SAPO: Single-Rollout Autoregressive Policy Optimization for Agentic Reinforcement Learning

Agentic reinforcement learning (RL) has become a critical stage in the post-training of large language models. Existing critic-free, group-relative methods estimate policy advantages from multiple rollouts, avoiding the substantial memory overhead of conventional proximal policy optimization (PPO) and achieving strong performance on long-horizon interactive tasks. Despite their success, recent studies revealed three limitations: (1) Lack explicit value generalization and effective temporal credit assignment; (2) Suffer from potential advantage collapse in long-horizon complex tasks; (3) Require a costly trade-off between sampling budget and policy performance. In this work, we propose Single-rollout Autoregressive Policy Optimization (SAPO), a low-memory and compute-efficient framework in which the policy and value functions share a single autoregressive backbone. SAPO exploits the autoregressive structure of LLMs to produce policy and value predictions at distinct causal boundaries with shared parameters, while independently optimizing the PPO objectives and auxiliary on-policy SARSA objectives. To robustly estimate the contribution of each turn, we further introduce a trajectory-level generalized advantage estimator that combines lambda-returns with batch normalization. Experiments across ALFWorld and WebShop with Qwen2.5-1.5B/7B show that SAPO trains stably and outperforms PPO and GRPO by mean +15.1 and +12.1 percentage points, respectively, while eliminating the memory cost of a separate critic model and reducing per-iteration runtime by 33.2% over PPO.

PlanPO: Group Planning-Aware Policy Optimization for Multi-Turn Agentic LLMs

Group-relative policy optimization has emerged as a key paradigm for training agentic large language models (LLMs) on multi-turn interactive tasks. However, most existing variants fail to distinguish advantages among successful trajectories even when these trajectories differ substantially in their interaction efficiency. For instance, circuitous successes are often assigned the identical outcome reward, causing advantage collapse and severe performance bottlenecks. To this end, we propose Group Planning-aware Policy Optimization (PlanPO), a simple yet effective RL method for learning generalizable planning abilities beyond task-specific high-quality behavior patterns. Specifically, PlanPO introduces coarse-to-fine advantage signals, which capture the relative differences in trajectory-level lengths and turn-level response lengths conditioned on successful trajectories sampled for the same task. Within the group-relative optimization structure, this enables agents to actively learn generalizable and deliberate behaviors spanning interaction planning and textual generation from high-quality rollouts, without degenerating into vanilla length minimization. Experimentally, PlanPO improves over GRPO by 27.2\% on average across the challenging multi-turn benchmarks ALFWorld, WebShop, and SciWorld, outperforming recent powerful baselines while incurring negligible additional training cost.

Learning Task-relevant Sequence Representations via Intrinsic Dynamics Characteristics in Reinforcement Learning

Learning task-relevant state representations is crucial to solving the problem of scene generalization in visual deep reinforcement learning. Prior work typically establishes a self-supervised auxiliary learner, introducing elements (e.g., rewards and actions) to extract task-relevant state information from observations through behavioral similarity metrics. However, the methods often ignore the inherent relationships between the elements (e.g., dynamics relationships) that are essential for learning accurate representations, and they are also limited to single-step metrics, which impedes the discrimination of short-term similar task/behavior information in long-term dynamics transitions. To solve the issues, we propose an intrinsic dynamic characteristics-driven sequence representation learning method (DSR) over a common DRL frame. Concretely, inspired by the fact of state transition in the underlying system, it constrains the optimization of the encoder via modeling the dynamics equations related to the state transition, which prompts the latent encoding information to satisfy the state transition process and thereby distinguishes state space and noise space. Further, to refine the ability of encoding similar tasks based on dynamics constraints, DSR also sequentially models inherent dynamics equation relationships from the perspective of sequence elements' frequency domain and multi-step prediction. Finally, experimental results show that DSR has achieved a significant performance boost in the Distracting DMControl Benchmark, with an average of 78.9% over the backbone baseline. Further results indicate that it also achieves the best performance in real-world autonomous driving tasks in the CARLA simulator. Moreover, the qualitative analysis results of t-SNE visualization validate that our method possesses superior representation ability on visual tasks.

Episodic Reinforcement Learning with Expanded State-reward Space

Empowered by deep neural networks, deep reinforcement learning (DRL) has demonstrated tremendous empirical successes in various domains, including games, health care, and autonomous driving. Despite these advancements, DRL is still identified as data-inefficient as effective policies demand vast numbers of environmental samples. Recently, episodic control (EC)-based model-free DRL methods enable sample efficiency by recalling past experiences from episodic memory. However, existing EC-based methods suffer from the limitation of potential misalignment between the state and reward spaces for neglecting the utilization of (past) retrieval states with extensive information, which probably causes inaccurate value estimation and degraded policy performance. To tackle this issue, we introduce an efficient EC-based DRL framework with expanded state-reward space, where the expanded states used as the input and the expanded rewards used in the training both contain historical and current information. To be specific, we reuse the historical states retrieved by EC as part of the input states and integrate the retrieved MC-returns into the immediate reward in each interactive transition. As a result, our method is able to simultaneously achieve the full utilization of retrieval information and the better evaluation of state values by a Temporal Difference (TD) loss. Empirical results on challenging Box2d and Mujoco tasks demonstrate the superiority of our method over a recent sibling method and common baselines. Further, we also verify our method's effectiveness in alleviating Q-value overestimation by additional experiments of Q-value comparison.

Sequential Action-Induced Invariant Representation for Reinforcement Learning

How to accurately learn task-relevant state representations from high-dimensional observations with visual distractions is a realistic and challenging problem in visual reinforcement learning. Recently, unsupervised representation learning methods based on bisimulation metrics, contrast, prediction, and reconstruction have shown the ability for task-relevant information extraction. However, due to the lack of appropriate mechanisms for the extraction of task information in the prediction, contrast, and reconstruction-related approaches and the limitations of bisimulation-related methods in domains with sparse rewards, it is still difficult for these methods to be effectively extended to environments with distractions. To alleviate these problems, in the paper, the action sequences, which contain task-intensive signals, are incorporated into representation learning. Specifically, we propose a Sequential Action--induced invariant Representation (SAR) method, in which the encoder is optimized by an auxiliary learner to only preserve the components that follow the control signals of sequential actions, so the agent can be induced to learn the robust representation against distractions. We conduct extensive experiments on the DeepMind Control suite tasks with distractions while achieving the best performance over strong baselines. We also demonstrate the effectiveness of our method at disregarding task-irrelevant information by deploying SAR to real-world CARLA-based autonomous driving with natural distractions. Finally, we provide the analysis results of generalization drawn from the generalization decay and t-SNE visualization. Code and demo videos are available at https://github.com/DMU-XMU/SAR.git.