Research map / Reinforcement learning / RL from feedback
Random Policy Valuation is Enough for LLM Reasoning with Verifiable Rewards
Haoran He, Yuxiao Ye, Qingpeng Cai, Chen Hu, Binxing Jiao, Daxin Jiang, Ling Pan
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 |
|---|---|---|---|
| use ROVER instead of best baseline (DAPO) | Qwen3-8B-Base · AIME 2024 | pass@1: 20.8 -> 30.6 | Improved |
| use ROVER instead of REINFORCE++ | Qwen3-8B-Base · HMMT 2025 | pass@1: 7.1 -> 14.6 | Improved |
| use ROVER instead of GRPO | Qwen3-4B-Base · average pass@1 | pass@1: 35.3 -> 36.5 | Improved |
| use ROVER instead of baselines at large k | Qwen3-8B-Base · AIME24/AIME25/HMMT25 pass@256 | pass@256: +16.8 over best baseline | Improved |
| use ROVER instead of GRPO | Qwen3-8B-Base · AIME24 distinct strategies | diversity: +6.8% | Improved |
| use ROVER instead of DAPO/GRPO baselines on out-of-distribution task | Qwen3-8B-Base · GPQA-diamond | pass@1: 46.6 -> 50.2 | Improved |
| remove the uniform-policy successor Q term (beta=0) (ablation) | ROVER · AIME24/25, HMMT25 (Qwen3-4B-Base) | pass@k | Improved |
| scale successor Q term too large (beta=5.0) (ablation) | ROVER · AIME24/25, HMMT25 (Qwen3-4B-Base) | pass@k | Hurt |
| vary beta between 0.2 and 1.0 (ablation) | ROVER · AIME24/25, HMMT25 (Qwen3-4B-Base) | pass@k | No clear effect |
| use softmax temperature rho=1 vs much lower or higher (ablation) | ROVER · Countdown (Qwen2.5-3B) | test score | Improved |
| use ROVER instead of GRPO with KL 0.01 on Countdown | Qwen2.5-3B · Countdown | distinct solutions: 3 -> 17 | Improved |
Closest papers by meaning
- Reinforcement Learning with Verifiable Rewards Implicitly Incentivizes Correct Reasoning in Base LLMs (ICLR 2026 · similarity 0.93)
- Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model? (NeurIPS 2025 · similarity 0.92)
- Reinforcement Learning for Reasoning in Large Language Models with One Training Example (NeurIPS 2025 · similarity 0.92)
- On the Direction of RLVR Updates for LLM Reasoning: Identification and Exploitation (ICLR 2026 · similarity 0.92)
- Tricks or Traps? A Deep Dive into RL for LLM Reasoning (ICLR 2026 · similarity 0.91)
- What are Key Factors for Updates in RL for LLM Reasoning? (NeurIPS 2026 · similarity 0.91)
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.
