Research map / Reinforcement learning / RL from feedback
HiPO: Self-Hint Policy Optimization for RLVR
Deng Qiyuan, Kehai Chen, Min Zhang, Zhongwen Xu
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 on-policy self-hint rollouts | GRPO · Average (5 benchmarks) | avg@32: 49.8 -> 54.8 | Improved |
| add on-policy self-hint rollouts | GRPO · CMIMC 2025 | avg@32: 43.5 -> 53.8 | Improved |
| add on-policy self-hint rollouts | GRPO · BRUMO 2025 | avg@32: 41.7 -> 46.6 | Improved |
| add on-policy self-hint rollouts | GRPO · AIME 2024 | avg@32: 72.1 -> 76.7 | Improved |
| HiPO versus DAPO dynamic sampling | DAPO · Average (5 benchmarks) | avg@32: 53.7 -> 54.8 | Improved |
| add on-policy self-hint rollouts | GRPO · Average (5 benchmarks) | maj@32: 57.0 -> 61.7 | Improved |
| use fixed hint ratio 0.05 instead of random ratio in [0.05,0.45] (ablation) | HiPO · AIME 2024 / 2025 training curves | Avg@16 | Hurt |
| use fixed hint ratio 0.80 instead of random ratio (ablation) | HiPO · AIME 2024 / 2025 training curves | Avg@16 | Hurt |
| use static off-policy hints from the base model (first 20% tokens) instead of on-policy hints (ablation) | HiPO · AIME 2024 / 2025 training curves | Avg@16 | Hurt |
| add hint-based group replacement | GRPO · Learnable group ratio | learnable groups: 17.5% -> 75.8% | Improved |
| add on-policy self-hint rollouts | GRPO · HumanEval (OOD) | Pass@1: 71.3 -> 71.3 | No clear effect |
| add on-policy self-hint rollouts | Qwen3-8B · HumanEval (OOD) | Pass@1: 70.7 -> 71.3 | Improved |
Closest papers by meaning
- TraPO: A Semi-Supervised Reinforcement Learning Framework for Boosting LLM Reasoning (ICLR 2026 · similarity 0.91)
- Multi-Step Likelihood-Ratio Correction for Reinforcement Learning with Verifiable Rewards (NeurIPS 2026 · similarity 0.90)
- DeepSearch: Overcome the Bottleneck of Reinforcement Learning with Verifiable Rewards via Monte Carlo Tree Search (ICLR 2026 · similarity 0.90)
- Reinforcement Learning for Reasoning in Large Language Models with One Training Example (NeurIPS 2025 · similarity 0.90)
- Depth-Breadth Synergy in RLVR: Unlocking LLM Reasoning Gains with Adaptive Exploration (ICML 2026 · similarity 0.90)
- The Reasoning Boundary Paradox: How Reinforcement Learning Constrains Language Models (NeurIPS 2026 · 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.
