Long-Horizon Model-Based Offline Reinforcement Learning Without Conservatism

Popular offline reinforcement learning (RL) methods rely on conservatism, either by penalizing out-of-dataset actions or by restricting planning horizons. In this work, we question the universality of this principle and instead revisit a complementary one: a Bayesian perspective. Rather than enforcing conservatism, the Bayesian approach tackles epistemic uncertainty in offline data by modeling a posterior distribution over plausible world models and training a history-dependent agent to maximize expected rewards, enabling test-time generalization. We first illustrate, in a bandit setting, that Bayesianism excels on low-quality datasets where conservatism fails. We then scale the principle to realistic tasks, identifying key design choices, such as layer normalization in the world model and adaptive long-horizon planning, that mitigate compounding error and value overestimation. These yield our practical algorithm, Neubay, grounded in the neutral Bayesian principle. On D4RL and NeoRL benchmarks, Neubay generally matches or surpasses leading conservative algorithms, achieving new state-of-the-art on 7 datasets. Notably, it succeeds with planning horizons of several hundred steps, challenging common belief. Finally, we characterize when Neubay is preferable to conservatism, laying the foundation for a new direction in offline and model-based RL.