Efficient RL Training for LLMs with Experience Replay

Charles Arnal, Vivien Cabannes, Taco Cohen, Julia Kempe, Rémi Munos
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:3570-3599, 2026.

Abstract

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.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-arnal26a, title = {Efficient {RL} Training for {LLM}s with Experience Replay}, author = {Arnal, Charles and Cabannes, Vivien and Cohen, Taco and Kempe, Julia and Munos, R\'{e}mi}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {3570--3599}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/arnal26a/arnal26a.pdf}, url = {https://proceedings.mlr.press/v306/arnal26a.html}, abstract = {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.} }
Endnote
%0 Conference Paper %T Efficient RL Training for LLMs with Experience Replay %A Charles Arnal %A Vivien Cabannes %A Taco Cohen %A Julia Kempe %A Rémi Munos %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-arnal26a %I PMLR %P 3570--3599 %U https://proceedings.mlr.press/v306/arnal26a.html %V 306 %X 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.
APA
Arnal, C., Cabannes, V., Cohen, T., Kempe, J. & Munos, R.. (2026). Efficient RL Training for LLMs with Experience Replay. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:3570-3599 Available from https://proceedings.mlr.press/v306/arnal26a.html.

Related Material