A novel policy gradient method is introduced, Hindsight Policy Optimization (HPO), that projects both the current policy distribution and the hindsight distribution into an intent space and extracts low-variance learning signals from the Wasserstein distance between them.
Abstract
Reinforcement learning (RL) has become a widely adopted technique for improving large language models (LLMs) on complex tasks. Despite this progress, existing RL methods still face challenges in training agents with longer-horizon interactions. One major bottleneck is distinguishing the contribution of different actions in long-horizon interaction, leading to high optimization variance. To address this, we introduce a novel policy gradient method, Hindsight Policy Optimization (HPO), that projects both the current policy distribution and the hindsight distribution into an intent space and extracts low-variance learning signals from the Wasserstein distance between them. We theoretically and empirically show that aggregating semantically similar states and actions in the intent space yields a bounded-variance estimator and improves policy performance stably. Our code is available online.
Reinforcement learning (RL) has achieved strong results in improving large language models (LLMs) on tasks with stationary, verifiable rewards, such as mathematical reasoning and code execution. In these settings, the environment follows fixed rules and does not adapt strategically to the agent. Strategic dialogue differs in this respect: the environment is another agent that adapts to the policy, and success depends on the interaction between the two sides. Despite this interactive nature, current RL approaches typically train a target agent against a fixed counterpart or simulator. We find that this training paradigm encourages the policy to exploit counterpart-specific regularities rather than learn strategies that generalize across counterparts. We call this problem the static-counterpart mismatch, which we quantify directly in our experiments. To address it, we propose Isolated Bilateral Reinforcement Learning (IB-RL), in which the two roles coevolve through joint rollouts while each role optimizes its own reward through fully independent advantages, action masks, and update paths. We evaluate frozen policies against fully independent held-out counterparts in both domains. On Vehicle TeleSales, IB-RL achieves 89.6% Success@1, compared to 84.6% for the best unilateral RL baseline. On Deal-or-NoDeal, it reaches 98.4% agreement against DeepSeek V4 Pro, compared to 86.4% for the best unilateral baseline. These results indicate that jointly training both roles with strict peragent isolation produces policies that generalize more effectively to unseen counterparts.
Senhao Wang, Chenghao Cai, Haitao Hu et al.· 0 citations
This work formalizes the hybrid LLM-planner and RL-controller architecture as a Goal-Augmented Markov Decision Process and shows that when the LLM per-state progress score is used as a bounded potential function, the resulting shaping term preserves the optimal policy set even when the LLM scores are inaccurate.
Christophe D. Hounwanou, John Emeka Eze, Yaé Ulrich Gaba· 0 citations
Production large language model (LLM) based systems such as coding agents, web navigators, and tool-calling assistants operate over multiple turns of interaction with users, tools, and environments. Pretrained LLMs, depending on their size, can either underperform in these settings due to misalignment with the system's interaction mechanics, or, when capable, incur prohibitive latency. Fine-tuning right-sized models addresses both accuracy and latency, but training such multi-turn agents requires Reinforcement Learning (RL), where the model acts as a policy optimizing long-horizon outcomes across sequential interactions. This poses challenges absent from single-turn settings: credit assignment over long trajectories, reward design for sparse and delayed feedback, state and context management as observation histories grow, environment scaling for parallel rollout collection, and training stability under prompt/environment distribution shift. This hands-on problem-solving tutorial provides both a rigorous algorithmic and practical introduction to multi-turn RL finetuning for LLMs. Using Amazon SageMaker AI, participants progress through four labs: (1) environment and reward function design, (2) multi-turn trajectory collection and Group Relative Policy Optimization (GRPO)-based training, (3) reward densification and credit assignment strategies, and (4) evaluation, failure diagnosis and deployment. We cover state-of-the-art multi-turn RL finetuning algorithms, turn-level vs. trajectory-level reward design, and production grade monitoring for reward hacking detection. The tutorial targets machine learning (ML) engineers, data scientists, and researchers who build agentic LLM systems. No prior RL experience is required. All materials will be publicly available on GitHub.
Zhe Wang, Sapana Chaudhary, Jiayu Li et al.· Proceedings of the 32nd ACM...· 0 citations
Terminal agents are a compelling application of large language models (LLMs), with the potential to integrate deeply into users'daily workflows. Reinforcement learning (RL) is a key technique for improving their capabilities, making scalable training environments a central challenge. Since public real-user interaction data are scarce, synthetic environments provide a practical alternative, but often suffer from domain gaps and limited fidelity, leading to poor generalization. Existing work mainly scales the quantity and diversity of synthetic environments, while reward-signal quality and the mechanisms governing generalization remain under-explored. We study how RL improves terminal agents and propose the Agentic Compositional Generalization hypothesis: rather than teaching new domain-specific skills from scratch, RL primarily shapes high-level decision-making behaviors that compose and route low-level skills acquired during pre-training and supervised fine-tuning (SFT). This account is consistent with our empirical results and suggests that verifier quality, which determines which behaviors are reinforced, is more important than simply increasing environment quantity or diversity. Motivated by this insight, we propose River, a simple training recipe that improves reward quality by filtering low-quality environments and augmenting outcome rewards with process-level behavior regularization. Using this recipe, our RL-trained agent achieves the best performance among evaluated open-source RL-trained 8B models across four terminal-agent benchmarks. River also generalizes across model families, scales, agent harnesses, and RL objectives. Using fewer than 30% of the TMax training environments, River improves RL gains by 106% and 30% on average for models ranging from 2B to 27B on Terminal-Bench-Lite and Terminal-Bench-v2.1, respectively.
Yi-Fan Yao, Bo Pang, Xuan-Phi Nguyen et al.· 0 citations
Progress-conditioned Group Policy Optimization is proposed, which uses first-visit observation coverage only when all samples in a group receive zero outcome reward, and consistently improves over group-based baselines, with particularly large gains on hard tasks.
Kaibing Yang, Guangfeng Cai, Shengtian Yang et al.· 0 citations
Off-policy reinforcement learning (RL) has become increasingly sample-efficient, enabling applications such as RL fine-tuning of Vision-Language-Action models into reliable, high-performing policies. World models offer a further lever for sample efficiency, as they predict state changes rather than actions alone, but their success has largely been confined to supervised policy learning. Prior model-based RL methods often optimize the policy or value function directly on imagined rollouts, which is prone to compounding bias and struggles to scale to large, high-dimensional problems such as real-world robotics, a problem that worsens with task horizon and visual complexity. In this work, we instead ask whether we can leverage world models directly on top of standard Q-learning to improve performance, while remaining trained and grounded in the real, online setting. We propose QWM, a framework that leverages world models to perform test-time search over imagined trajectories on top of Q-learning to select high-value actions during both online rollouts and evaluation. Since the policy and value function are trained only on real transitions, QWM avoids compounding model bias while still gaining the sample-efficiency benefits of predictive search. On challenging manipulation benchmarks Robomimic and LIBERO, QWM significantly outperforms strong prior state-of-the-art methods on both sample efficiency and performance.
Perry Dong, Yueru Jia, Chelsea Finn et al.· 0 citations