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

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.

Many decision-making systems must learn from past data because real-world trial and error can be costly or unsafe. Most current methods handle this by being deliberately cautious: they avoid behavior that was not well represented in the data. This improves safety, but it can also trap the system into repeating limited past behavior, especially when the data is weak or incomplete. This paper revisits a different idea: instead of assuming unfamiliar actions are bad, the agent reasons over many possible ways the world might work and adapts when it is tested. We show that this approach is especially helpful when good actions are rare or missing from the data. We also identify a key ingredient for making it work in realistic tasks: the agent must imagine long sequences of future events, while stopping when its predictions become unreliable. Our method, NEUBAY, turns this idea into a practical algorithm. Across standard benchmarks, it is competitive with leading deliberately cautious methods and performs especially well on low-quality or moderately covered data. More broadly, this work points toward offline learning agents that are not just safely trained from past data, but can also continually adapt after deployment.