Long-Horizon Model-Based Offline Reinforcement Learning Without Explicit Conservatism

Tianwei Ni, Esther Derman, Vineet Jain, Vincent Taboga, Siamak Ravanbakhsh, Pierre-Luc Bacon
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:92875-92924, 2026.

Abstract

Popular offline reinforcement learning (RL) methods rely on explicit conservatism, penalizing out-of-dataset actions or restricting rollout horizons. We question the universality of this principle and revisit a complementary Bayesian perspective for test-time adaptation. By modeling a posterior over world models and training a history-dependent agent to maximize expected return, the Bayesian approach directly addresses epistemic uncertainty without explicit conservatism. We first illustrate in a bandit setting that Bayesianism excels on low-quality datasets where conservatism fails. Scaling to realistic tasks, we find that long-horizon rollouts are essential to control value overestimation once conservatism is removed. We introduce design choices that enable learning from long-horizon rollouts while mitigating compounding model errors, yielding our algorithm, NEUBAY, grounded in the neutral Bayesian principle. On D4RL and NeoRL benchmarks, NEUBAY is competitive with leading conservative algorithms, achieving new state-of-the-art on 7 datasets with rollout horizons of several hundred steps. Finally, we characterize datasets by quality and coverage to identify when NEUBAY is preferable to conservative methods.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-ni26a, title = {Long-Horizon Model-Based Offline Reinforcement Learning Without Explicit Conservatism}, author = {Ni, Tianwei and Derman, Esther and Jain, Vineet and Taboga, Vincent and Ravanbakhsh, Siamak and Bacon, Pierre-Luc}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {92875--92924}, 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/ni26a/ni26a.pdf}, url = {https://proceedings.mlr.press/v306/ni26a.html}, abstract = {Popular offline reinforcement learning (RL) methods rely on explicit conservatism, penalizing out-of-dataset actions or restricting rollout horizons. We question the universality of this principle and revisit a complementary Bayesian perspective for test-time adaptation. By modeling a posterior over world models and training a history-dependent agent to maximize expected return, the Bayesian approach directly addresses epistemic uncertainty without explicit conservatism. We first illustrate in a bandit setting that Bayesianism excels on low-quality datasets where conservatism fails. Scaling to realistic tasks, we find that long-horizon rollouts are essential to control value overestimation once conservatism is removed. We introduce design choices that enable learning from long-horizon rollouts while mitigating compounding model errors, yielding our algorithm, NEUBAY, grounded in the neutral Bayesian principle. On D4RL and NeoRL benchmarks, NEUBAY is competitive with leading conservative algorithms, achieving new state-of-the-art on 7 datasets with rollout horizons of several hundred steps. Finally, we characterize datasets by quality and coverage to identify when NEUBAY is preferable to conservative methods.} }
Endnote
%0 Conference Paper %T Long-Horizon Model-Based Offline Reinforcement Learning Without Explicit Conservatism %A Tianwei Ni %A Esther Derman %A Vineet Jain %A Vincent Taboga %A Siamak Ravanbakhsh %A Pierre-Luc Bacon %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-ni26a %I PMLR %P 92875--92924 %U https://proceedings.mlr.press/v306/ni26a.html %V 306 %X Popular offline reinforcement learning (RL) methods rely on explicit conservatism, penalizing out-of-dataset actions or restricting rollout horizons. We question the universality of this principle and revisit a complementary Bayesian perspective for test-time adaptation. By modeling a posterior over world models and training a history-dependent agent to maximize expected return, the Bayesian approach directly addresses epistemic uncertainty without explicit conservatism. We first illustrate in a bandit setting that Bayesianism excels on low-quality datasets where conservatism fails. Scaling to realistic tasks, we find that long-horizon rollouts are essential to control value overestimation once conservatism is removed. We introduce design choices that enable learning from long-horizon rollouts while mitigating compounding model errors, yielding our algorithm, NEUBAY, grounded in the neutral Bayesian principle. On D4RL and NeoRL benchmarks, NEUBAY is competitive with leading conservative algorithms, achieving new state-of-the-art on 7 datasets with rollout horizons of several hundred steps. Finally, we characterize datasets by quality and coverage to identify when NEUBAY is preferable to conservative methods.
APA
Ni, T., Derman, E., Jain, V., Taboga, V., Ravanbakhsh, S. & Bacon, P.. (2026). Long-Horizon Model-Based Offline Reinforcement Learning Without Explicit Conservatism. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:92875-92924 Available from https://proceedings.mlr.press/v306/ni26a.html.

Related Material