NeurIPS 2025 Papers

In order to enable fluid and natural human-machine speech interaction, existing full-duplex conversational systems often adopt modular architectures with auxiliary components such as voice activity detectors, interrupters, conversation state predictors, or multiple LLMs. These systems, however, suffer from error accumulation across modules and struggle with key challenges such as context-dependent barge-in and echo cancellation. Recent approaches, most notably Moshi, simplify the pipeline by injecting audio codecs into the token space of a single LLM. However, such methods still incur significant performance degradation when operating on the speech rather than text modality. In this paper, we introduce SALMONN-omni, the first single, standalone full-duplex speech LLM that operates without audio codecs in its token space. It features a novel dynamic thinking mechanism within the LLM backbone, enabling the model to learn when to transition between speaking and listening states. Experiments on widely used benchmarks for spoken question answering and open-domain dialogue show that SALMONN-omni achieves at least 30\% relative performance improvement over existing open-source full-duplex models and performs highly competitively to half-duplex and turn-based systems, despite using substantially less training data. Moreover, SALMONN-omni demonstrates strong performance in complex conversational scenarios, including turn-taking, backchanneling, echo cancellation and context-dependent barge-in, with further improvements achieved through reinforcement learning. Some demo conversations between user and SALMONN-omni are provided in the following repository https://github.com/bytedance/SALMONN.

Risk-averse total-reward Markov Decision Processes (MDPs) offer a promising framework for modeling and solving undiscounted infinite-horizon objectives. Existing model-based algorithms for risk measures like the entropic risk measure (ERM) and entropic value-at-risk (EVaR) are effective in small problems, but require full access to transition probabilities. We propose a Q-learning algorithm to compute the optimal stationary policy for total-reward ERM and EVaR objectives with strong convergence and performance guarantees. The algorithm and its optimality are made possible by ERM's dynamic consistency and elicitability. Our numerical results on tabular domains demonstrate quick and reliable convergence of the proposed Q-learning algorithm to the optimal risk-averse value function.

Off-policy evaluation (OPE) estimates the performance of a target policy using offline data collected from a behavior policy, and is crucial in domains such as robotics or healthcare where direct interaction with the environment is costly or unsafe. Existing OPE methods are ineffective for high-dimensional, long-horizon problems, due to exponential blow-ups in variance from importance weighting or compounding errors from learned dynamics models. To address these challenges, we propose STITCH-OPE, a model-based generative framework that leverages denoising diffusion for long-horizon OPE in high-dimensional state and action spaces. Starting with a diffusion model pre-trained on the behavior data, STITCH-OPE generates synthetic trajectories from the target policy by guiding the denoising process using the score function of the target policy. STITCH-OPE proposes two technical innovations that make it advantageous for OPE: (1) prevents over-regularization by subtracting the score of the behavior policy during guidance, and (2) generates long-horizon trajectories by stitching partial trajectories together end-to-end. We provide a theoretical guarantee that under mild assumptions, these modifications result in an exponential reduction in variance versus long-horizon trajectory diffusion. Experiments on the D4RL and OpenAI Gym benchmarks show substantial improvement in mean squared error, correlation, and regret metrics compared to state-of-the-art OPE methods.

The increasing energy demands and carbon footprint of large-scale AI require intelligent workload management in globally distributed data centers. Yet progress is limited by the absence of benchmarks that realistically capture the interplay of time-varying environmental factors (grid carbon intensity, electricity prices, weather), detailed data center physics (CPUs, GPUs, memory, HVAC energy), and geo-distributed network dynamics (latency and transmission costs). To bridge this gap, we present DCcluster-Opt: an open-source, high-fidelity simulation benchmark for sustainable, geo-temporal task scheduling. DCcluster-Opt combines curated real-world datasets, including AI workload traces, grid carbon intensity, electricity markets, weather across 20 global regions, cloud transmission costs, and empirical network delay parameters with physics-informed models of data center operations, enabling rigorous and reproducible research in sustainable computing. It presents a challenging scheduling problem where a top-level coordinating agent must dynamically reassign or defer tasks that arrive with resource and service-level agreement requirements across a configurable cluster of data centers to optimize multiple objectives. The environment also models advanced components such as heat recovery. A modular reward system enables an explicit study of trade-offs among carbon emissions, energy costs, service level agreements, and water use. It provides a Gymnasium API with baseline controllers, including reinforcement learning and rule-based strategies, to support reproducible ML research and a fair comparison of diverse algorithms. By offering a realistic, configurable, and accessible testbed, DCcluster-Opt accelerates the development and validation of next-generation sustainable computing solutions for geo-distributed data centers.

Visual reasoning in multimodal large language models (MLLMs) has primarily been studied in passive, static settings, limiting their effectiveness in real-world physical environments where an embodied agent must contend with incomplete information due to occlusion or a limited field of view. Humans, in contrast, leverage their embodiment to actively explore and interact with their environment—moving, examining, and manipulating objects—to gather information through a closed-loop process integrating perception, reasoning, and action. Inspired by this capability, we introduce the Active Visual Reasoning (AVR) task, extending visual reasoning to a paradigm of embodied interaction in partially observable environments. AVR necessitates embodied agents to: (1) actively acquire information via sequential physical actions, (2) integrate observations across multiple steps for coherent reasoning, and (3) dynamically adjust decisions based on evolving visual feedback. To rigorously evaluate AVR, we introduce CLEVR-AVR, a simulation benchmark featuring multi-round interactive environments designed to assess both reasoning correctness and information-gathering efficiency. We present AVR-152k, a large-scale dataset that offers rich Chain-of-Thought (CoT) annotations detailing iterative reasoning for uncertainty identification, action-conditioned information gain prediction, and information-maximizing action selection, crucial for training agents in a higher-order Markov Decision Process. Building on this, we develop PhysVLM-AVR, an embodied MLLM achieving state-of-the-art performance on CLEVR-AVR, embodied reasoning (OpenEQA, RoboVQA), and passive visual reasoning (GeoMath, Geometry30K). Our analysis also reveals that current embodied MLLMs, despite detecting information incompleteness, struggle to actively acquire and integrate new information through interaction, highlighting a fundamental gap in active reasoning capabilities.

Large Language Models (LLMs) achieve superior performance through Chain-of-Thought (CoT) reasoning, but these token-level reasoning chains are computationally expensive and inefficient. In this paper, we introduce Compressed Latent Reasoning (CoLaR), a novel framework that dynamically compresses reasoning processes in latent space through a two-stage training approach. First, during supervised fine-tuning, CoLaR extends beyond next-token prediction by incorporating an auxiliary next compressed embedding prediction objective. This process merges embeddings of consecutive tokens using a compression factor $c$ randomly sampled from a predefined range, and trains a specialized latent head to predict distributions of subsequent compressed embeddings. Second, we enhance CoLaR through reinforcement learning (RL) that leverages the latent head's non-deterministic nature to explore diverse reasoning paths and exploit more compact ones. This approach enables CoLaR to: i) **perform reasoning at a dense latent level** (i.e., silently), substantially reducing reasoning chain length, and ii) **dynamically adjust reasoning speed** at inference time by simply prompting the desired compression factor. Extensive experiments across four mathematical reasoning datasets demonstrate that CoLaR achieves 14.1% higher accuracy than latent-based baseline methods at comparable compression ratios, and reduces reasoning chain length by 53.3% with only 4.8% performance degradation compared to explicit CoT method. Moreover, when applied to more challenging mathematical reasoning tasks, our RL-enhanced CoLaR demonstrates performance gains of up to 5.4% while dramatically reducing latent reasoning chain length by 82.8%. The code and models will be released upon acceptance.

The design of biological sequences is essential for engineering functional biomolecules that contribute to advancements in human health and biotechnology. Recent advances in diffusion models, with their generative power and efficient conditional sampling, have made them a promising approach for sequence generation. To enhance model performance on limited data and enable multi-objective design and optimization, reinforcement learning (RL)-based fine-tuning has shown great potential. However, existing post-sampling and fine-tuning methods either lack stability in discrete optimization when avoiding gradients or incur high computational costs when employing gradient-based approaches, creating significant challenges for achieving both control and stability in the tuning process. To address these limitations, we propose GLID$^2$E, a gradient-free RL-based tuning approach for discrete diffusion models. Our method introduces a clipped likelihood constraint to regulate the exploration space and implements reward shaping to better align the generative process with design objectives, ensuring a more stable and efficient tuning process. By integrating these techniques, GLID$^2$E mitigates training instabilities commonly encountered in RL and diffusion-based frameworks, enabling robust optimization even in challenging biological design tasks. In the DNA sequence and protein sequence design systems, GLID$^2$E achieves competitive performance in function-based design while maintaining computational efficiency and a flexible tuning mechanism.

We propose ReinFlow, a simple yet effective online reinforcement learning (RL) framework that fine-tunes a family of flow matching policies for continuous robotic control. Derived from rigorous RL theory, ReinFlow injects learnable noise into a flow policy’s deterministic path, converting the flow into a discrete-time Markov Process for exact and straightforward likelihood computation. This conversion facilitates exploration and ensures training stability, enabling ReinFlow to fine-tune diverse flow model variants stably, including Rectified Flow [34] and Shortcut Models [18], particularly at very few or even one denoising step. We benchmark ReinFlow in representative locomotion and manipulation tasks, including long- horizon planning with visual input and sparse reward. The episode reward of Rectified Flow policies obtained an average net growth of 135.36% after fine-tuning in challenging legged locomotion tasks while saving denoising steps and 82.63% of wall time compared to state-of-the-art diffusion RL fine-tuning method DPPO [42]. The success rate of the Shortcut Model policies in state and visual manipulation tasks achieved an average net increase of 40.34% after fine-tuning with ReinFlow at four or even one denoising step, whose performance is comparable to fine-tuned DDIM policies while saving computation time for an average of 23.20% . Code, model, and checkpoints available on the project website: https://reinflow.github.io/

Accurate spatial reasoning in outdoor environments—covering geometry, object pose, and inter-object relationships—is fundamental to downstream tasks such as mapping, motion forecasting, and high-level planning in autonomous driving. We introduce SURDS, a large-scale benchmark designed to systematically evaluate the spatial reasoning capabilities of vision language models (VLMs). Built on the nuScenes dataset, SURDS comprises 41,080 vision–question–answer training instances and 9,250 evaluation samples, spanning six spatial categories: orientation, depth estimation, pixel-level localization, pairwise distance, lateral ordering, and front–behind relations. We benchmark leading general-purpose VLMs, including GPT, Gemini, and Qwen, revealing persistent limitations in fine-grained spatial understanding. To address these deficiencies, we go beyond static evaluation and explore whether alignment techniques can improve spatial reasoning performance. Specifically, we propose a reinforcement learning–based alignment scheme leveraging spatially grounded reward signals—capturing both perception-level accuracy (location) and reasoning consistency (logic). We further incorporate final-answer correctness and output-format rewards to guide fine-grained policy adaptation. Our GRPO-aligned variant achieves overall score of 40.80 in SURDS benchmark. Notably, it outperforms proprietary systems such as GPT-4o (13.30) and Gemini-2.0-flash (35.71). To our best knowledge, this is the first study to demonstrate that reinforcement learning–based alignment can significantly and consistently enhance the spatial reasoning capabilities of VLMs in real-world driving contexts. We release the SURDS benchmark, evaluation toolkit, and GRPO alignment code through: https://github.com/XiandaGuo/Drive-MLLM.

Online reinforcement learning (RL) with complex function approximations such as transformers and deep neural networks plays a significant role in the modern practice of artificial intelligence. Despite its popularity and importance, balancing the fundamental trade-off between exploration and exploitation remains a long-standing challenge; in particular, we are still in lack of efficient and practical schemes that are backed by theoretical performance guarantees. Motivated by recent developments in exploration via optimistic regularization, this paper provides an interpretation of the principle of optimism through the lens of primal-dual optimization. From this fresh perspective, we set forth a new value-incentivized actor-critic (VAC) method, which optimizes a single easy-to-optimize objective integrating exploration and exploitation --- it promotes state-action and policy estimates that are both consistent with collected data transitions and result in higher value functions. Theoretically, the proposed VAC method has near-optimal regret guarantees under linear Markov decision processes (MDPs) in both finite-horizon and infinite-horizon settings, which can be extended to the general function approximation setting under appropriate assumptions.

Large Language Models (LLMs) are increasingly used to simulate human users in interactive settings such as therapy, education, and social role-play. While these simulations enable scalable training and evaluation of AI agents, off-the-shelf LLMs often drift from their assigned personas, contradict earlier statements, or abandon role-appropriate behavior. We introduce a unified framework for evaluating and improving persona consistency in LLM-generated dialogue. We define three automatic metrics—prompt-to-line consistency, line-to-line consistency, and Q\&A consistency—that capture different types of persona drift and validate each against human annotations. Using these metrics as reward signals, we apply multi-turn reinforcement learning to fine-tune LLMs for three user roles: a patient, a student, and a social chat partner. Our method reduces inconsistency by over 55%, resulting in more coherent, faithful, and trustworthy simulated users.

Recent advances in reinforcement learning (RL) have strengthened the reasoning capabilities of vision-language models (VLMs). However, enhancing policy exploration to better scale test-time compute remains largely underexplored. In addition, VLMs continue to struggle with imperfect visual perception, which in turn affects the subsequent reasoning process. To this end, we propose **NoisyRollout**, a simple yet effective data augmentation method that mixes trajectories from both clean and moderately distorted images during RL training. By injecting targeted diversity in visual perception and the resulting reasoning patterns, NoisyRollout promotes better policy exploration through vision-oriented inductive biases, ultimately leading to more robust reasoning behaviors. We further adopt a noise annealing schedule that gradually reduces distortion strength over training, leveraging noisy signals early on while ensuring training stability in later stages. Crucially, our method is easy-to-adopt—**requiring no additional training cost and no modifications to the RL objective**. Extensive experiments on $2$ distinct training datasets demonstrate that NoisyRollout achieves state-of-the-art performance among open-source RL-tuned models across $5$ out-of-domain reasoning and perception benchmarks. Furthermore, we validate the effectiveness of NoisyRollout across model sizes ($7$B and $32$B) and data scales (from $1$K to $6$K), highlighting its generalizability and scalability.

Large language models are quickly becoming the foundation for intelligent agents that are capable of using tools. However, training such agents is challenging because it requires human creation and annotation of a diverse set of tasks, tools, and evaluation criteria. In this paper, we propose the Self-Challenging Agent framework for training an agent on high-quality tasks that are generated by itself. The agent first plays the role of challenger and generates a task after interacting with the given tools. The tasks take the form of a novel general class of problems termed Code-as-Task, which are defined by an instruction, a verification function and solution and failure cases which serve as tests, allowing to filter only for high-quality tasks. The agent then takes an executor role and trains on those tasks with reinforcement learning using the evaluation feedback as a reward. We show our method improves the performance of Llama-3.1-8B-Instruct on two existing multi-turn tool-use agent benchmarks, M$^3$ToolEval and TauBench, with a two-fold average success rate increase, despite using only self-generated training data.

In the zero-shot policy transfer setting in reinforcement learning, the goal is to train an agent on a fixed set of training environments so that it can generalise to similar, but unseen, testing environments. Previous work has shown that policy distillation after training can sometimes produce a policy that outperforms the original in the testing environments. However, it is not yet entirely clear why that is, or what data should be used to distil the policy. In this paper, we prove, under certain assumptions, a generalisation bound for policy distillation after training. The theory provides two practical insights: for improved generalisation, you should 1) train an ensemble of distilled policies, and 2) distil it on as much data from the training environments as possible. We empirically verify that these insights hold in more general settings, when the assumptions required for the theory no longer hold. Finally, we demonstrate that an ensemble of policies distilled on a diverse dataset can generalise significantly better than the original agent.

Learning representations for solutions of constrained optimization problems (COPs) with unknown cost functions is challenging, as models like (Variational) Autoencoders struggle to enforce constraints when decoding structured outputs. We propose an Inverse Optimization Latent Variable Model (IO-LVM) that learns a latent space of COP cost functions from observed solutions and reconstructs feasible outputs by solving a COP with a solver in the loop. Our approach leverages estimated gradients of a Fenchel-Young loss through a non-differentiable deterministic solver to shape the latent space. Unlike standard Inverse Optimization or Inverse Reinforcement Learning methods, which typically recover a single or context-specific cost function, IO-LVM captures a distribution over cost functions, enabling the identification of diverse solution behaviors arising from different agents or conditions not available during the training process. We validate our method on real-world datasets of ship and taxi routes, as well as paths in synthetic graphs, demonstrating its ability to reconstruct paths and cycles, predict their distributions, and yield interpretable latent representations.

Foundation models have recently shown strong potential as web agents, capable of interpreting high-level instructions and interacting with complex web interfaces. However, existing training paradigms for these agents often rely on predefined task datasets and curated demonstrations, limiting their scalability, adaptability, and capacity for self-improvement. In this work, we introduce *Self-guided hierArchical exploration for Generalist wEb agents* (SAGE), a new training framework designed to support autonomous skill acquisition through self-guided hierarchical exploration. Our method introduces a three-tier exploration strategy: a pre-exploration phase to build structural understanding of web environments, a top-level exploration strategy to generate a self-evolving curriculum of tasks from easy to hard, and a low-level exploration mechanism that combines planning-based rollouts with step-wise learning to improve policy efficiency. Together, these components form a scalable, supervision-free framework for web agent training. Experimental results on WebVoyager and WebArena demonstrate that our method significantly outperforms prior approaches, enabling foundation model agents to learn complex web tasks with greater generalization and robustness.

Offline meta-RL usually tackles generalization by inferring task beliefs from high-quality samples or warmup explorations. The restricted form limits their generality and usability since these supervision signals are expensive and even infeasible to acquire in advance for unseen tasks. Learning directly from the raw text about decision tasks is a promising alternative to leverage a much broader source of supervision. In the paper, we propose **T**ext-to-**D**ecision **A**gent (**T2DA**), a simple and scalable framework that supervises offline meta-RL with natural language. We first introduce a generalized world model to encode multi-task decision data into a dynamics-aware embedding space. Then, inspired by CLIP, we predict which textual description goes with which decision embedding, effectively bridging their semantic gap via contrastive language-decision pre-training and aligning the text embeddings to comprehend the environment dynamics. After training the text-conditioned generalist policy, the agent can directly realize zero-shot text-to-decision generation in response to language instructions. Comprehensive experiments on MuJoCo and Meta-World benchmarks show that T2DA facilitates high-capacity zero-shot generalization and outperforms various types of baselines. Our code is available at [https://github.com/NJU-RL/T2DA](https://github.com/NJU-RL/T2DA).

Enhancing the reasoning capabilities of large language models effectively using reinforcement learning (RL) remains a crucial challenge. Existing approaches primarily adopt two contrasting advantage estimation granularities: token-level methods (e.g., PPO) aim to provide fine-grained advantage signals but suffer from inaccurate estimation due to difficulties in training an accurate critic model. On the other extreme, trajectory-level methods (e.g., GRPO) solely rely on a coarse-grained advantage signal from the final reward, leading to imprecise credit assignment. To address these limitations, we propose Segment Policy Optimization (SPO), a novel RL framework that leverages segment-level advantage estimation at an intermediate granularity, achieving a better balance by offering more precise credit assignment than trajectory-level methods and requiring fewer estimation points than token-level methods, enabling accurate advantage estimation based on Monte Carlo (MC) without a critic model. SPO features three components with novel strategies: (1) flexible segment partition; (2) accurate segment advantage estimation; and (3) policy optimization using segment advantages, including a novel probability-mask strategy. We further instantiate SPO for two specific scenarios: (1) SPO-chain for short chain-of-thought (CoT), featuring novel cutpoint-based partition and chain-based advantage estimation, achieving $6$-$12$ percentage point improvements in accuracy over PPO and GRPO on GSM8K. (2) SPO-tree for long CoT, featuring novel tree-based advantage estimation, which significantly reduces the cost of MC estimation, achieving $7$-$11$ percentage point improvements over GRPO on MATH500 under 2K and 4K context evaluation. We make our code publicly available at https://github.com/AIFrameResearch/SPO.

Machine Learning models in real-world applications must continuously learn new tasks to adapt to shifts in the data-generating distribution. Yet, for Continual Learning (CL), models often struggle to balance learning new tasks (plasticity) with retaining previous knowledge (memory stability). Consequently, they are susceptible to Catastrophic Forgetting, which degrades performance and undermines the reliability of deployed systems. In the Bayesian CL literature, variational methods tackle this challenge by employing a learning objective that recursively updates the posterior distribution while constraining it to stay close to its previous estimate. Nonetheless, we argue that these methods may be ineffective due to compounding approximation errors over successive recursions. To mitigate this, we propose new learning objectives that integrate the regularization effects of multiple previous posterior estimations, preventing individual errors from dominating future posterior updates and compounding over time. We reveal insightful connections between these objectives and Temporal-Difference methods, a popular learning mechanism in Reinforcement Learning and Neuroscience. Experiments on challenging CL benchmarks show that our approach effectively mitigates Catastrophic Forgetting, outperforming strong Variational CL methods.