Research map / Reinforcement learning / RL from feedback
CRPO: Character-centric Group Relative Policy Optimization for Role-aware Reasoning in Role-playing Agents
Yihong Tang, Kehai Chen, Liang Yue, Benyou Wang, Min Zhang
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 dual-stream advantage (ablation) | CRPO (Qwen3-8B) · CharacterBench | average score: 3.811 -> 4.043 | Improved |
| add entropy-aware adaptive exploitation (ablation) | CRPO (Qwen3-8B) · CharacterBench | average score: 3.962 -> 4.043 | Improved |
| add contrastive anchor sampling (ablation) | CRPO (Qwen3-8B) · CharacterBench | average score: 3.942 -> 4.043 | Improved |
| add contrastive anchor sampling (ablation) | CRPO (Qwen3-8B) · CharacterBench Engagement | score: 3.025 -> 3.130 | Improved |
| add dual-stream advantage (ablation) | CRPO (Qwen3-8B) · CharacterBench Attribute Consistency (Bot) | score: 4.294 -> 4.994 | Improved |
| add entropy-aware adaptive exploitation (ablation) | CRPO (Qwen3-8B) · CharacterBench Fact Accuracy | score: 2.553 -> 2.581 | Improved |
| use CRPO instead of GRPO | Qwen3-8B · CharacterBench | average score: 3.829 -> 4.043 | Improved |
| use CRPO instead of GRPO | Llama-3.2-3B-Instruct · CharacterBench | average score: 3.781 -> 3.832 | Improved |
| add SFT on role-play data | Qwen3-8B · CharacterBench | average score: 3.701 -> 3.673 | No clear effect |
| train with CRPO | Qwen3-8B · SocialBench (unseen) | average score: 0.769 -> 0.802 | Improved |
| train with CRPO | Qwen3-8B · SocialBench SU | score: 0.230 -> 0.350 | Improved |
| increase number of contrastive anchors beyond one (ablation) | CRPO · CharacterBench | average score | No clear effect |
Later papers that compared with this method
1 later accepted paper reports CRPO as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Breaking Safety Paradox with Feasible Dual Policy Iteration (ICLR 2026 · better on 2 of 2 shared results)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- +GDPO — GDPO: Group reward-Decoupled Normalization Policy Optimization for Multi-reward RL Optimization (ICML 2026) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Stronger-MAS: Multi-Agent Reinforcement Learning for Collaborative LLMs (ICLR 2026 · similarity 0.87)
- Thinking in Character: Advancing Role-Playing Agents with Role-Aware Reasoning (NeurIPS 2025 · similarity 0.87)
- All Roads Lead to Rome: Incentivizing Divergent Thinking in Vision-Language Models (CVPR 2026 · similarity 0.85)
- Group-in-Group Policy Optimization for LLM Agent Training (NeurIPS 2025 · similarity 0.85)
- TIC-GRPO: Provable and Efficient Optimization for Reinforcement Learning from Human Feedback (NeurIPS 2026 · similarity 0.85)
- Agentic Reinforced Policy Optimization (ICLR 2026 · similarity 0.84)
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.
