Research map / Reinforcement learning / RL from feedback
GRPO is Secretly a Process Reward Model
Michael Sullivan, Alexander Koller
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 |
|---|---|---|---|
| scale token loss by inverse process-set size (lambda-GRPO) (ablation) | GRPO · toy tree n=1, r_neg=-1.0 | target path frequency: 0.0000 -> 0.7535 | Improved |
| scale token loss by inverse process-set size (ablation) | GRPO · toy tree n=2, r_neg=-1.5 | target path frequency: 0.0000 -> 0.3820 | Improved |
| scale token loss by inverse process-set size (ablation) | GRPO · toy tree n=1, r_neg=-0.5 | target path frequency: 0.3995 -> 0.5663 | Improved |
| scale token loss by inverse process-set size (ablation) | GRPO · Qwen average accuracy, beta=0.0 | accuracy: 0.4837 -> 0.5576 | Improved |
| scale token loss by inverse process-set size (ablation) | GRPO · Qwen average accuracy, beta=0.04 | accuracy: 0.5394 -> 0.5739 | Improved |
| scale token loss by inverse process-set size (ablation) | GRPO · Llama average accuracy, beta=0.0 | accuracy: 0.0842 -> 0.1001 | Improved |
| scale token loss by inverse process-set size (ablation) | GRPO · Llama average accuracy, beta=0.04 | accuracy: 0.0996 -> 0.0973 | Hurt |
| scale token loss by inverse process-set size (ablation) | GRPO · Qwen MATH-500, beta=0.04 | accuracy: 0.8660 -> 0.8340 | Hurt |
| scale token loss by inverse process-set size (ablation) | GRPO · Qwen AIME24, beta=0.04 | accuracy: 0.3000 -> 0.4000 | Improved |
| scale token loss by inverse process-set size (ablation) | GRPO · Llama AMC23, beta=0.04 | accuracy: 0.1750 -> 0.0750 | Hurt |
| scale token loss by inverse process-set size (ablation) | GRPO · validation accuracy during training | validation accuracy: more than 10% increase in less than half the training steps | Improved |
| train Qwen with GRPO instead of leaving base model (beta=0.0) | DeepSeek-R1-Distill-Qwen-1.5B · average accuracy | accuracy: 0.5175 -> 0.4837 | Hurt |
Closest papers by meaning
- $f$-GRPO & Beyond: Divergence-Based Reinforcement Learning Algorithms for General LLM Alignment (NeurIPS 2026 · similarity 0.90)
- TIC-GRPO: Provable and Efficient Optimization for Reinforcement Learning from Human Feedback (NeurIPS 2026 · similarity 0.89)
- PAIR: Prefix-Aware Internal Reward Model for Multi-Turn Agent Optimization (NeurIPS 2026 · similarity 0.89)
- Learn from the Gap: Differential-Aware Advantage Pruning with Adaptive Rollout Sampling for GRPO (NeurIPS 2026 · similarity 0.89)
- Unveiling Implicit Advantage Symmetry: Why GRPO Struggles with Exploration and Difficulty Adaptation (NeurIPS 2026 · similarity 0.88)
- Submodular Reinforcement Learning (ICLR 2024 · similarity 0.87)
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.
