Research map / Reinforcement learning / RL from feedback
Advantage Collapse in Group Relative Policy Optimization: Diagnosis and Mitigation
Xixiang He, Qiyao Sun, Ao Cheng, Xingming Li, Xuanyu Ji, Hailun Lu, Runke Huang, Qingyong Hu
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 AVSPO virtual samples to GRPO | GRPO, Qwen2.5-0.5B to 14B · six math benchmarks | advantage collapse rate: 0.28-0.45 -> 0.11-0.18 | Improved |
| add AVSPO to GRPO | Qwen2.5-Math-7B · MATH-500 | pass@1 accuracy: 65.0 -> 74.1 | Improved |
| add AVSPO to GRPO on out-of-domain MMLU-Pro | Qwen2.5-3B · MMLU-Pro | accuracy: 37.4 -> 38.6 | No clear effect |
| use stratified virtual rewards instead of random uniform (ablation) | AVSPO, Qwen2.5-Math-1.5B · MATH-500 | accuracy: 62.1 -> 67.2 | Improved |
| use stratified virtual rewards instead of fixed partial credit (ablation) | AVSPO, Qwen2.5-Math-1.5B · MATH-500 | accuracy: 63.5 -> 67.2 | Improved |
| augment all-incorrect groups in addition to all-correct groups (ablation) | AVSPO, Qwen2.5-Math-1.5B · MATH-500 | accuracy: 60.8 -> 67.2 | Improved |
| augment all-correct groups in addition to all-incorrect groups (ablation) | AVSPO, Qwen2.5-Math-1.5B · MATH-500 | accuracy: 63.2 -> 67.2 | Improved |
| use adaptive threshold instead of fixed tau=0.5 (ablation) | AVSPO, Qwen2.5-Math-1.5B · MATH-500 | final accuracy: 65.4 -> 67.2 | Improved |
| increase alpha from 0.5 to 1.0 (ablation) | AVSPO · MATH-500 | accuracy: drops 2.1% | Hurt |
| decrease alpha from 0.5 to 0.3 (ablation) | AVSPO · MATH-500 | ACR: residual ACR = 0.24 | Hurt |
| vary initial threshold in [0.3,0.7] and eta in [0.005,0.02] (ablation) | AVSPO · MATH-500 | accuracy: less than 1% accuracy variation | No clear effect |
| augment collapsed groups (AVSPO) instead of filtering them (Filter-Drop) (ablation) | GRPO, Qwen2.5-Math-7B · MATH | accuracy: 67.1 -> 74.1 | Improved |
Closest papers by meaning
- Revisiting Group Relative Policy Optimization: Insights into On-Policy and Off-Policy Training (ICLR 2026 · similarity 0.92)
- The Capability Frontier of GRPO (NeurIPS 2026 · similarity 0.91)
- Learn from the Gap: Differential-Aware Advantage Pruning with Adaptive Rollout Sampling for GRPO (NeurIPS 2026 · similarity 0.90)
- Unveiling Implicit Advantage Symmetry: Why GRPO Struggles with Exploration and Difficulty Adaptation (NeurIPS 2026 · similarity 0.90)
- GVPO: Group Variance Policy Optimization for Large Language Model Post-Training (NeurIPS 2025 · similarity 0.90)
- TIC-GRPO: Provable and Efficient Optimization for Reinforcement Learning from Human Feedback (NeurIPS 2026 · similarity 0.89)
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.
