IsoCompute Playbook: Optimally Scaling Sampling Compute for LLM RL

arXiv 2603.12151·high signal

First Chinchilla-style prescriptions for RL post-training compute allocation, framing sampling as a constrained optimization over three resources: parallel rollouts per problem, problems per batch, and update steps. Compute-optimal rollouts increase predictably with budget then saturate; easy problems benefit from 'solution sharpening' while hard problems benefit from 'coverage expansion'. Directly actionable for practitioners running RLHF or RLVR: the playbook tells you how to redistribute sampling budget at any compute tier.