Research map / Large language models / Language, safety and interpretability
A Stitch in Time Saves Nine: Proactive Self-Refinement for Language Models
Jinyi Han, Xinyi Wang, Haiquan Zhao, tingyun li, Zishang Jiang, Sihang Jiang, Jiaqing Liang, Xin Alex Lin, Weikang Zhou, Zeye Sun, Fei Yu, Yanghua Xiao
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 |
|---|---|---|---|
| add RL-trained proactive self-refinement (PASR) over vanilla decoding | Qwen2.5-7B · 10-task average | average score: 55.9 -> 61.7 | Improved |
| add RL-trained proactive self-refinement (PASR) over vanilla decoding | Qwen3-8B · 10-task average | average score: 60.9 -> 69.1 | Improved |
| add reinforcement learning rather than prompting for in-generation self-refinement (ablation) | Qwen2.5-7B · 10-task average | average score: 40.2 -> 61.7 | Improved |
| add reinforcement learning rather than prompting for in-generation self-refinement (ablation) | Qwen3-8B · 10-task average | average score: 51.4 -> 69.1 | Improved |
| add reinforcement learning rather than instruction fine-tuning (ablation) | Qwen2.5-7B · 10-task average | average score: 57.0 -> 61.7 | Improved |
| add reinforcement learning rather than instruction fine-tuning (ablation) | Qwen3-8B · 10-task average | average score: 52.6 -> 69.1 | Improved |
| add multi-answer comparison reward (ablation) | PASR (Qwen2.5-7B) · 10-task average | average score: 57.2 -> 61.7 | Improved |
| add comparison-based refinement reward (ablation) | PASR (Qwen2.5-7B) · 10-task average | average score: 56.2 -> 61.7 | Improved |
| add comparison-based refinement reward (ablation) | PASR (Qwen2.5-7B) · Wino | accuracy: 57.0 -> 65.3 | Hurt |
| add multi-answer comparison reward (ablation) | PASR (Qwen2.5-7B) · GSM8K | accuracy: 75.7 -> 88.8 | Improved |
| add comparison-based reward (ablation) | PASR (Qwen2.5-7B) · AIME24 | accuracy: 10.0 -> 10.0 | No clear effect |
| add oracle feedback to Self-Refine | Qwen2.5-7B · 10-task average | average score: 57.5 -> 62.3 | Improved |
Closest papers by meaning
- Think Thrice Before You Act: Progressive Thought Refinement in Large Language Models (ICLR 2025 · similarity 0.89)
- Mind the Gap: Examining the Self-Improvement Capabilities of Large Language Models (ICLR 2025 · similarity 0.88)
- Beyond Prompt Engineering: A Reinforced Token-Level Input Refinement for Large Language Models (AAAI 2025 · similarity 0.88)
- Self-Improvement in Language Models: The Sharpening Mechanism (ICLR 2025 · similarity 0.88)
- Large Language Models Cannot Self-Correct Reasoning Yet (ICLR 2024 · similarity 0.88)
- RefactorBench: Evaluating Stateful Reasoning in Language Agents Through Code (ICLR 2025 · 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.
