Research map / Reinforcement learning / RL from feedback
IsoCompute Playbook: Optimally Scaling Sampling Compute for LLM RL
Zhoujun Cheng, Yutao Xie, Yuxiao Qu, Amrith Setlur, Shibo Hao, Varad Pimpalkhute, Tongtong Liang, Feng Yao, Zhengzhong Liu, Eric P. Xing, Virginia Smith, Ruslan Salakhutdinov and 3 more
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 |
|---|---|---|---|
| increase parallel rollouts per problem n as sampling compute grows | GRPO on Qwen2.5-7B-Instruct · Easy and Hard Guru-Math splits | validation avg@4 reward | Improved |
| raise n beyond the saturation point (n=2,048 vs n=512) (ablation) | GRPO on Qwen2.5-7B-Instruct · Easy Guru-Math split | validation avg@4 reward | No clear effect |
| use very large n=512 on the hard set (ablation) | GRPO on Qwen2.5-7B-Instruct · Hard Guru-Math split | final validation reward | Hurt |
| use larger n on the easy set (ablation) | GRPO on Qwen2.5-7B-Instruct · Easy split | worst@4 (sharpening) | Improved |
| use larger n on the hard set (ablation) | GRPO on Qwen2.5-7B-Instruct · Hard split | best@4 (coverage) | Improved |
| use larger n on the extremely hard set (pass@128 = 0) (ablation) | GRPO on Qwen2.5-7B-Instruct · Extremely Hard set | worst@4 and average reward | Hurt |
| use square-root learning-rate scaling with batch size instead of constant or linear (ablation) | GRPO on Qwen2.5-7B-Instruct · Easy set, B=8,192 | convergence stability | Improved |
| add KL and entropy regularization on the easy set (ablation) | GRPO on Qwen2.5-7B-Instruct · Easy split | training stability | Improved |
| add KL and entropy regularization on the hard set (ablation) | GRPO on Qwen2.5-7B-Instruct · Hard split | training stability | Hurt |
| add zero-variance filtering to KL and entropy terms (ablation) | GRPO on Qwen2.5-7B-Instruct · Hard split | training stability | No clear effect |
| decouple baseline estimation: 256 rollouts for advantages but update with 64 (ablation) | GRPO on Qwen2.5-7B-Instruct · Bp=128 | best validation reward | No clear effect |
| vary problems per batch Bp at fixed n (ablation) | GRPO on Qwen2.5-7B-Instruct · Easy set | validation reward | No clear effect |
Closest papers by meaning
- Scaling Test-Time Compute Without Verification or RL is Suboptimal (ICML 2025 · similarity 0.90)
- Compute-Optimal Scaling for Value-Based Deep RL (NeurIPS 2025 · similarity 0.89)
- SRL: Scaling Distributed Reinforcement Learning to Over Ten Thousand Cores (ICLR 2024 · similarity 0.89)
- UFO-RL: Uncertainty-Focused Optimization for Efficient Reinforcement Learning Data Selection (NeurIPS 2025 · similarity 0.88)
- Bigger, Regularized, Optimistic: scaling for compute and sample efficient continuous control (NeurIPS 2024 · similarity 0.88)
- Horizon Reduction Makes RL Scalable (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.
