Saved in:
| Main Authors: | , , , , , , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2603.12151 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866918384456695808 |
|---|---|
| author | Cheng, Zhoujun Xie, Yutao Qu, Yuxiao Setlur, Amrith Hao, Shibo Pimpalkhute, Varad Liang, Tongtong Yao, Feng Liu, Zhengzhong Xing, Eric Smith, Virginia Salakhutdinov, Ruslan Hu, Zhiting Killian, Taylor Kumar, Aviral |
| author_facet | Cheng, Zhoujun Xie, Yutao Qu, Yuxiao Setlur, Amrith Hao, Shibo Pimpalkhute, Varad Liang, Tongtong Yao, Feng Liu, Zhengzhong Xing, Eric Smith, Virginia Salakhutdinov, Ruslan Hu, Zhiting Killian, Taylor Kumar, Aviral |
| contents | While scaling laws guide compute allocation for LLM pre-training, analogous prescriptions for reinforcement learning (RL) post-training of large language models (LLMs) remain poorly understood. We study the compute-optimal allocation of sampling compute for on-policy RL methods in LLMs, framing scaling as a compute-constrained optimization over three resources: parallel rollouts per problem, number of problems per batch, and number of update steps. We find that the compute-optimal number of parallel rollouts per problem increases predictably with compute budget and then saturates. This trend holds across both easy and hard problems, though driven by different mechanisms: solution sharpening on easy problems and coverage expansion on hard problems. We further show that increasing the number of parallel rollouts mitigates interference across problems, while the number of problems per batch primarily affects training stability and can be chosen within a broad range. Validated across base models and data distributions, our results recast RL scaling laws as prescriptive allocation rules and provide practical guidance for compute-efficient LLM RL post-training. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2603_12151 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | IsoCompute Playbook: Optimally Scaling Sampling Compute for LLM RL Cheng, Zhoujun Xie, Yutao Qu, Yuxiao Setlur, Amrith Hao, Shibo Pimpalkhute, Varad Liang, Tongtong Yao, Feng Liu, Zhengzhong Xing, Eric Smith, Virginia Salakhutdinov, Ruslan Hu, Zhiting Killian, Taylor Kumar, Aviral Machine Learning Artificial Intelligence While scaling laws guide compute allocation for LLM pre-training, analogous prescriptions for reinforcement learning (RL) post-training of large language models (LLMs) remain poorly understood. We study the compute-optimal allocation of sampling compute for on-policy RL methods in LLMs, framing scaling as a compute-constrained optimization over three resources: parallel rollouts per problem, number of problems per batch, and number of update steps. We find that the compute-optimal number of parallel rollouts per problem increases predictably with compute budget and then saturates. This trend holds across both easy and hard problems, though driven by different mechanisms: solution sharpening on easy problems and coverage expansion on hard problems. We further show that increasing the number of parallel rollouts mitigates interference across problems, while the number of problems per batch primarily affects training stability and can be chosen within a broad range. Validated across base models and data distributions, our results recast RL scaling laws as prescriptive allocation rules and provide practical guidance for compute-efficient LLM RL post-training. |
| title | IsoCompute Playbook: Optimally Scaling Sampling Compute for LLM RL |
| topic | Machine Learning Artificial Intelligence |
| url | https://arxiv.org/abs/2603.12151 |