Research map / Reinforcement learning / RL from feedback
RiskPO: Risk-based Policy Optimization with Verifiable Reward for LLM Post-Training
Tao Ren, Jinyang Jiang, Hui Yang, Wan Tian, Minhao Zou, Guanghao Li, Zishi Zhang, Qinghao Wang, Shentao Qin, Yanjun Zhao, Rui Tao, Hui Shao and 1 more
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 |
|---|---|---|---|
| replace mean-based GRPO objective with MVaR risk objective | GRPO · Hard-level math avg (6 benchmarks) | Pass@1 avg: 40.41 -> 46.65 | Improved |
| replace mean objective with MVaR vs DAPO | DAPO · Hard-level math avg | Pass@1 avg: 43.87 -> 46.65 | Improved |
| add bundling of 5 questions (vs no bundling B=1) (ablation) | RiskPO · Easy-level math avg (MATH, GSM8K) | Pass@1 avg: 65.80 -> 68.25 | Improved |
| increase bundle size from 5 to 10 (ablation) | RiskPO · Easy-level math avg (MATH, GSM8K) | Pass@1 avg: 68.25 -> 66.65 | Hurt |
| use bundle size 2 instead of 5 (ablation) | RiskPO · Easy-level math avg (MATH, GSM8K) | Pass@1 avg: 68.25 -> 67.20 | Hurt |
| use risk-seeking objective instead of risk-averse (ablation) | RiskPO · MATH | Pass@1 / entropy: 52% -> 54% vs 52% -> 56% | Hurt |
| lower alpha from 0.2 to 0.1 (ablation) | RiskPO · Easy-level math avg (MATH, GSM8K) | Pass@1 avg: 68.25 -> 66.90 | Hurt |
| raise beta from 0.8 to 0.9 (ablation) | RiskPO · Easy-level math avg (MATH, GSM8K) | Pass@1 avg: 68.25 -> 66.95 | Hurt |
| set omega to 0 (no extra lower-tail weight) (ablation) | RiskPO · Easy-level math avg (MATH, GSM8K) | Pass@1 avg: 68.25 -> 67.20 | Hurt |
| set omega to 1.0 (ablation) | RiskPO · Easy-level math avg (MATH, GSM8K) | Pass@1 avg: 68.25 -> 67.85 | Hurt |
| use RiskPO instead of GRPO | GRPO · Pass@k on AMC and MATH500 | Pass@8/16 | Improved |
| use RiskPO instead of GRPO | GRPO · DAPOMATH-17K training | policy entropy | Improved |
Closest papers by meaning
- One-Way Policy Optimization for Self-Evolving LLMs (ICML 2026 · similarity 0.91)
- SetPO: Set-Level Policy Optimization for Diversity-Preserving LLM Reasoning (ICML 2026 · similarity 0.91)
- Act Only When It Pays: Efficient Reinforcement Learning for LLM Reasoning via Selective Rollouts (NeurIPS 2025 · similarity 0.90)
- GVPO: Group Variance Policy Optimization for Large Language Model Post-Training (NeurIPS 2025 · similarity 0.90)
- Accelerating RL for LLM Reasoning with Optimal Advantage Regression (NeurIPS 2025 · similarity 0.90)
- DAPO : Improving Multi-Step Reasoning Abilities of Large Language Models with Direct Advantage-Based Policy Optimization (NeurIPS 2025 · similarity 0.90)
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.
