Self-Correct on GSM · GSM | Correct: 21.26 -> 24.22 | Helped |
| Add a trained corrector | GPT-2 Large · CommonGen | Constraints: 91.38 -> 94.58 | Helped |
| Add a trained corrector | GPT-2 Medium · E2E | Constraints: 91.50 -> 98.77 | Helped |
| Add a trained corrector | GPT-2 Large · RealToxicityPrompts | Avg. Max. toxicity: 0.527 -> 0.171 | Helped |
| Provide natural-language feedback | Self-Correct on Multitask · Multitask | Correct: 74.31 -> 81.76 | Helped |
| Provide feedback elaborating missing constraint words | Self-Correct on CommonGen · CommonGen | Constraints: 94.58 -> 95.88 | Helped |
| Provide fine-grained toxicity-attribute feedback | Self-Correct on RealToxicityPrompts · RealToxicityPrompts | Avg. Max. toxicity: 0.171 -> 0.156 | Helped |
| Use uniform rather than proportional pair sampling (ablation) | Self-Correct on Multitask · Multitask | Correct: 78.24 -> 77.25 | Hurt |
| Remove value-improving pair selection (ablation) | Self-Correct on Multitask · Multitask | Correct: 78.24 -> 62.35 | Hurt |
| Remove corrector-generated exploration examples (ablation) | Self-Correct on GSM8k · GSM8k | Correct: 23.96 -> 17.60 | Hurt |
| Apply the GPT-Neo 1.3B-trained corrector to GPT-3 Instruct | GPT-3 Instruct · Multitask | Correct: 84.90 -> 90.90 | Helped |
Closest papers by meaning
- Self-Improvement in Language Models: The Sharpening Mechanism (ICLR 2025 · similarity 0.89)
- Small Language Model Can Self-Correct (AAAI 2024 · similarity 0.87)
- Training Language Models to Self-Correct via Reinforcement Learning (ICLR 2025 · similarity 0.87)
- Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo (ICLR 2025 · similarity 0.87)
- Mind the Gap: Examining the Self-Improvement Capabilities of Large Language Models (ICLR 2025 · similarity 0.87)
- Self-Adapting Language Models (NeurIPS 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.