Research map / Reinforcement learning / RL from feedback
Reinforcement Learning from Dynamic Critic Feedback for Free-Form Generations
Mian Wu, Gavin Zhang, Sewon Min, Sergey Levine, Aviral Kumar
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 reliable external validation (vs random labels) (ablation) | RLAC · Biography 8-sentence, Qwen3-4B | FactScore: 0.607 -> 0.817 | Improved |
| add adversarial critic training (vs static critic) (ablation) | RLAC · Biography 8-sentence, Qwen3-4B | # Corr: 17.77 -> 21.58 | Improved |
| add adversarial critic training (vs static critic) (ablation) | RLAC · Biography 8-sentence, Qwen3-4B | FactScore: 0.825 -> 0.817 | No clear effect |
| train critic adversarially instead of static (ablation) | RLAC · HumanEval Pass@1, Qwen2.5-Coder-7B-Instruct | Pass@1: 92.1 -> 93.3 | Improved |
| train critic adversarially instead of static (ablation) | RLAC · Code critic detection rate | detection precision: 42.3% to 33.9% | Improved |
| replace exhaustive verification with adversarial critic | FactTune-FS · Biography 8-sentence, Qwen3-8B | verification calls: 438,949 -> 76,800 | Improved |
| replace exhaustive verification with adversarial critic | FactTune-FS · Biography 8-sentence, Qwen3-8B | FactScore: 0.867 -> 0.889 | Improved |
| replace reward model with adversarial critic | ArmoRM · Biography 8-sentence, Qwen3-8B | FactScore: 0.723 -> 0.889 | Improved |
| replace enumerative test rewards with adversarial critic | AceCoder-Rule · Code average, Qwen2.5-Coder-7B-Base | average score: 52.3 -> 53.2 | Improved |
| replace enumerative test rewards with adversarial critic | AceCoder-Rule · Code validator calls | test cases executed: 7.86 million -> 192 thousand | Improved |
| use AceCoder-RM reward model | Qwen2.5-Coder-7B-Instruct · HumanEval Base | Pass@1: 91.5 -> 89.0 | Hurt |
| use GPT-4o reference solutions in validator instead of Qwen | RLAC · Code average, Qwen2.5-Coder-7B-Instruct | average score: 56.6 -> 56.5 | No clear effect |
Closest papers by meaning
- Teaching Language Models to Critique via Reinforcement Learning (ICML 2025 · similarity 0.91)
- Reward Hacking in Rubric-Based Reinforcement Learning (NeurIPS 2026 · similarity 0.88)
- RLEF: Grounding Code LLMs in Execution Feedback with Reinforcement Learning (ICML 2025 · similarity 0.88)
- From Verifiable Dot to Reward Chain: Harnessing Verifiable Reference-based Rewards for Reinforcement Learning of Open-ended Generation (ICLR 2026 · similarity 0.88)
- Learning to Generate Unit Test via Adversarial Reinforcement Learning (ICLR 2026 · similarity 0.87)
- Critique-Coder: Enhancing Coder Models by Critique Reinforcement Learning (ICLR 2026 · 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.
