Functional Natural Policy Gradients

Reinforcement learning has become a central tool for large language model (LLM) post-training, where policy gradient methods are routinely deployed off-policy, even though vanilla policy gradient assumes on-policy sampling. We study when such off-policy deployment is justified, and what roles its core ingredients, including importance sampling, KL regularization, and baselines, play in handling the resulting distribution shift. We show that for a general class of off-policy policy gradient updates, consistency, in the sense that the optimal policy remains a stationary point, requires one of two structures at the optimal policy: trajectory-level importance weighting, which suffers from the well-known curse of horizon, or KL regularization with a baseline equal to the optimal soft value, instantiated by the grouped mean. Focusing on the KL route, we identify it with the group-relative policy gradient (GRPG) update, establish an equivalence to a trajectory-level Bellman residual minimization objective, and obtain a finite-sample regret bound under all-policy coverage. This also gives a theoretical account of a variant of the empirically successful GRPO objective. We further uncover a hidden role of baselines in the off-policy setting: a constant shift in the baseline implements pessimism, the standard remedy in offline RL for partial data coverage. We instantiate this via the asymmetric REINFORCE objective, and connect it to the game-theoretic offline RL framework, showing a regret bound with single-policy coverage. Together, these results offer a unified theoretical view of off-policy policy gradient in LLM post-training and connect it to classical ideas in offline RL.

Distributed reinforcement learning trains on data from stale, buggy, or mismatched actors, producing actions with high surprisal (negative log-probability) under the learner's policy. The core difficulty is not surprising data per se, but negative learning from surprising data. High-surprisal failures can dominate finite-batch updates through large perpendicular components, while high-surprisal successes reveal opportunities the current policy would otherwise miss. The Delightful Policy Gradient (DG) separates these cases by gating each update with delight, the product of advantage and surprisal, suppressing rare failures and preserving rare successes without behavior probabilities. In a tabular analysis, DG suppresses the perpendicular second moment of high-surprisal failures by a policy-overlap factor that vanishes as the learner improves. The advantage sign is essential for surprisal-based filtering: any learner-probability-only gate that suppresses rare failures also suppresses rare successes. On MNIST with simulated staleness, DG without off-policy correction outperforms importance-weighted PG with exact behavior probabilities. On a transformer sequence task with staleness, actor bugs, reward corruption, and rare discovery, DG often achieves nearly order-of-magnitude lower error. When all four frictions act simultaneously, its sample-efficiency advantage is order-of-magnitude and grows with task complexity.

Policy learning has received substantial attention with the goal of learning policies from observational data for decision-making. A majority of work in this space has focused on developing algorithms for computing policies that minimize regret compared to the optimal policy. However, in many practical settings, there is insufficient data to obtain low regret. As a result, recent work has shifted attention to alternative objectives, most notably, studying whether it is possible to learn an improving policy that statistically significantly outperforms baseline policies. We argue that there is substantial merit in studying a broader range of policy learning problems. When there is insufficient data to learn an improving policy, there may still be useful questions that can be answered. To this end, we provide a mathematical framework for studying the relationships between policy learning problems. We formalize three problems within our framework: beyond the optimal policy problem and the improving policy problem, we also propose the policy existence problem, which aims to determine if an improving policy exists. Within our framework, we show that the policy existence problem reduces to the improving policy problem, which in turn reduces to the optimal policy problem; these reductions prove that each problem is at least as easy as the next one (in sample complexity). A key question remains: is this hardness strict? We provide partial answers. First, the gap between the optimal policy and improving policy problems is strict. For the improving policy and policy existence problems, we prove that a sublinear polynomial gap exists under natural conditions on improving policy learning algorithms. Thus, we may be able to answer questions about the existence of an improving policy even when we cannot find one. These results highlight the value in studying a broader range of policy learning problems.