BASIS: Batchwise Advantage Estimation from Single-Rollout Information Sharing for LLM Reasoning
For researchers and practitioners training LLMs with reinforcement learning, BASIS offers a computationally efficient method that improves sample efficiency without requiring multiple rollouts per prompt.
BASIS introduces a critic-free post-training algorithm that improves value function estimation by sharing information across prompts in a batch, reducing MSE by 69% compared to REINFORCE++ with a single rollout and achieving lower MSE than group mean estimators with 8 rollouts, leading to better policy optimization with less training time.
Reinforcement learning with verifiable rewards has become a standard recipe for improving the reasoning abilities of large language models. Existing algorithms face a tradeoff between computational efficiency and sample efficiency in value estimation and policy learning. We introduce BASIS, a critic-free post-training algorithm designed to address this tradeoff. At each online training step, BASIS samples only one rollout per prompt, but leverages rich information across prompts in the entire batch to improve value function estimation. Our experiments demonstrate that BASIS reduces MSE in value function estimation by 69% compared to REINFORCE++, a representative single-rollout baseline, and achieves lower MSE with one rollout than group mean estimators with 8 rollouts. This improvement in value estimation translates to better policy optimization: using substantially less training time, BASIS achieves performance close to multi-rollout GRPO-type baselines and often outperforms single-rollout REINFORCE-type baselines.