NeurIPS 2025 论文列表

A long-standing goal in AI is to develop agents capable of solving diverse tasks across a range of environments, including those never seen during training. Two dominant paradigms address this challenge: (i) reinforcement learning (RL), which learns policies via trial and error, and (ii) optimal control, which plans actions using a known or learned dynamics model. However, their comparative strengths in the offline setting—where agents must learn from reward-free trajectories—remain underexplored. In this work, we systematically evaluate RL and control-based methods on a suite of navigation tasks, using offline datasets of varying quality. On the RL side, we consider goal-conditioned and zero-shot methods. On the control side, we train a latent dynamics model using the Joint Embedding Predictive Architecture (JEPA) and employ it for planning. We investigate how factors such as data diversity, trajectory quality, and environment variability influence the performance of these approaches. Our results show that model-free RL benefits most from large amounts of high-quality data, whereas model-based planning generalizes better to unseen layouts and is more data-efficient, while achieving trajectory stitching performance comparable to leading model-free methods. Notably, planning with a latent dynamics model proves to be a strong approach for handling suboptimal offline data and adapting to diverse environments.

Bilevel reinforcement learning (BRL) has emerged as a powerful framework for aligning generative models, yet its theoretical foundations, especially sample complexity bounds, remain underexplored. In this work, we present the first sample complexity bound for BRL, establishing a rate of $\mathcal{O}(\epsilon^{-3})$ in continuous state-action spaces. Traditional MDP analysis techniques do not extend to BRL due to its nested structure and non-convex lower-level problems. We overcome these challenges by leveraging the Polyak-Łojasiewicz (PL) condition and the MDP structure to obtain closed-form gradients, enabling tight sample complexity analysis. Our analysis also extends to general bi-level optimization settings with non-convex lower levels, where we achieve state-of-the-art sample complexity results of $\mathcal{O}(\epsilon^{-3})$ improving upon existing bounds of $\mathcal{O}(\epsilon^{-6})$. Additionally, we address the computational bottleneck of hypergradient estimation by proposing a fully first-order, Hessian-free algorithm suitable for large-scale problems.

We propose a framework for hypothesis testing on conditional probability distributions, which we then use to construct *statistical tests of functionals of conditional distributions*. These tests identify the inputs where the functionals differ with high probability, and include tests of conditional moments or two-sample tests. Our key idea is to transform confidence bounds of a learning method into a test of conditional expectations. We instantiate this principle for kernel ridge regression (KRR) with subgaussian noise. An intermediate data embedding then enables more general tests — including *conditional two-sample tests* — via kernel mean embeddings of distributions. To have guarantees in this setting, we generalize existing pointwise-in-time or time-uniform confidence bounds for KRR to previously-inaccessible yet essential cases such as infinite-dimensional outputs with non-trace-class kernels. These bounds also circumvent the need for independent data, allowing for instance online sampling. To make our tests readily applicable in practice, we introduce bootstrapping schemes leveraging the parametric form of testing thresholds identified in theory to avoid tuning inaccessible parameters. We illustrate the tests on examples, including one in process monitoring and comparison of dynamical systems. Overall, our results establish a comprehensive foundation for conditional testing on functionals, from theoretical guarantees to an algorithmic implementation, and advance the state of the art on confidence bounds for vector-valued least squares estimation.

Multimodal large language models (MLLMs) have shown promising capabilities in reasoning tasks, yet still struggle significantly with complex problems requiring explicit self-reflection and self-correction, especially compared to their unimodal text-based counterparts. Existing reflection methods are simplistic and struggle to generate meaningful, instructive feedback, as the reasoning ability and knowledge limits of pre-trained models are largely fixed during initial training. To overcome these challenges, we propose \textit{multimodal \textbf{S}elf-\textbf{R}eflection enhanced reasoning with Group Relative \textbf{P}olicy \textbf{O}ptimization} \textbf{SRPO}, a two-stage reflection-aware reinforcement learning (RL) framework explicitly designed to enhance multimodal LLM reasoning. In the first stage, we construct a high-quality, reflection-focused dataset under the guidance of an advanced MLLM, which generates reflections based on initial responses to help the policy model to learn both reasoning and self-reflection. In the second stage, we introduce a novel reward mechanism within the GRPO framework that encourages concise and cognitively meaningful reflection while avoiding redundancy. Extensive experiments across multiple multimodal reasoning benchmarks—including MathVista, MathVision, Mathverse, and MMMU-Pro—using Qwen-2.5-VL-7B and Qwen-2.5-VL-32B demonstrate that SRPO significantly outperforms state-of-the-art models, achieving notable improvements in both reasoning accuracy and reflection quality.

Retrieval-augmented generation (RAG) enhances large language models (LLMs) by integrating external knowledge, yet traditional RAG systems struggle with static workflows and limited adaptability for complex, multistep reasoning tasks. Agentic RAG systems, such as DeepResearch, address these issues through dynamic retrieval, iterative context refinement, and adaptive workflows. However, recent methods like Search-R1, which rely on outcome-based reinforcement learning, face challenges such as low exploration efficiency, gradient conflict, and sparse reward signals. To tackle these limitations, we introduce ReasonRAG, a novel method that leverages RAG-ProGUIDE—a high-quality dataset providing fine-grained, process-level rewards for query generation, evidence extraction, and answer generation. By employing process-supervised reinforcement learning, ReasonRAG enhances LLMs’ autonomous capabilities in search, query generation, evidence extraction, and answer synthesis. Experimental results show that ReasonRAG, utilizing RAG-ProGUIDE, outperforms existing approaches like Search-R1 and traditional RAG systems, achieving superior performance on five benchmark datasets with only 5k training instances—significantly fewer than the 90k required by Search-R1. Our code is available at https://github.com/Applied-Machine-Learning-Lab/ReasonRAG.

The task of Knowlegde-Based Visual Question Answering (KB-VQA) requires the model to understand visual features and retrieve external knowledge. Retrieval-Augmented Generation (RAG) have been employed to address this problem through knowledge base querying. However, existing work demonstrate two limitations: insufficient interactivity during knowledge retrieval and ineffective organization of retrieved information for Visual-Language Model (VLM). To address these challenges, we propose a three-stage visual language model with Process, Retrieve and Filter (VLM-PRF) framework. For interactive retrieval, VLM-PRF uses reinforcement learning (RL) to guide the model to strategically process information via tool-driven operations. For knowledge filtering, our method trains the VLM to transform the raw retrieved information into into task-specific knowledge. With a dual reward as supervisory signals, VLM-PRF successfully enable model to optimize retrieval strategies and answer generation capabilities simultaneously. Experiments on two datasets demonstrate the effectiveness of our framework.

Comprehending natural language and following human instructions are critical capabilities for intelligent agents. However, the flexibility of linguistic instructions induces substantial ambiguity across language-conditioned tasks, severely degrading algorithmic performance. To address these limitations, we present a novel method named DAIL (Distributional Aligned Learning), featuring two key components: distributional policy and semantic alignment. Specifically, we provide theoretical results that the value distribution estimation mechanism enhances task differentiability. Meanwhile, the semantic alignment module captures the correspondence between trajectories and linguistic instructions. Extensive experimental results on both structured and visual observation benchmarks demonstrate that DAIL effectively resolves instruction ambiguities, achieving superior performance to baseline methods. Our implementation is available at https://github.com/RunpengXie/Distributional-Aligned-Learning.

The scaling of model and data sizes has reshaped the AI landscape, establishing finetuning pretrained models as the standard paradigm for solving downstream tasks. However, dominant finetuning methods typically rely on weight adaptation, often lack interpretability, and depend on heuristically chosen hyperparameters. In this paper, we take a different perspective and shift the focus from weights to activation functions, viewing them through the lens of spline operators. We propose Curvature Tuning (CT), an interpretable and principled steering method that modulates a model's decision boundary by injecting a single hyperparameter into its activation functions. We show that CT provably adjusts model decision boundary curvature and, more fundamentally, projects a model onto a space of smooth functions—thereby complementing current finetuning methods, whose effect lies primarily in feature adaptation. Making this hyperparameter trainable gives rise to a novel and highly parameter-efficient finetuning method. Empirically, CT improves both generalization and robustness. For example, it boosts downstream accuracy of ResNet-50/152 by 8.59\%/8.34\% over linear probing and 4.64\%/1.70\% over LoRA across 12 datasets, and improves robust accuracy on the $\ell_{\infty}$ benchmark from RobustBench by 1032.64\%/1494.46\%. Our code is available at https://github.com/Leon-Leyang/curvature-tuning.

Sparse-reward reinforcement learning (RL) can model a wide range of highly complex tasks. Solving sparse-reward tasks is RL's core premise — requiring efficient exploration coupled with long-horizon credit assignment — and overcoming these challenges is key for building self-improving agents with superhuman ability. We argue that solving complex and high-dimensional tasks requires solving simpler tasks that are *relevant* to the target task. In contrast, most prior work designs strategies for selecting exploratory tasks with the objective of solving *any* task, making exploration of challenging high-dimensional, long-horizon tasks intractable. We find that the sense of direction, necessary for effective exploration, can be extracted from existing reinforcement learning algorithms, without needing any prior information. Based on this finding, we propose a method for _**di**rected **s**parse-reward goal-**co**nditioned **ve**ry long-horizon **R**L_ (DISCOVER), which selects exploratory goals in the direction of the target task. We connect DISCOVER to principled exploration in bandits, formally bounding the time until the target task becomes achievable in terms of the agent's initial distance to the target, but independent of the volume of the space of all tasks. Empirically, we perform a thorough evaluation in high-dimensional simulated environments. We find that the directed goal selection of DISCOVER solves exploration problems that are beyond the reach of prior state-of-the-art exploration methods in RL.

Recent advances in diffusion models have revolutionized text-guided image editing, yet existing editing methods face critical challenges in hyperparameter identification. To get the reasonable editing performance, these methods often require the user to brute-force tune multiple interdependent hyperparameters, such as inversion timesteps and attention modification, \textit{etc.} This process incurs high computational costs due to the huge hyperparameter search space. We consider searching optimal editing's hyperparameters as a sequential decision-making task within the diffusion denoising process. Specifically, we propose a reinforcement learning framework, which establishes a Markov Decision Process that dynamically adjusts hyperparameters across denoising steps, integrating editing objectives into a reward function. The method achieves time efficiency through proximal policy optimization while maintaining optimal hyperparameter configurations. Experiments demonstrate significant reduction in search time and computational overhead compared to existing brute-force approaches, advancing the practical deployment of a diffusion-based image editing framework in the real world.

Multi-modal Large Language Models (MLLMs) excel at single-image tasks but struggle with multi-image understanding due to cross-modal misalignment, leading to hallucinations (context omission, conflation, and misinterpretation). Existing methods using Direct Preference Optimization (DPO) constrain optimization to a solitary image reference within the input sequence, neglecting holistic context modeling. To address this, we propose Context-to-Cue Direct Preference Optimization (CcDPO), a multi-level preference optimization framework that enhances per-image perception in multi-image settings by zooming into visual clues—from sequential context to local details. Our approach features two sequentially dependent components: (i) Context-Level Optimization: By introducing low-cost sequence preference pairs, we optimize the model to distinguish between complete and disrupted multi-image contexts, thereby correcting cognitive biases in MLLMs’ multi-image understanding. (ii) Needle-Level Optimization: By integrating region-specific visual prompts with multimodal preference supervision, we direct the model’s attention to critical visual details, effectively suppressing perceptual biases toward fine-grained visual information. To support scalable optimization, we also construct MultiScope-42k, an automatically generated multi-image dataset with hierarchical preference pairs. Experiments show that CcDPO significantly reduces hallucinations and yields consistent performance gains across general single- and multi-image tasks. Codes are available at https://github.com/LXDxmu/CcDPO.

Positive-Unlabeled (PU) learning refers to a specific weakly-supervised learning paradigm that induces a binary classifier with a few positive labeled instances and massive unlabeled instances. To handle this task, the community has proposed dozens of PU learning methods with various techniques, demonstrating strong potential. In this paper, we conduct a comprehensive study to investigate the basic characteristics of current PU learning methods. We organize them into two fundamental families of PU learning, including *disambiguation-free empirical risks*, which approximate the expected risk of supervised learning, and *pseudo-labeling methods*, which estimate pseudo-labels for unlabeled instances. First, we make an empirical analysis on disambiguation-free empirical risks such as uPU, nnPU, and DistPU, and suggest a novel risk-consistent set-aware empirical risk from the perspective of aggregate supervision. Second, we make an empirical analysis of pseudo-labeling methods to evaluate the potential of pseudo-label estimation techniques and widely applied generic tricks in PU learning. Finally, based on those empirical findings, we propose a general framework of PU learning by integrating the set-aware empirical risk with pseudo-labeling. Compared with existing PU learning methods, the proposed framework can be a practical benchmark in PU learning.

We consider how to construct state abstractions compatible with a given set of abstract actions, to obtain a well-formed abstract Markov decision process (MDP). We show that the Bellman equation suggests that abstract states should represent distributions over states in the ground MDP; we characterize the conditions under which the resulting process is Markov and approximately model-preserving, derive algorithms for constructing and planning with the abstract MDP, and apply them to a visual maze task. We generalize these results to the factored actions case, characterizing the conditions that result in factored abstract states and apply the resulting algorithm to Montezuma's Revenge. These results provide a powerful and principled framework for constructing neurosymbolic abstract Markov decision processes.

Offline goal-conditioned reinforcement learning (GCRL) offers a practical learning paradigm in which goal-reaching policies are trained from abundant state–action trajectory datasets without additional environment interaction. However, offline GCRL still struggles with long-horizon tasks, even with recent advances that employ hierarchical policy structures, such as HIQL. Identifying the root cause of this challenge, we observe the following insight. Firstly, performance bottlenecks mainly stem from the high-level policy’s inability to generate appropriate subgoals. Secondly, when learning the high-level policy in the long-horizon regime, the sign of the advantage estimate frequently becomes incorrect. Thus, we argue that improving the value function to produce a clear advantage estimate for learning the high-level policy is essential. In this paper, we propose a simple yet effective solution: _**Option-aware Temporally Abstracted**_ value learning, dubbed **OTA**, which incorporates temporal abstraction into the temporal-difference learning process. By modifying the value update to be _option-aware_, our approach contracts the effective horizon length, enabling better advantage estimates even in long-horizon regimes. We experimentally show that the high-level policy learned using the OTA value function achieves strong performance on complex tasks from OGBench, a recently proposed offline GCRL benchmark, including maze navigation and visual robotic manipulation environments. Our code is available at https://github.com/ota-v/ota-v

We introduce *OpenVLThinker*, one of the first open-source large vision–language models (LVLMs) to exhibit sophisticated chain-of-thought reasoning, achieving notable performance gains on challenging visual reasoning tasks. While text-based reasoning models (e.g., Deepseek R1) show promising results in text-only tasks, distilling their reasoning into LVLMs via supervised fine-tuning (SFT) often results in performance degradation due to imprecise visual grounding. Conversely, purely reinforcement learning (RL)-based methods face a large search space, hindering the emergence of reflective behaviors in smaller models (e.g., 7B LVLMs). Surprisingly, alternating between SFT and RL ultimately results in significant performance improvements after a few iterations. Our analysis reveals that the base model rarely exhibits reasoning behaviors initially, but SFT effectively surfaces these latent actions and narrows the RL search space, accelerating the development of reasoning capabilities. Each subsequent RL stage further refines the model's reasoning skills, producing higher-quality SFT data for continued self-improvement. OpenVLThinker-7B consistently advances performance across six benchmarks demanding mathematical and general reasoning, notably improving MathVista by 3.2\%, EMMA by 1.4\%, and HallusionBench by 2.7\%. Beyond demonstrating the synergy between SFT and RL for complex reasoning tasks, our findings provide early evidence towards achieving R1-style reasoning in multimodal contexts.

This paper introduces the Real-DRL framework for safety-critical autonomous systems, enabling runtime learning of a deep reinforcement learning (DRL) agent to develop safe and high-performance action policies in real plants while prioritizing safety. The Real-DRL consists of three interactive components: a DRL-Student, a PHY-Teacher, and a Trigger. The DRL-Student is a DRL agent that innovates in the dual self-learning and teaching-to-learn paradigm and the safety-status-dependent batch sampling. On the other hand, PHY-Teacher is a physics-model-based design of action policies that focuses solely on safety-critical functions. PHY-Teacher is novel in its real-time patch for two key missions: i) fostering the teaching-to-learn paradigm for DRL-Student and ii) backing up the safety of real plants. The Trigger manages the interaction between the DRL-Student and the PHY-Teacher. Powered by the three interactive components, the Real-DRL can effectively address safety challenges that arise from the unknown unknowns and the Sim2Real gap. Additionally, Real-DRL notably features i) assured safety, ii) automatic hierarchy learning (i.e., safety-first learning and then high-performance learning), and iii) safety-informed batch sampling to address the experience imbalance caused by corner cases. Experiments with a real quadruped robot, a quadruped robot in Nvidia Isaac Gym, and a cart-pole system, along with comparisons and ablation studies, demonstrate the Real-DRL's effectiveness and unique features.

Holdout validation and hyperparameter tuning from data is a long-standing problem in offline reinforcement learning (RL). A standard framework is to use off-policy evaluation (OPE) methods to evaluate and select the policies, but OPE either incurs exponential variance (e.g., importance sampling) or has hyperparameters on their own (e.g., FQE and model-based). We focus on hyperparameter tuning for OPE itself, which is even more under-investigated. Concretely, we select among candidate value functions ("model-free") or dynamics models ("model-based") to best assess the performance of a target policy. We develop: (1) new model-free and model-based selectors with theoretical guarantees, and (2) a new experimental protocol for empirically evaluating them. Compared to the model-free protocol in prior works, our new protocol allows for more stable generation and better control of candidate value functions in an optimization-free manner, and evaluation of model-free and model-based methods alike. We exemplify the protocol on Gym-Hopper, and find that our new model-free selector, LSTD-Tournament, demonstrates promising empirical performance.

We present Prototypical Pair Network (ProtoPairNet), a novel interpretable architecture that combines deep learning with case-based reasoning to predict continuous targets. While prototype-based models have primarily addressed image classification with discrete outputs, extending these methods to continuous targets, such as regression, poses significant challenges. Existing architectures which rely heavily on one-to-one comparison with prototypes lack the directional information necessary for continuous predictions. Our method redefines the role of prototypes in such tasks by incorporating prototypical pairs into the reasoning process. Predictions are derived based on the input's relative dissimilarities to these pairs, leveraging an intuitive geometric interpretation. Our method further reduces the complexity of the reasoning process by relying on the single most relevant pair of prototypes, rather than all prototypes in the model as was done in prior works. Our model is versatile enough to be used in both vision-based regression and continuous control in reinforcement learning. Our experiments demonstrate that ProtoPairNet achieves performance on par with its black-box counterparts across these tasks. Comprehensive analyses confirm the meaningfulness of prototypical pairs and the faithfulness of our model’s interpretations, and extensive user studies highlight our model's improved interpretability over existing methods.

Continual learning is often motivated by the idea, known as the big world hypothesis, that the "world is bigger" than the agent. Recent problem formulations capture this idea by explicitly constraining an agent relative to the environment. These constraints lead to solutions in which the agent continually adapts to best use its limited capacity, rather than converging to a fixed solution. However, explicit constraints can be ad hoc, difficult to incorporate, and limiting to the effectiveness of scaling up the agent's capacity. In this paper, we characterize a problem setting in which an agent, regardless of its capacity, is implicitly constrained by being embedded in the environment. In particular, we introduce a computationally-embedded perspective that represents an embedded agent as an automaton simulated within a universal (formal) computer. We prove that such an automaton is implicitly constrained and that it is equivalent to an agent that interacts with a partially observable Markov decision process over a countably infinite state-space. We then propose an objective for this setting, which we call interactivity, that measures an agent's ability to continually adapt its behaviour and to continually learn new predictions. We develop a reinforcement learning algorithm for maximizing interactivity and a synthetic benchmark to experimentation on continual learning. Our results indicate that deep nonlinear networks struggle to sustain interactivity whereas deep linear networks can achieve higher interactivity as capacity increases.