Research map / Reinforcement learning / RL from feedback
Scaf-GRPO: Scaffolded Group Relative Policy Optimization for Enhancing LLM Reasoning
Xichen Zhang, Sitong Wu, Yinghao Zhu, Haoru Tan, Shaozuo Yu, Ziyi He, Jiaya Jia
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 scaffolded hints (Scaf-GRPO) to vanilla GRPO | Qwen2.5-Math-7B · Math benchmarks average | pass@1 (%): 45.2 -> 50.9 | Improved |
| add Scaf-GRPO to vanilla GRPO | Qwen2.5-Math-1.5B · Math benchmarks average | pass@1 (%): 37.6 -> 41.5 | Improved |
| add Scaf-GRPO to vanilla GRPO | Llama-3.2-3B-Instruct · Math benchmarks average | pass@1 (%): 26.1 -> 28.8 | Improved |
| add Scaf-GRPO to vanilla GRPO | DeepSeek-R1-Distill-Qwen-1.5B · Math benchmarks average | pass@1 (%): 50.6 -> 53.6 | Improved |
| use progressive hierarchy K->P->S instead of solution-only hint (ablation) | Scaf-GRPO · Math benchmarks average | pass@1 (%): 48.4 -> 50.9 | Improved |
| add the solution hint to the K->P hierarchy (ablation) | Scaf-GRPO · Math benchmarks average | pass@1 (%): 48.0 -> 50.9 | Improved |
| add the knowledge hint to the P->S hierarchy (ablation) | Scaf-GRPO · Math benchmarks average | pass@1 (%): 49.2 -> 50.9 | Improved |
| add the planning hint to the K->S hierarchy (ablation) | Scaf-GRPO · Math benchmarks average | pass@1 (%): 48.6 -> 50.9 | Improved |
| deliver hints incrementally in chunks instead of full hint at once (ablation) | Scaf-GRPO · Math benchmarks average | pass@1 (%): 47.7 -> 50.9 | Improved |
| add a guidance exemption period (no scaffolding from the very beginning vs exemption) (ablation) | Scaf-GRPO · Math benchmarks average | pass@1 (%) | Improved |
| vary guidance exemption duration from 10% to 40% (ablation) | Scaf-GRPO · Math benchmarks average | pass@1 (%): plateau between 49.5% and 50.9% | No clear effect |
| train on harder filtered data (ablation) | Vanilla GRPO · Math benchmarks average (Qwen2.5-Math-7B) | pass@1 (%): 43.9 -> 44.1 | No clear effect |
Closest papers by meaning
- Group-in-Group Policy Optimization for LLM Agent Training (NeurIPS 2025 · similarity 0.91)
- Advancing LLM Reasoning with Natural Language and Numerical Feedback (ICML 2026 · similarity 0.91)
- DAPO : Improving Multi-Step Reasoning Abilities of Large Language Models with Direct Advantage-Based Policy Optimization (NeurIPS 2025 · similarity 0.91)
- Stronger-MAS: Multi-Agent Reinforcement Learning for Collaborative LLMs (ICLR 2026 · similarity 0.91)
- WS-GRPO: Weakly-Supervised Group-Relative Policy Optimization for Rollout-Efficient Reasoning (ICML 2026 · similarity 0.91)
- Plan Then Action: High-Level Planning Guidance Reinforcement Learning for LLM Reasoning (ICML 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.
