Not All Rollouts are Useful: Down-Sampling Rollouts in LLM Reinforcement Learning (2504.13818v2)

Abstract: Reinforcement learning with verifiable rewards (RLVR) has emerged as a powerful paradigm for enhancing reasoning capabilities in LLMs. However, it is constrained by a fundamental asymmetry in computation and memory requirements: rollout generation is embarrassingly parallel and memory-light, whereas policy updates are communication-heavy and memory-intensive. To address this, we introduce PODS (Policy Optimization with Down-Sampling). PODS produces numerous rollouts in parallel, then trains on only an informative subset, preserving learning signals while slashing update cost. We instantiate PODS with max-variance down-sampling, a principled criterion that maximises reward diversity and show it admits an $O(n\log n)$ solution. Empirically, coupling PODS with Group Relative Policy Optimization (GRPO) achieves superior performance over standard GRPO across different reasoning benchmarks and hardware environments.