Research map / Reinforcement learning / RL from feedback
GTPO and GRPO-S: Token and Sequence-Level Reward Shaping with Policy Entropy
Hongze Tan, Zihan Wang, Jianfei Pan, Jinghao Lin, Hao Wang, Yifan Wu, Tao Chen, Zhihang Zheng, Zhihao Tang, Haihua Yang
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 |
|---|---|---|---|
| use GTPO entropy-weighted token rewards instead of DAPO | Qwen2.5-7B · AIME 2024 | Pass@32: 35.51 -> 64.89 | Improved |
| use GTPO instead of GRPO | Qwen2.5-7B · AIME 2024 | Pass@32: 46.01 -> 64.89 | Improved |
| use GTPO instead of DAPO | Qwen2.5-32B · AIME 2024 | Pass@32: 59.02 -> 68.91 | Improved |
| use GTPO instead of DAPO w/ Forking Tokens | Qwen2.5-7B · AIME 2024 | Pass@32: 54.16 -> 64.89 | Improved |
| use GRPO-S instead of DAPO | Qwen2.5-7B · AIME 2025 | Mean@32: 13.33 -> 18.33 | Improved |
| use GTPO instead of DAPO | Qwen2.5-32B · AIME 2025 | Pass@32: 26.67 -> 36.67 | Improved |
| use GTPO instead of DAPO | Qwen2.5-7B · MATH 500 | Mean@32: 78.85 -> 80.16 | Improved |
| use GRPO-S instead of DAPO | Qwen2.5-32B · MATH 500 | Mean@32: 84.45 -> 85.79 | Improved |
| use GTPO on the smaller 7B model rather than 32B (gain in Pass@32 over baseline) | GTPO · AIME 2024 | Pass@32: +29.4 points on the 7B model, compared to +9.9 points on the 32B model | Improved |
| increase entropy bonus weight beta_2 from 0.1 to 0.2 (ablation) | GRPO-S · AIME 2024 | Mean@32 and Pass@32: 0.1 -> 0.2: clear decline in overall performance | Hurt |
| vary reward shaping hyperparameters across tested configurations (ablation) | GTPO and GRPO-S · AIME 2024 | Mean@32 and Pass@32 | Improved |
Closest papers by meaning
- STARE: Surprisal-Guided Token-Level Advantage Reweighting for Policy Entropy Stability (NeurIPS 2026 · similarity 0.91)
- Geometric-Mean Policy Optimization (ICLR 2026 · similarity 0.91)
- TIC-GRPO: Provable and Efficient Optimization for Reinforcement Learning from Human Feedback (NeurIPS 2026 · similarity 0.90)
- STAPO: Stabilizing Reinforcement Learning for LLMs by Silencing Rare Spurious Tokens (NeurIPS 2026 · similarity 0.89)
- Beyond the 80/20 Rule: High-Entropy Minority Tokens Drive Effective Reinforcement Learning for LLM Reasoning (NeurIPS 2025 · similarity 0.88)
- Act Only When It Pays: Efficient Reinforcement Learning for LLM Reasoning via Selective Rollouts (NeurIPS 2025 · similarity 0.88)
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.
