Trust Region Masking for Long-Horizon LLM Reinforcement Learning

• URL: http://arxiv.org/abs/2512.23075v1
• Date: Sun, 28 Dec 2025 20:41:59 GMT
• Title: Trust Region Masking for Long-Horizon LLM Reinforcement Learning
• Authors: Yingru Li, Jiacai Liu, Jiawei Xu, Yuxuan Tong, Ziniu Li, Baoxiang Wang,
• Abstract summary: Policy gradient methods for large language models optimize a surrogate objective computed from samples of a rollout policy.<n>When $_textroll ne _$, there is approximation error between the surrogate and the true objective.<n>We propose Trust Region Masking (TRM), which excludes entire sequences from gradient computation if any token violates the trust region.
• Score: 20.589897184824878
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Policy gradient methods for large language models optimize a surrogate objective computed from samples of a rollout policy $π_{\text{roll}}$. When $π_{\text{roll}} \ne π_θ$, there is approximation error between the surrogate and the true objective. Prior work has shown that this off-policy mismatch is unavoidable in modern LLM-RL due to implementation divergence, mixture-of-experts routing discontinuities, and distributed training staleness. Classical trust region bounds on the resulting error scale as $O(T^2)$ with sequence length $T$, rendering them vacuous for long-horizon tasks. We derive two tighter bounds: a Pinsker-Marginal bound scaling as $O(T^{3/2})$ and a Mixed bound scaling as $O(T)$. Crucially, both bounds depend on $D_{kl}^{tok,max}$ -- the maximum token-level KL divergence across all positions in a sequence. This is inherently a sequence-level quantity: it requires examining the entire trajectory to compute, and therefore cannot be controlled by token-independent methods like PPO clipping. We propose Trust Region Masking (TRM), which excludes entire sequences from gradient computation if any token violates the trust region, providing the first non-vacuous monotonic improvement guarantees for long-horizon LLM-RL.

Related papers

• Rank-Aware Spectral Bounds on Attention Logits for Stable Low-Precision Training [0.0]
  Attention scores in transformers are bilinear forms $S_ij = x_itop M x_j / sqrtd_h$ whose maximum magnitude governs overflow risk in low-precision training.<n>We derive a emphrank-aware concentration inequality: when the interaction matrix $M = WQ WKtop$ has rank $r ll d$, tail probabilities for $max_i,j|S_ij|$ decay as $exp(-d22/(
  arXiv  Detail & Related papers  (2026-02-21T14:29:22Z)
• Robust Layerwise Scaling Rules by Proper Weight Decay Tuning [50.11170157029911]
  In modern scale-invariant architectures, training quickly enters an degrading-governed steady state.<n>We introduce a weight-decay scaling rule for AdamW that preserves sublayer gain across widths.<n>Our results extend $mu$P beyond the near-init regime by explicitly controlling the steady-state scales set by parameters.
  arXiv  Detail & Related papers  (2025-10-17T02:58:35Z)
• Arithmetic-Mean $μ$P for Modern Architectures: A Unified Learning-Rate Scale for CNNs and ResNets [9.94514344279733]
  Arithmetic-Mean $mu$P constrains not each individual layer but the network-wide average one-step pre-activation second moment to a constant scale.<n>We prove that, for one- and two-dimensional convolutional networks, the maximal-update learning rate satisfies $etastar(L)propto L-3/2$; with zero padding, boundary effects are constant-level as $Ngg k$.
  arXiv  Detail & Related papers  (2025-10-05T19:22:50Z)
• Mean-Field Sampling for Cooperative Multi-Agent Reinforcement Learning [8.400105595501158]
  We propose a new $textttSUBPLE-MFQ$ ($textbfSubsample$-$textbfMean-$textbfF$ield-$textbfQ$-learning) and a decentralized randomized policy for a system with $n$ agents.<n>We prove that this learned policy converges to the optimal policy on the order of $tilde$O (1/sqrtk)$ as the number of subsampled agents $k$ increases.
  arXiv  Detail & Related papers  (2024-12-01T03:45:17Z)
• Best-of-Both-Worlds Algorithms for Linear Contextual Bandits [11.94312915280916]
  We study best-of-both-worlds algorithms for $K$-armed linear contextual bandits. Our algorithms deliver near-optimal regret bounds in both the adversarial and adversarial regimes.
  arXiv  Detail & Related papers  (2023-12-24T08:27:30Z)
• Scalable Primal-Dual Actor-Critic Method for Safe Multi-Agent RL with General Utilities [12.104551746465932]
  We investigate safe multi-agent reinforcement learning, where agents seek to collectively maximize an aggregate sum of local objectives while satisfying their own safety constraints. Our algorithm converges to a first-order stationary point (FOSP) at the rate of $mathcalOleft(T-2/3right)$. In the sample-based setting, we demonstrate that, with high probability, our algorithm requires $widetildemathcalOleft(epsilon-3.5right)$ samples to achieve an $epsilon$-FOSP.
  arXiv  Detail & Related papers  (2023-05-27T20:08:35Z)
• Restarted Bayesian Online Change-point Detection for Non-Stationary Markov Decision Processes [12.229154524476405]
  We introduce a variant of the Restarted Bayesian Online Change-Point Detection algorithm (R-BOCPD) We propose an improved version of the UCRL2 algorithm for MDPs with state transition kernel sampled from a multinomial distribution. We show that R-BOCPD-UCRL2 enjoys a favorable regret bound of $Oleft(D O sqrtA T K_T logleft (fracTdelta right) + fracK_Tdeltaminlimits_ell.
  arXiv  Detail & Related papers  (2023-04-01T05:26:41Z)
• Adversarial Online Multi-Task Reinforcement Learning [12.421997449847153]
  We consider the adversarial online multi-task reinforcement learning setting. In each of $K$ episodes the learner is given an unknown task taken from a finite set of $M$ unknown finite-horizon MDP models. The learner's objective is to generalize its regret with respect to the optimal policy for each task.
  arXiv  Detail & Related papers  (2023-01-11T02:18:26Z)
• Horizon-Free and Variance-Dependent Reinforcement Learning for Latent Markov Decision Processes [62.90204655228324]
  We study regret minimization for reinforcement learning (RL) in Latent Markov Decision Processes (LMDPs) with context in hindsight. We design a novel model-based algorithmic framework which can be instantiated with both a model-optimistic and a value-optimistic solver.
  arXiv  Detail & Related papers  (2022-10-20T21:32:01Z)
• Near-Optimal Regret Bounds for Multi-batch Reinforcement Learning [54.806166861456035]
  We study the episodic reinforcement learning (RL) problem modeled by finite-horizon Markov Decision Processes (MDPs) with constraint on the number of batches. We design a computational efficient algorithm to achieve near-optimal regret of $tildeO(sqrtSAH3Kln (1/delta))$tildeO(cdot) hides logarithmic terms of $(S,A,H,K)$ in $K$ episodes. Our technical contribution are two-fold: 1) a near-optimal design scheme to explore
  arXiv  Detail & Related papers  (2022-10-15T09:22:22Z)
• Settling the Sample Complexity of Model-Based Offline Reinforcement Learning [50.5790774201146]
  offline reinforcement learning (RL) learns using pre-collected data without further exploration. Prior algorithms or analyses either suffer from suboptimal sample complexities or incur high burn-in cost to reach sample optimality. We demonstrate that the model-based (or "plug-in") approach achieves minimax-optimal sample complexity without burn-in cost.
  arXiv  Detail & Related papers  (2022-04-11T17:26:19Z)
• Nearly Minimax Optimal Reinforcement Learning for Linear Mixture Markov Decision Processes [91.38793800392108]
  We study reinforcement learning with linear function approximation where the underlying transition probability kernel of the Markov decision process (MDP) is a linear mixture model. We propose a new, computationally efficient algorithm with linear function approximation named $textUCRL-VTR+$ for the aforementioned linear mixture MDPs. To the best of our knowledge, these are the first computationally efficient, nearly minimax optimal algorithms for RL with linear function approximation.
  arXiv  Detail & Related papers  (2020-12-15T18:56:46Z)
• Provably Efficient Safe Exploration via Primal-Dual Policy Optimization [105.7510838453122]
  We study the Safe Reinforcement Learning (SRL) problem using the Constrained Markov Decision Process (CMDP) formulation. We present an provably efficient online policy optimization algorithm for CMDP with safe exploration in the function approximation setting.
  arXiv  Detail & Related papers  (2020-03-01T17:47:03Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.