NeurIPS 2025 论文列表

Recent efforts to employ large language models (LLMs) as agents have demonstrated promising results in a wide range of multi-step agent tasks. However, existing agents lack an effective experience reuse approach to leverage historical completed tasks. In this paper, we propose a novel experience reuse framework MetaFlowLLM, which constructs a hierarchical experience tree from historically completed tasks. Each node in this experience tree is presented as a MetaFlow which contains static execution workflow and subtask required by agents to complete dynamically. Then, we propose a Hierarchical MetaFlow Merging algorithm to construct the hierarchical experience tree. When accomplishing a new task, MetaFlowLLM can first retrieve the most relevant MetaFlow node from the experience tree and then execute it accordingly. To effectively generate valid MetaFlows from historical data, we further propose a reinforcement learning pipeline to train the MetaFlowGen. Extensive experimental results on AppWorld and WorkBench demonstrate that integrating with MetaFlowLLM, existing agents (e.g., ReAct, Reflexion) can gain substantial performance improvement with reducing execution costs. Notably, MetaFlowLLM achieves an average success rate improvement of 32.3% on AppWorld and 6.2% on WorkBench, respectively.

Continuous time systems are often modeled using discrete time dynamics but this requires a small simulation step to maintain accuracy. In turn, this requires a large planning horizon which leads to computationally demanding planning problems and reduced performance. Previous work in model-free reinforcement learning has partially addressed this issue using action repeats where a policy is learned to determine a discrete action duration. Instead we propose to control the continuous decision timescale directly by using temporally-extended actions and letting the planner treat the duration of the action as an additional optimization variable along with the standard action variables. This additional structure has multiple advantages. It speeds up simulation time of trajectories and, importantly, it allows for deep horizon search in terms of primitive actions while using a shallow search depth in the planner. In addition, in the model-based reinforcement learning (MBRL) setting, it reduces compounding errors from model learning and improves training time for models. We show that this idea is effective and that the range for action durations can be automatically selected using a multi-armed bandit formulation and integrated into the MBRL framework. An extensive experimental evaluation both in planning and in MBRL, shows that our approach yields faster planning, better solutions, and that it enables solutions to problems that are not solved in the standard formulation.

In this work, we introduce CAD-Coder, a novel framework that reformulates text-to-CAD as the generation of CadQuery scripts—a Python-based, parametric CAD language. This representation enables direct geometric validation, a richer modeling vocabulary, and seamless integration with existing LLMs. To further enhance code validity and geometric fidelity, we propose a two-stage learning pipeline: (1) supervised fine-tuning on paired text–CadQuery data, and (2) reinforcement learning with Group Reward Policy Optimization (GRPO), guided by a CAD-specific reward comprising both a geometric reward (Chamfer Distance) and a format reward. We also introduce a chain-of-thought (CoT) planning process to improve model reasoning, and construct a large-scale, high-quality dataset of 110K text–CadQuery–3D model triplets and 1.5K CoT samples via an automated pipeline. Extensive experiments demonstrate that CAD-Coder enables LLMs to generate diverse, valid, and complex CAD models directly from natural language, advancing the state of the art of text-to-CAD generation and geometric reasoning.

Reinforcement learning (RL) is increasingly used to align large language models (LLMs). Off-policy methods offer greater implementation simplicity and data efficiency than on-policy techniques, but often result in suboptimal performance. In this work, we study the intermediate range of algorithms between off-policy RL and supervised fine-tuning by analyzing a simple off-policy REINFORCE algorithm, where the advantage is defined as $A=r-V$, with $r$ a reward and $V$ some tunable baseline. Intuitively, lowering $V$ emphasizes high-reward samples, while raising it penalizes low-reward ones more heavily. We first provide a theoretical analysis of this off-policy REINFORCE algorithm, showing that when the baseline $V$ lower-bounds the expected reward, the algorithm enjoys a policy improvement guarantee. Our analysis reveals that while on-policy updates can safely leverage both positive and negative signals, off-policy updates benefit from focusing more on positive rewards than on negative ones. We validate our findings experimentally in a controlled stochastic bandit setting and through fine-tuning state-of-the-art LLMs on reasoning tasks.

Vision-Language-Action (VLA) models have demonstrated strong multi-modal reasoning capabilities, enabling direct action generation from visual perception and language instructions in an end-to-end manner. However, their substantial computational cost poses a challenge for real-time robotic control, where rapid decision-making is essential. This paper introduces VLA-Cache, a training-free inference acceleration method that reduces computational overhead by adaptively caching and reusing static visual tokens across frames. Exploiting the temporal continuity in robotic manipulation, VLA-Cache identifies minimally changed tokens between adjacent frames and reuses their cached key-value representations, thereby circumventing redundant computations. Additionally, to maintain action precision, VLA-Cache selectively re-computes task-relevant tokens that are environmentally sensitive, ensuring the fidelity of critical visual information. To further optimize efficiency, we introduce a layer adaptive token reusing strategy that dynamically adjusts the reuse ratio based on attention concentration across decoder layers, prioritizing critical tokens for recomputation. Extensive experiments on two simulation platforms (LIBERO and SIMPLER) and a real-world robotic system demonstrate that VLA-Cache achieves up to 1.7× speedup in CUDA latency and a 15\% increase in control frequency, with negligible loss on task success rate. The code and videos can be found at our project page: https://vla-cache.github.io.

Recent advances in group-based reinforcement learning (RL) have driven frontier large language models (LLMs) in single-turn tasks like mathematical reasoning. However, their scalability to multi-turn LLM agent training remains limited. Unlike static tasks, agent-environment interactions unfold over many steps and often yield sparse or delayed rewards, making credit assignment across individual steps significantly more challenging. In this work, we propose Group-in-Group Policy Optimization (GiGPO), a novel RL algorithm that achieves fine-grained credit assignment for LLM agents while preserving the appealing properties of group-based RL: critic-free, low memory, and stable convergence. GiGPO introduces a two-level structure for estimating relative advantage: (i) At the episode-level, GiGPO computes macro relative advantages based on groups of complete trajectories; (ii) At the step-level, GiGPO introduces an anchor state grouping mechanism that retroactively constructs step-level groups by identifying repeated environment states across trajectories. Actions stemming from the same state are grouped together, enabling micro relative advantage estimation. This hierarchical structure effectively captures both global trajectory quality and local step effectiveness without relying on auxiliary models or additional rollouts. We evaluate GiGPO on challenging agent benchmarks, including ALFWorld and WebShop, as well as tool-integrated reasoning on search-augmented QA tasks, using Qwen2.5-1.5B/3B/7B-Instruct. Crucially, GiGPO delivers fine-grained per-step credit signals, achieves performance gains of > 12\% on ALFWorld and > 9\% on WebShop over GRPO, and obtains superior performance on QA tasks (42.1\% on 3B and 47.2\% on 7B): all while maintaining the same GPU memory overhead, identical LLM rollout, and incurring little to no additional time cost.

Reinforcement learning (RL) for collaborative agents capable of cooperating with humans to accomplish tasks has long been a central goal in the RL community. While prior approaches have made progress in adapting collaborative agents to diverse human partners, they often focus solely on optimizing task performance and overlook human preferences—despite the fact that such preferences often diverge from the reward-maximization objective of the environment. Addressing this discrepancy poses significant challenges: humans typically provide only a small amount of offline, preference-related feedback and are unable to engage in online interactions, resulting in a distributional mismatch between the agent’s online learning process and the offline human data. To tackle this, we formulate the problem as an online&offline reinforcement learning problem that jointly integrates online generalization and offline preference learning, entirely under an offline training regime. We propose a simple yet effective training framework built upon existing RL algorithms that alternates between offline preference learning and online generalization recovery, ensuring the stability and alignment of both learning objectives. We evaluate our approach on a benchmark built upon the Overcooked environment—a standard environment for human-agent collaboration—and demonstrate remarkable performance across diverse preference styles and cooperative scenarios.

Reinforcement Learning enables agents to learn optimal behaviors through interactions with environments. However, real-world environments are typically non-stationary, requiring agents to continuously adapt to new tasks and changing conditions. Although Continual Reinforcement Learning facilitates learning across multiple tasks, existing methods often suffer from catastrophic forgetting and inefficient knowledge utilization. To address these challenges, we propose Continual Knowledge Adaptation for Reinforcement Learning (CKA-RL), which enables the accumulation and effective utilization of historical knowledge. Specifically, we introduce a Continual Knowledge Adaptation strategy, which involves maintaining a task-specific knowledge vector pool and dynamically using historical knowledge to adapt the agent to new tasks. This process mitigates catastrophic forgetting and enables efficient knowledge transfer across tasks by preserving and adapting critical model parameters. Additionally, we propose an Adaptive Knowledge Merging mechanism that combines similar knowledge vectors to address scalability challenges, reducing memory requirements while ensuring the retention of essential knowledge. Experiments on three benchmarks demonstrate that the proposed CKA-RL outperforms state-of-the-art methods, achieving an improvement of 4.20% in overall performance and 8.02% in forward transfer. The source code is available at https://github.com/Fhujinwu/CKA-RL.

This paper provides the first expert sample complexity characterization for learning a Nash equilibrium from expert data in Markov Games. We show that a new quantity named the *single policy deviation concentrability coefficient* is unavoidable in the non-interactive imitation learning setting, and we provide an upper bound for behavioral cloning (BC) featuring such coefficient. BC exhibits substantial regret in games with high concentrability coefficient, leading us to utilize expert queries to develop and introduce two novel solution algorithms: MAIL-BRO and MURMAIL. The former employs a best response oracle and learns an $\varepsilon$-Nash equilibrium with $\mathcal{O}(\varepsilon^{-4})$ expert and oracle queries. The latter bypasses completely the best response oracle at the cost of a worse expert query complexity of order $\mathcal{O}(\varepsilon^{-8})$. Finally, we provide numerical evidence, confirming our theoretical findings.

While visual autoregressive modeling (VAR) strategies have shed light on image generation with the autoregressive models, their potential for segmentation, a task that requires precise low-level spatial perception, remains unexplored. Inspired by the multi-scale modeling of classic Mask2Former-based models, we propose Seg-VAR, a novel framework that rethinks segmentation as a conditional autoregressive mask generation problem. This is achieved by replacing the discriminative learning with the latent learning process. Specifically, our method incorporates three core components: (1) an image encoder generating latent priors from input images, (2) a spatial-aware seglat (a latent expression of segmentation mask) encoder that maps segmentation masks into discrete latent tokens using a location-sensitive color mapping to distinguish instances, and (3) a decoder reconstructing masks from these latents. A multi-stage training strategy is introduced: first learning seglat representations via image-seglat joint training, then refining latent transformations, and finally aligning image-encoder-derived latents with seglat distributions. Experiments show Seg-VAR outperforms previous discriminative and generative methods on various segmentation tasks and validation benchmarks. By framing segmentation as a sequential hierarchical prediction task, Seg-VAR opens new avenues for integrating autoregressive reasoning into spatial-aware vision systems.

Understanding the behavior of deep reinforcement learning (DRL) agents—particularly as task and agent sophistication increase—requires more than simple comparison of reward curves, yet standard methods for behavioral analysis remain underdeveloped in DRL. We apply tools from neuroscience and ethology to study DRL agents in a novel, complex, partially observable environment, ForageWorld, designed to capture key aspects of real-world animal foraging—including sparse, depleting resource patches, predator threats, and spatially extended arenas. We use this environment as a platform for applying joint behavioral and neural analysis to agents, revealing detailed, quantitatively grounded insights into agent strategies, memory, and planning. Contrary to common assumptions, we find that model-free RNN-based DRL agents can exhibit structured, planning-like behavior purely through emergent dynamics—without requiring explicit memory modules or world models. Our results show that studying DRL agents like animals—analyzing them with neuroethology-inspired tools that reveal structure in both behavior and neural dynamics—uncovers rich structure in their learning dynamics that would otherwise remain invisible. We distill these tools into a general analysis framework linking core behavioral and representational features to diagnostic methods, which can be reused for a wide range of tasks and agents. As agents grow more complex and autonomous, bridging neuroscience, cognitive science, and AI will be essential—not just for understanding their behavior, but for ensuring safe alignment and maximizing desirable behaviors that are hard to measure via reward. We show how this can be done by drawing on lessons from how biological intelligence is studied.

Learning from self-sampled data and sparse environmental feedback remains a fundamental challenge in training self-evolving agents. Temporal credit assignment mitigates this issue by transforming sparse feedback into dense supervision signals. However, previous approaches typically depend on domain-specific value functions for credit assignment, which suffer from poor sample efficiency and limited generalization. In this work, we propose to leverage pre-trained knowledge from large language models (LLMs) to transform sparse rewards into dense training signals (i.e., the advantage function) through retrospective in-context learning (RICL). We further propose an online learning framework, RICOL, which iteratively refines the policy based on the credit assignment results from RICL. We empirically demonstrate that RICL can accurately estimate the advantage function with limited samples and effectively identify critical states for temporal credit assignment. Extended evaluation on the BabyAI benchmark shows that RICOL significantly improves sample efficiency compared to traditional online RL algorithms while achieving performance comparable to imitation learning from expert demonstartions. Our findings highlight the potential of leveraging LLMs for temporal credit assignment, paving the way for more sample-efficient and generalizable RL paradigms.

We develop hybrid memory architectures for general-purpose sequence processing neural networks, that combine key-value memory using softmax attention (KV-memory) with fast weight memory through dynamic synaptic modulation (FW-memory)---the core principles of quadratic and linear transformers, respectively. These two memory systems have complementary but individually limited properties: KV-memory offers precise retrieval but is constrained by quadratic complexity in sequence length, while FW-memory supports arbitrarily long sequences and enables more expressive computation but sacrifices precise recall. We propose and compare three methods to blend these two systems into a single memory system, differing in how and when input information is delivered to each system, to leverage the strengths of both. We conduct experiments on general language modeling and retrieval tasks by training 340M- and 1.3B-parameter models from scratch, as well as on synthetic algorithmic tasks designed to precisely illustrate the benefits of certain hybrid methods over others. We also evaluate our hybrid memory systems on reinforcement learning in partially observable environments. Overall, we demonstrate how a well-designed hybrid can overcome the limitations of its individual components, offering new insights into the design principle of neural memory systems.

Deploying reinforcement learning (RL) safely in the real world is challenging, as policies trained in simulators must face the inevitable *sim-to-real gap*. Robust safe RL techniques are provably safe, however difficult to scale, while domain randomization is more practical yet prone to unsafe behaviors. We address this gap by proposing SPiDR, short for Sim-to-real via Pessimistic Domain Randomization—a scalable algorithm with provable guarantees for safe sim-to-real transfer. SPiDR uses domain randomization to incorporate the uncertainty about the sim-to-real gap into the safety constraints, making it versatile and highly compatible with existing training pipelines. Through extensive experiments on sim-to-sim benchmarks and two distinct real-world robotic platforms, we demonstrate that SPiDR effectively ensures safety despite the sim-to-real gap while maintaining strong performance.

Dexterous robotic hands are essential for performing complex manipulation tasks, yet remain difficult to train due to the challenges of demonstration collection and high-dimensional control. While reinforcement learning (RL) can alleviate the data bottleneck by generating experience in simulation, it typically relies on carefully designed, task-specific reward functions, which hinder scalability and generalization. Thus, contemporary works in dexterous manipulation have often bootstrapped from reference trajectories. These trajectories specify target hand poses that guide the exploration of RL policies and object poses that enable dense, task-agnostic rewards. However, sourcing suitable trajectories---particularly for dexterous hands---remains a significant challenge. Yet, the precise details in explicit reference trajectories are often unnecessary, as RL ultimately refines the motion. Our key insight is that modern vision-language models (VLMs) already encode the commonsense spatial and semantic knowledge needed to specify tasks and guide exploration effectively. Given a task description (e.g., “open the cabinet”) and a visual scene, our method uses an off-the-shelf VLM to first identify task-relevant keypoints (e.g., handles, buttons) and then synthesize 3D trajectories for hand motion and object motion. Subsequently, we train a low-level residual RL policy in simulation to track these coarse trajectories or ``scaffolds'' with high fidelity. Across a number of simulated tasks involving articulated objects and semantic understanding, we demonstrate that our method is able to learn robust dexterous manipulation policies. Moreover, we showcase that our method transfers to real-world robotic hands without any human demonstrations or handcrafted rewards.

Existing imitation learning methods decouple perception and action, which overlooks the causal reciprocity between sensory representations and action execution that humans naturally leverage for adaptive behaviors. To bridge this gap, we introduce Action-Guided Diffusion Policy (DP-AG), a unified representation learning that explicitly models a dynamic interplay between perception and action through probabilistic latent dynamics. DP-AG encodes latent observations into a Gaussian posterior via variational inference and evolves them using an action-guided SDE, where the Vector–Jacobian Product (VJP) of the diffusion policy's noise predictions serves as a structured stochastic force driving latent updates. To promote bidirectional learning between perception and action, we introduce a cycle-consistent contrastive loss that organizes the gradient flow of the noise predictor into a coherent perception–action loop, enforcing mutually consistent transitions in both latent updates and action refinements. Theoretically, we derive a variational lower bound for the action-guided SDE, and prove that the contrastive objective enhances continuity in both latent and action trajectories. Empirically, DP-AG significantly outperforms state-of-the-art methods across simulation benchmarks and real-world UR5 manipulation tasks. As a result, our DP-AG offers a promising step toward bridging biological adaptability and artificial policy learning. Code is available on our project website: https://jingwang18.github.io/dp-ag.github.io/.

World models allow agents to simulate the consequences of actions in imagined environments for planning, control, and long-horizon decision-making. However, existing autoregressive world models struggle with visually coherent predictions due to disrupted spatial structure, inefficient decoding, and inadequate motion modeling. In response, we propose Scale-wise Autoregression with Motion PrOmpt (SAMPO), a hybrid framework that combines visual autoregressive modeling for intra-frame generation with causal modeling for next-frame generation. Specifically, SAMPO integrates temporal causal decoding with bidirectional spatial attention, which preserves spatial locality and supports parallel decoding within each scale. This design significantly enhances both temporal consistency and rollout efficiency. To further improve dynamic scene understanding, we devise an asymmetric multi-scale tokenizer that preserves spatial details in observed frames and extracts compact dynamic representations for future frames, optimizing both memory usage and model performance. Additionally, we introduce a trajectory-aware motion prompt module that injects spatiotemporal cues about object and robot trajectories, focusing attention on dynamic regions and improving temporal consistency and physical realism. Extensive experiments show that SAMPO achieves competitive performance in action-conditioned video prediction and model-based control, improving generation quality with 4.4× faster inference. We also evaluate SAMPO's zero-shot generalization and scaling behavior, demonstrating its ability to generalize to unseen tasks and benefit from larger model sizes.

Approaches for goal-conditioned reinforcement learning (GCRL) often use learned state representations to extract goal-reaching policies. Two frameworks for representation structure have yielded particularly effective GCRL algorithms: (1) *contrastive representations*, in which methods learn "successor features" with a contrastive objective that performs inference over future outcomes, and (2) *temporal distances*, which link the (quasimetric) distance in representation space to the transit time from states to goals. We propose an approach that unifies these two frameworks, using the structure of a quasimetric representation space (triangle inequality) with the right additional constraints to learn successor representations that enable optimal goal-reaching. Unlike past work, our approach is able to exploit a **quasimetric** distance parameterization to learn **optimal** goal-reaching distances, even with **suboptimal** data and in **stochastic** environments. This gives us the best of both worlds: we retain the stability and long-horizon capabilities of Monte Carlo contrastive RL methods, while getting the free stitching capabilities of quasimetric network parameterizations. On existing offline GCRL benchmarks, our representation learning objective improves performance on stitching tasks where methods based on contrastive learning struggle, and on noisy, high-dimensional environments where methods based on quasimetric networks struggle.

Signal Temporal Logic (STL) is a powerful specification language for describing complex temporal behaviors of continuous signals, making it well-suited for high-level robotic task descriptions. However, generating executable plans for STL tasks is challenging, as it requires consideration of the coupling between the task specification and the system dynamics. Existing approaches either follow a model-based setting that explicitly requires knowledge of the system dynamics or adopt a task-oriented data-driven approach to learn plans for specific tasks. In this work, we address the problem of generating executable STL plans for systems with unknown dynamics. We propose a hierarchical planning framework that enables zero-shot generalization to new STL tasks by leveraging only task-agnostic trajectory data during offline training. The framework consists of three key components: (i) decomposing the STL specification into several progresses and time constraints, (ii) searching for timed waypoints that satisfy all progresses under time constraints, and (iii) generating trajectory segments using a pre-trained diffusion model and stitching them into complete trajectories. We formally prove that our method guarantees STL satisfaction, and simulation results demonstrate its effectiveness in generating dynamically feasible trajectories across diverse long-horizon STL tasks. Project Page: https://cps-sjtu.github.io/Zero-Shot-STL/