Trust Region Masking for Long-Horizon LLM Reinforcement Learning

TL;DR

This paper introduces Trust Region Masking, a novel method that provides the first non-vacuous guarantees for long-horizon reinforcement learning with large language models by controlling sequence-level divergences.

Contribution

It derives new divergence bounds that scale better with sequence length and proposes Trust Region Masking to ensure monotonic improvement in long-horizon LLM-RL.

Findings

Derived divergence bounds with improved scaling ($O(T^{3/2})$, $O(T)$)

Proposed Trust Region Masking to control sequence divergence

Achieved the first non-vacuous guarantees for long-horizon LLM-RL

Abstract

Policy gradient methods for Large Language Models optimize a policy $\pi_\theta$ via a surrogate objective computed from samples of a rollout policy $\pi_{\text{roll}}$. However, modern LLM-RL pipelines suffer from unavoidable implementation divergences -- backend discrepancies, Mixture-of-Experts routing discontinuities, and distributed training staleness -- causing off-policy mismatch ($\pi_{\text{roll}} \neq \pi_\theta$) and approximation errors between the surrogate and the true objective. We demonstrate that classical trust region bounds on this error scale as $O(T^2)$ with sequence length $T$, rendering them vacuous for long-horizon tasks. To address this, we derive a family of bounds -- both KL-based and TV-based -- including a Pinsker-Marginal bound ($O(T^{3/2})$), a Mixed bound ($O(T)$), and an Adaptive bound that strictly generalizes the Pinsker-Marginal bound via per-position importance-ratio decomposition. Taking the minimum over all bounds yields the tightest known guarantee across all divergence regimes. Crucially, all bounds depend on the maximum token-level divergence $D_{\mathrm{KL}}^{\mathrm{tok,max}}$ (or $D_{\mathrm{TV}}^{\mathrm{tok,max}}$), a sequence-level quantity that cannot be controlled by token-independent methods like PPO clipping. We propose Trust Region Masking (TRM), which masks entire sequences violating the trust region, enabling the first non-vacuous monotonic improvement guarantees for long-horizon LLM-RL.