Research map / Reinforcement learning / RL from feedback
VeriRole: Verifiable Role-Awareness through Hint-Guided Reinforcement Learning
Zongsheng Wang, Kaili Sun, Bowen Wu, qun yu, Ying Li, Xu Chen, Baoxun Wang
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 hint reward (ablation) | Qwen2.5-32B-GRPO · RAIDEN | average score: 0.6598 -> 0.8268 | Improved |
| add hint reward (ablation) | Qwen2.5-32B-GRPO · RAIDEN Topic Advancement | score: 0.2089 -> 0.6865 | Improved |
| add accuracy reward (ablation) | Qwen2.5-32B-GRPO · RAIDEN | average score: 0.8061 -> 0.8268 | Improved |
| train on RAIDEN plus Situation Puzzle instead of Situation Puzzle only (ablation) | Qwen2.5-32B-GRPO · RAIDEN | average score: 0.7111 -> 0.8268 | Improved |
| add Situation Puzzle data to RAIDEN-only training (ablation) | Qwen2.5-32B-GRPO · RAIDEN | average score: 0.8143 -> 0.8268 | Improved |
| use GRPO instead of SFT on replies | Qwen2.5-32B · RAIDEN | average score: 0.6809 -> 0.8268 | Improved |
| use GRPO instead of SFT on replies | Qwen2.5-32B · RAIDEN Topic Advancement | score: 0.2835 -> 0.6865 | Improved |
| apply GRPO to instruct model | Qwen2.5-14B · RAIDEN | average score: 0.6302 -> 0.7725 | Improved |
| apply GRPO to role-play model | Peach-9B-roleplay · RAIDEN | average score: 0.3611 -> 0.6183 | Improved |
| apply GRPO to a model with inherent reasoning | Qwen3-32B · RAIDEN Topic Advancement | score | Hurt |
| enable thinking mode | Qwen3-32B · RAIDEN | average score: 0.6989 -> 0.6796 | No clear effect |
| GRPO training before SFT on psychotherapist data | Qwen2.5-32B · RAIDEN after post-training SFT | average score: 0.6366 -> 0.8006 | Improved |
Closest papers by meaning
- Thinking in Character: Advancing Role-Playing Agents with Role-Aware Reasoning (NeurIPS 2025 · similarity 0.86)
- Consistently Simulating Human Personas with Multi-Turn Reinforcement Learning (NeurIPS 2025 · similarity 0.84)
- Generalizing Verifiable Instruction Following (NeurIPS 2025 · similarity 0.84)
- RLVER: Reinforcement Learning with Verifiable Emotion Rewards for Empathetic Agents (ICLR 2026 · similarity 0.84)
- MMRole: A Comprehensive Framework for Developing and Evaluating Multimodal Role-Playing Agents (ICLR 2025 · similarity 0.84)
- DMT-RoleBench: A Dynamic Multi-Turn Dialogue Based Benchmark for Role-Playing Evaluation of Large Language Model and Agent (AAAI 2025 · 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.
