Search papers, labs, and topics across Lattice.
This paper investigates the use of experience replay in LLM post-training, challenging the common belief that on-policy data is strictly necessary. They formalize the design of replay buffers as a trade-off between data staleness, diversity, and generation cost. Results demonstrate that a carefully designed replay buffer can significantly reduce inference compute during training without sacrificing, and sometimes even improving, final model performance and policy entropy.
Reusing old data can actually *improve* LLM training, slashing compute costs without hurting performance.
While Experience Replay - the practice of storing rollouts and reusing them multiple times during training - is a foundational technique in general RL, it remains largely unexplored in LLM post-training due to the prevailing belief that fresh, on-policy data is essential for high performance. In this work, we challenge this assumption. We present a systematic study of replay buffers for LLM post-training, formalizing the optimal design as a trade-off between staleness-induced variance, sample diversity and the high computational cost of generation. We show that strict on-policy sampling is suboptimal when generation is expensive. Empirically, we show that a well-designed replay buffer can drastically reduce inference compute without degrading - and in some cases even improving - final model performance, while preserving policy entropy.
