Research map / Large language models / Instruction and fine-tuning
On the Entropy Dynamics in Reinforcement Fine-Tuning of Large Language Models
Shumin Wang, Yuexiang Xie, Wenhao Zhang, Yuchang Sun, Yanxi Chen, Yaliang Li, Yanyong Zhang
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 |
|---|---|---|---|
| apply ClipB (batch-normalized entropy-discriminator clipping) on top of GRPO | GRPO, Qwen2.5-7B-Instruct · Average over 5 math benchmarks | Avg@K: 39.36 -> 41.38 | Improved |
| apply ClipV (vocabulary-normalized clipping) on top of GRPO | GRPO, Qwen2.5-7B-Instruct · Average over 5 math benchmarks | Avg@K: 39.36 -> 40.61 | Improved |
| apply ClipV on top of GRPO | GRPO, Qwen2.5-14B-Instruct · Average over 5 math benchmarks | Avg@K: 45.57 -> 49.44 | Improved |
| apply ClipB on top of GRPO | GRPO, Qwen2.5-14B-Instruct · Average over 5 math benchmarks | Avg@K: 45.57 -> 48.81 | Improved |
| apply ClipB on top of GRPO | GRPO, Qwen2.5-7B-Instruct · DAPO500 | Pass@8: 76.8 -> 80.2 | Improved |
| apply ClipB on top of GRPO | GRPO, Qwen2.5-7B-Instruct · AIME24 | Pass@32: 50.00 -> 56.67 | Improved |
| compare other entropy methods (Forking Tokens, Rewarding Unlikely) against GRPO | GRPO, Qwen2.5-7B-Instruct · Average over 5 math benchmarks | Avg@K: 39.36 -> 38.33 (Forking Tokens), 38.78 (Rewarding Unlikely) | Hurt |
| apply Clip-Cov or KL-Cov on top of GRPO | GRPO, Qwen2.5-7B-Instruct · Average over 5 math benchmarks | Avg@K: 39.36 -> 39.37 (Clip-Cov), 39.54 (KL-Cov) | No clear effect |
| retain only updates for tokens with S*>0 in positive samples (ablation) | GRPO training · Entropy dynamics, DAPO-Math-17k | entropy | Hurt |
| mask gradients of S*>0 tokens in positive samples (ablation) | GRPO training · Entropy dynamics, DAPO-Math-17k | entropy | Mixed |
| increase mu threshold in ClipB/ClipV (ablation) | ClipB, ClipV · Clip fraction, DAPO-Math-17k | clip fraction | Improved |
| apply ClipB to GRPO | GRPO, Qwen2.5-7B-Instruct · Pass-rate distribution, test set | share of all-fail questions (%): 66.7 -> 63.3 | Improved |
Closest papers by meaning
- SCOPE-RL: Stable and Quantitative Control of Policy Entropy in RL Post-Training (NeurIPS 2026 · similarity 0.91)
- Reasoning with Exploration: An Entropy Perspective (AAAI 2026 · similarity 0.90)
- Follow the Winners: Conservative Policy Improvement with the Cross-Entropy Method for Critic-Free RFT (NeurIPS 2026 · similarity 0.90)
- Learning Dynamics of LLM Finetuning (ICLR 2025 · similarity 0.89)
- Vanishing Gradients in Reinforcement Finetuning of Language Models (ICLR 2024 · similarity 0.89)
- AEM: Adaptive Entropy Modulation for Multi-Turn Agentic Reinforcement Learning (NeurIPS 2026 · 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.
