BASIS: Single-Rollout RL Training for LLM Reasoning Cuts Compute While Sharing Cross-Prompt Signal

arXiv·medium signal

Gong et al. introduce BASIS, a critic-free RL post-training algorithm that samples only one rollout per prompt but leverages rich cross-prompt information sharing within each batch. This addresses the core tradeoff between computational efficiency (fewer rollouts) and sample efficiency (better value estimates) in reinforcement learning with verifiable rewards for reasoning. The approach maintains training quality while significantly reducing generation cost.