Research map / Reinforcement learning / RL from feedback
SERL: Self-Examining Reinforcement Learning on Open-Domain
Weixuan Ou, Yanzhao Zheng, Shuoshuo Sun, Wei Zhang, Baohua Dong, Hangcheng Zhu, Ruohui Huang, Gang Yu, Pengwei Yan, Yifan Qiao
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 reward for Judge (self-consistency) (ablation) | SERL (Qwen3-8B) · Summarization | win rate: -9.34% when removed | Improved |
| add reward for Judge (self-consistency) (ablation) | SERL (Qwen3-8B) · Open Writing | win rate: -10.66% when removed | Improved |
| add reward for Judge (self-consistency) (ablation) | SERL (Qwen3-8B) · AlpacaEval 2.0 | LC win rate: 59.90% -> 54.50% when removed | Improved |
| add reward for Actor (Copeland win-rate ranking) (ablation) | SERL (Qwen3-8B) · Summarization | win rate: -35.34% when removed | Improved |
| add reward for Actor (Copeland win-rate ranking) (ablation) | SERL (Qwen3-8B) · Open Writing | win rate: -20.34% when removed | Improved |
| add reward for Actor (Copeland win-rate ranking) (ablation) | SERL (Qwen3-8B) · AlpacaEval 2.0 | LC win rate: 59.90% -> 51.47% when removed | Improved |
| add position bias mitigation (PBMM) (ablation) | SERL (Qwen3-8B) · Summarization | win rate: -15.00% when removed | Improved |
| add position bias mitigation (PBMM) (ablation) | SERL (Qwen3-8B) · Open Writing | win rate: -2.66% when removed | Improved |
| add position bias mitigation (PBMM) (ablation) | SERL (Qwen3-8B) · AlpacaEval 2.0 | LC win rate: 59.90% -> 54.01% when removed | Improved |
| add length control module (LCM) (ablation) | SERL (Qwen3-8B) · Summarization training | average output length: surpassing 1,000 tokens without LCM vs around 600 tokens with LCM | Improved |
| train Qwen3-8B with SERL | Qwen3-8B · AlpacaEval 2.0 | LC win rate: 52.37% -> 59.90% | Improved |
| train Qwen3-8B with SERL | Qwen3-8B · AlpacaEval 2.0 | win rate: 55.07% -> 69.88% | Improved |
Closest papers by meaning
- SeRL: Self-play Reinforcement Learning for Large Language Models with Limited Data (NeurIPS 2025 · similarity 0.94)
- LaSeR: Reinforcement Learning with Last-Token Self-Rewarding (ICLR 2026 · similarity 0.90)
- Enhancing the Outcome Reward-based RL Training of MLLMs with Self-Consistency Sampling (NeurIPS 2025 · similarity 0.89)
- Self-Aligned Reward: Towards Effective and Efficient Reasoners (ICLR 2026 · similarity 0.89)
- SELF-EVOLVED REWARD LEARNING FOR LLMS (ICLR 2025 · similarity 0.89)
- Revisiting Reinforcement Learning for LLM Reasoning from A Cross-Domain Perspective (NeurIPS 2025 · similarity 0.88)
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.
