NeurIPS 2025 论文列表

The challenge of learning with noisy labels is significant in machine learning, as it can severely degrade the performance of prediction models if not addressed properly. This paper introduces a novel framework that conceptualizes noisy label correction as a reinforcement learning (RL) problem. The proposed approach, Reinforcement Learning for Noisy Label Correction (RLNLC), defines a comprehensive state space representing data and their associated labels, an action space that indicates possible label corrections, and a reward mechanism that evaluates the efficacy of label corrections. RLNLC learns a deep feature representation based policy network to perform label correction through reinforcement learning, utilizing an actor-critic method. The learned policy is subsequently deployed to iteratively correct noisy training labels and facilitate the training of the prediction model. The effectiveness of RLNLC is demonstrated through extensive experiments on multiple benchmark datasets, where it consistently outperforms existing state-of-the-art techniques for learning with noisy labels.

Reinforcement learning (RL) has become an effective approach for fine-tuning large language models (LLMs), particularly to enhance their reasoning capabilities. However, RL fine-tuning remains highly resource-intensive, and existing work has largely overlooked the problem of data efficiency. In this paper, we propose two techniques to improve data efficiency in LLM RL fine-tuning: difficulty-targeted online data selection and rollout replay. We introduce the notion of adaptive difficulty to guide online data selection, prioritizing questions of moderate difficulty that are more likely to yield informative learning signals. To estimate adaptive difficulty efficiently, we develop an attention-based framework that requires rollouts for only a small reference set of questions. The adaptive difficulty of the remaining questions is then estimated based on their similarity to this set. To further reduce rollout cost, we introduce a rollout replay mechanism inspired by experience replay in traditional RL. This technique reuses recent rollouts, lowering per-step computation while maintaining stable updates. Experiments across 6 LLM-dataset combinations show that our method reduces RL fine-tuning time by 23% to 62% while reaching the same level of performance as the original GRPO algorithm. Our code repository is available at https://github.com/ASTRAL-Group/data-efficient-llm-rl/.

Post-training has demonstrated its importance in enhancing the reasoning capabilities of large language models (LLMs). The primary post-training methods can be categorized into supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT). SFT is efficient and well-suited for small language models, but it may lead to overfitting and limit the reasoning abilities of larger models. In contrast, RFT generally yields better generalization but depends heavily on the strength of the base model. To address the limitations of SFT and RFT, we propose Unified Fine-Tuning (UFT), a novel post-training paradigm that unifies SFT and RFT into a single, integrated process. UFT enables the model to effectively explore solutions while incorporating informative supervision signals, bridging the gap between memorizing and thinking underlying existing methods. Notably, UFT outperforms both SFT and RFT in general, regardless of model sizes. Furthermore, we theoretically prove that UFT breaks RFT's inherent exponential sample complexity bottleneck, showing for the first time that unified training can exponentially accelerate convergence on long-horizon reasoning tasks.

Recent research on Reasoning of Large Language Models (LLMs) has sought to further enhance their performance by integrating meta-thinking—enabling models to monitor, evaluate, and control their reasoning processes for more adaptive and effective problem-solving. However, current single-agent work lacks a specialized design for acquiring meta-thinking, resulting in low efficacy. To address this challenge, we introduce Reinforced Meta-thinking Agents (ReMA), a novel framework that leverages Multi-Agent Reinforcement Learning (MARL) to elicit meta-thinking behaviors, encouraging LLMs to think about thinking. ReMA decouples the reasoning process into two hierarchical agents: a high-level meta-thinking agent responsible for generating strategic oversight and plans, and a low-level reasoning agent for detailed executions. Through iterative reinforcement learning with aligned objectives, these agents explore and learn collaboration, leading to improved generalization and robustness. Empirical results from single-turn experiments demonstrate that ReMA outperforms single-agent RL baselines on complex reasoning tasks, including competitive-level mathematical benchmarks and LLM-as-a-Judge benchmarks. Additionally, we further extend ReMA to multi-turn interaction settings, leveraging turn-level ratio and parameter sharing to improve efficiency. Comprehensive ablation studies further illustrate the evolving dynamics of each distinct agent, providing valuable insights into how the meta-thinking reasoning process enhances the reasoning capabilities of LLMs.

Multi-modal Large Language Models (MLLMs) have shown remarkable capabilities across a wide range of vision-language tasks. However, due to the restricted input resolutions, MLLMs face significant challenges in precisely understanding and localizing visual details in high-resolution images---particularly when dealing with extra-small objects embedded in cluttered contexts. To address this issue, we propose FineRS, a two-stage MLLM-based reinforcement learning framework for jointly reasoning and segmenting extremely small objects within high-resolution scenes. FineRS adopts a coarse-to-fine pipeline comprising Global Semantic Exploration (GSE) and Localized Perceptual Refinement (LPR). Specifically, GSE performs instruction-guided reasoning to generate a textural response and a coarse target region, while LPR refines this region to produce an accurate bounding box and segmentation mask. To couple the two stages, we introduce a locate-informed retrospective reward, where LPR's outputs are used to optimize GSE for more robust coarse region exploration. Additionally, we present FineRS-4k, a new dataset for evaluating MLLMs on attribute-level reasoning and pixel-level segmentation on subtle, small-scale targets in complex high-resolution scenes. Experimental results on FineRS-4k and public datasets demonstrate that our method consistently outperforms state-of-the-art MLLM-based approaches on both instruction-guided segmentation and visual reasoning tasks.

Test-time scaling in agentic tasks often relies on generating long reasoning traces ("think" more) before acting, but this does not allow agents to acquire new information from the environment or adapt behavior over time. In this work, we propose scaling test-time interaction, an untapped dimension for test-time scaling that increases the agent's interaction horizon to enable rich behaviors such as exploration, backtracking, and dynamic re-planning within a single rollout. To demonstrate the promise of this scaling dimension, we situate our study in the domain of web agents. We first show that even prompting-based interaction scaling can improve task success on web benchmarks non-trivially. Building on this, we introduce TTI, a curriculum-based online reinforcement learning (RL) approach that trains agents by adaptively adjusting their interaction lengths during rollout. Using a Gemma 3 12B model, TTI sets a new state-of-the-art among open-source agents trained on public data on WebVoyager and WebArena. Case studies further reveal that TTI enables agents to balance exploration and exploitation adaptively. Our results establish interaction scaling as a powerful, complementary axis to scaling per-action compute, offering new avenues for training robust and adaptive agents.

Estimating the Kullback--Leibler (KL) divergence between language models has many applications, e.g., reinforcement learning from human feedback (RLHF), interpretability, and knowledge distillation. However, computing the exact KL divergence between two arbitrary language models is intractable. Thus, practitioners often resort to sampling-based estimators. While it is easy to fashion a simple Monte Carlo (MC) estimator that provides an unbiased estimate of the KL divergence between language models, this estimator notoriously suffers from high variance and can even result in a negative estimate of the KL divergence, a non-negative quantity. In this paper, we introduce a Rao--Blackwellized estimator that is unbiased and provably has variance less than or equal to that of the standard Monte Carlo estimator. In an empirical study on sentiment-controlled fine-tuning, we show that our estimator provides more stable KL estimates and reduces variance substantially. Additionally, we derive an analogous Rao--Blackwellized estimator of the gradient of the KL divergence, which leads to more stable training and produces models that more frequently appear on the Pareto frontier of reward vs. KL compared to the ones trained with the MC estimator of the gradient.

Behavior cloning has shown promise for robot manipulation, but real-world demonstrations are costly to acquire at scale. While simulated data offers a scalable alternative, particularly with advances in automated demonstration generation, transferring policies to the real world is hampered by various simulation and real domain gaps. In this work, we propose a unified sim-and-real co-training framework for learning generalizable manipulation policies that primarily leverages simulation and only requires a few real-world demonstrations. Central to our approach is learning a domain-invariant, task-relevant feature space. Our key insight is that aligning the joint distributions of observations and their corresponding actions across domains provides a richer signal than aligning observations (marginals) alone. We achieve this by embedding an Optimal Transport (OT)-inspired loss within the co-training framework, and extend this to an Unbalanced OT framework to handle the imbalance between abundant simulation data and limited real-world examples. We validate our method on challenging manipulation tasks, showing it can leverage abundant simulation data to achieve up to a 30\% improvement in the real-world success rate and even generalize to scenarios seen only in simulation.

Human-in-the-loop (HIL) imitation learning enables agents to learn complex behaviors safely through real-time human intervention. However, existing methods struggle to efficiently leverage agent-generated data due to dynamically evolving trajectory distributions and imperfections caused by human intervention delays, often failing to faithfully imitate the human expert policy. In this work, we propose Faithful Dynamic Imitation Learning (FaithDaIL) to address these challenges. We formulate HIL imitation learning as an online non-convex problem and employ dynamic regret minimization to adapt to the shifting data distribution and track high-quality policy trajectories. To ensure faithful imitation of the human expert despite training on mixed agent and human data, we introduce an unbiased imitation objective and achieve it by weighting the behavior distribution relative to the human expert's as a proxy reward. Extensive experiments on MetaDrive and CARLA driving benchmarks demonstrate that FaithDaIL achieves state-of-the-art performance in safety and task success with significantly reduced human intervention data compared to prior HIL baselines.

Diffusion and flow models have emerged as powerful generative approaches capable of modeling diverse and multimodal behavior. However, applying these models to offline RL remains challenging due to the iterative nature of their noise sampling processes, making policy optimization difficult. In this paper, we introduce Scalable Offline Reinforcement Learning (SORL), a new offline RL algorithm that leverages shortcut models – a novel class of generative models – to scale both training and inference. SORL's policy can capture complex data distributions and can be trained simply and efficiently in a one-stage training procedure. At test time, SORL supports both sequential and parallel inference scaling by using the learned Q-function as a verifier. We demonstrate that SORL achieves strong performance across a range of offline RL tasks and exhibits positive scaling behavior with increased test-time compute.

DeepSeek-R1 has demonstrated remarkable effectiveness in incentivizing reasoning and generalization capabilities of large language models (LLMs) through reinforcement learning. Nevertheless, the potential of reasoning-induced computation has not been thoroughly explored in the context of image quality assessment (IQA), a task depending critically on visual reasoning. In this paper, we introduce VisualQuality-R1, a reasoning-induced no-reference IQA (NR-IQA) model, and we train it with reinforcement learning to rank, a learning algorithm tailored to the intrinsically relative nature of visual quality. Specifically, for a pair of images, we employ group relative policy optimization to generate multiple quality scores for each image. These estimates are used to compute comparative probabilities of one image having higher quality than the other under the Thurstone model. Rewards for each quality estimate are defined using continuous fidelity measures rather than discretized binary labels. Extensive experiments show that the proposed VisualQuality-R1 consistently outperforms discriminative deep learning-based NR-IQA models as well as a recent reasoning-induced quality regression method. Moreover, VisualQuality-R1 is capable of generating contextually rich, human-aligned quality descriptions, and supports multi-dataset training without requiring perceptual scale realignment. These features make VisualQuality-R1 especially well-suited for reliably measuring progress in a wide range of image processing tasks like super-resolution and image generation.

Designing adaptive mechanisms to align individual and collective interests remains a central challenge in artificial social intelligence. Existing methods often struggle with modeling heterogeneous agents possessing persistent latent traits (e.g., skills, preferences) and dealing with complex multi-agent system dynamics. These challenges are compounded by the critical need for high sample efficiency due to costly real-world interactions. World Models, by learning to predict environmental dynamics, offer a promising pathway to enhance mechanism design in heterogeneous and complex systems. In this paper, we introduce a novel method named SWM-AP (Social World Model-Augmented Mechanism Design Policy Learning), which learns a social world model hierarchically modeling agents' behavior to enhance mechanism design. Specifically, the social world model infers agents' traits from their interaction trajectories and learns a trait-based model to predict agents' responses to the deployed mechanisms. The mechanism design policy collects extensive training trajectories by interacting with the social world model, while concurrently inferring agents' traits online during real-world interactions to further boost policy learning efficiency. Experiments in diverse settings (tax policy design, team coordination, and facility location) demonstrate that SWM-AP outperforms established model-based and model-free RL baselines in cumulative rewards and sample efficiency.

Recent advancements in large language models (LLMs) underscore the need for more comprehensive evaluation methods to accurately assess their reasoning capabilities. Existing benchmarks are often domain-specific and thus cannot fully capture an LLM’s general reasoning potential. To address this limitation, we introduce the **Knowledge Orthogonal Reasoning Gymnasium (KORGym)**, a dynamic evaluation platform inspired by KOR-Bench and Gymnasium. KORGym offers over fifty games in either textual or visual formats and supports interactive, multi-turn assessments with reinforcement learning scenarios. Using KORGym, we conduct extensive experiments on 19 LLMs and 8 VLMs, revealing consistent reasoning patterns within model families and demonstrating the superior performance of closed-source models. Further analysis examines the effects of modality, reasoning strategies, reinforcement learning techniques, and response length on model performance. We expect KORGym to become a valuable resource for advancing LLM reasoning research and developing evaluation methodologies suited to complex, interactive environments.

We study a stochastic principal-agent model. A principal and an agent interact in a stochastic environment, each privy to observations about the state not available to the other. The principal has the power of commitment, both to elicit information from the agent and to signal her own information. The players communicate with each other and then select actions independently. Both players are {\em far-sighted}, aiming to maximize their total payoffs over the entire time horizon. We consider both the computation and learning of the principal's optimal policy. The key challenge lies in enabling {\em history-dependent} policies, which are essential for achieving optimality in this model but difficult to cope with because of the exponential growth of possible histories as the size of the model increases; explicit representation of history-dependent policies is infeasible as a result. To address this challenge, we develop algorithmic techniques based on the concept of {\em inducible value set}. The techniques yield an efficient algorithm that computes an $\epsilon$-approximate optimal policy in time polynomial in $1/\epsilon$. We also present an efficient learning algorithm for an episodic reinforcement learning setting with unknown transition probabilities. The algorithm achieves sublinear regret $\widetilde{\mathcal{O}}(T^{2/3})$ for both players over $T$ episodes.

Recent advances in language modeling have demonstrated the effectiveness of State Space Models (SSMs) for efficient sequence modeling. While hybrid architectures such as Samba and the decoder-decoder architecture, YOCO, have shown promising performance gains over Transformers, prior works have not investigated the efficiency potential of representation sharing between SSM layers. In this paper, we introduce the Gated Memory Unit (GMU), a simple yet effective mechanism for efficient memory sharing across layers. We apply it to create SambaY, a decoder-hybrid-decoder architecture that incorporates GMUs in the cross-decoder to share memory readout states from a Samba-based self-decoder. SambaY significantly enhances decoding efficiency, preserves linear pre-filling time complexity, and boosts long-context performance, all while eliminating the need for explicit positional encoding. Through extensive scaling experiments, we demonstrate that our model exhibits a significantly lower irreducible loss compared to a strong YOCO baseline, indicating superior performance scalability under large-scale compute regimes. Our largest model enhanced with Differential Attention, Phi4-mini-Flash-Reasoning, achieves significantly better performance than Phi4-mini-Reasoning on reasoning tasks such as Math500, AIME24/25, and GPQA Diamond without any reinforcement learning, while delivering up to 10× higher decoding throughput on 2K-length prompts with 32K generation length under the vLLM inference framework. We release our training codebase on open-source data at https://github.com/microsoft/ArchScale.

We study the problem of offline imitation learning in Markov decision processes (MDPs), where the goal is to learn a well-performing policy given a dataset of state-action pairs generated by an expert policy. Complementing a recent line of work on this topic that assumes that the expert policy belongs to a tractable class of known policies, we approach this problem from a new angle and leverage another type of structural assumption about the environment. Specifically, for the class of linear $Q^\pi$-realizable MDPs, we introduce a new algorithm called primal-dual offline imitation learning (PDOIL), which is guaranteed to match the performance of any expert up to an additive error $\varepsilon$ with access to $\mathcal{O}(\varepsilon^{-2})$ samples. Moreover, we extend this result to possibly non-linear $Q^\pi$-realizable MDPs at the cost of a worst sample complexity of order $\mathcal{O}(\varepsilon^{-4})$. Finally, our analysis suggests a new loss function for training critic networks from expert data in deep imitation learning. Empirical evaluations on standard benchmarks demonstrate that the neural net implementation of PDOIL is superior to behavior cloning and competitive with state-of-the-art algorithms.

Recent advances in reinforcement learning (RL) methods such as Grouped Relative Policy Optimization (GRPO) have strengthened the reasoning capabilities of Large Vision-Language Models (LVLMs). However, due to the inherent entanglement between visual and textual modalities, applying GRPO to LVLMs often leads to reward convergence across different responses to the same sample as training progresses, hindering effective gradient updates and causing the enhancement of chain-of-thought reasoning to stagnate or even collapse. To address this issue, we propose a progressive instruction evolution framework, EvolvedGRPO, to gradually generate more complex questions via editing instructions in an adversarial way, progressively aligned with the model’s evolving capabilities. Specifically, we design two instruction editing strategies across modalities, incorporating incrementally increasing editing instructions and RL-based adversarial data augmentation to improve the effectiveness of model training. To address GRPO's limitations on overly difficult problems, we first train on basic subproblem versions of complex multi-modal questions in both the visual and textual modalities, progressively increasing difficulty to enable prefix-style process rewards, effectively combining the strengths of both process rewards and group-wise relative rewards. Finally, EvolvedGRPO achieves state-of-the-art performance among open-source RL models on multi-modal reasoning tasks, even approaching the closed-source GPT-4o in reasoning capabilities, and demonstrates better performance on unseen LVLM general benchmarks. The Code for EvolvedGRPO is available at https://github.com/SHENZHEBEI/EvolvedGRPO.

In recent years, Multimodal Large Language Models (MLLMs) have been extensively utilized for multimodal reasoning tasks, including Graphical User Interface (GUI) automation. Unlike general offline multimodal tasks, GUI automation is executed in online interactive environments, necessitating step-by-step decision-making based on the real-time status of the environment. This task has a lower tolerance for decision-making errors at each step, as any mistakes may cumulatively disrupt the process and potentially lead to irreversible outcomes like deletions or payments. To address these issues, we introduce a pre-operative critic mechanism that provides effective feedback prior to the actual execution, by reasoning about the potential outcome and correctness of actions. Specifically, we propose a Suggestion-aware Group Relative Policy Optimization (S-GRPO) strategy to construct our pre-operative critic model GUI-Critic-R1, incorporating a novel suggestion reward to enhance the reliability of the model's feedback. Furthermore, we develop a reasoning-bootstrapping based data collection pipeline to create a GUI-Critic-Train and a GUI-Critic-Test, filling existing gaps in GUI critic data. Static experiments on the GUI-Critic-Test across both mobile and web domains reveal that our GUI-Critic-R1 offers significant advantages in critic accuracy compared to current MLLMs. Dynamic evaluation on GUI automation benchmark further highlights the effectiveness and superiority of our model, as evidenced by improved success rates and operational efficiency. The code is available at https://github.com/X-PLUG/MobileAgent/tree/main/GUI-Critic-R1.

Model-free reinforcement learning (RL) combined with diffusion models has achieved significant progress in addressing complex continuous control tasks. However, a persistent challenge in RL remains the accurate estimation of Q-values, which critically governs the efficacy of policy optimization. Although recent advances employ parametric distributions to model value distributions for enhanced estimation accuracy, current methodologies predominantly rely on unimodal Gaussian assumptions or quantile representations. These constraints introduce distributional bias between the learned and true value distributions, particularly in some tasks with a nonstationary policy, ultimately degrading performance. To address these limitations, we propose value diffusion reinforcement learning (VDRL), a novel model-free online RL method that utilizes the generative capacity of diffusion models to represent multimodal value distributions. The core innovation of VDRL lies in the use of the variational loss of diffusion-based value distribution, which is theoretically proven to be a tight lower bound for the optimization objective under the KL-divergence measurement. Furthermore, we introduce double value diffusion learning with sample selection to enhance training stability and further improve value estimation accuracy. Extensive experiments conducted on the MuJoCo benchmark demonstrate that VDRL significantly outperforms some SOTA model-free online RL baselines, showcasing its effectiveness and robustness.