NeurIPS 2025 Papers

Standard training metrics like loss fail to explain the emergence of complex capabilities in large language models. We take a spectral approach to investigate the geometry of learned representations across pretraining and post-training, measuring effective rank (RankMe) and eigenspectrum decay (αReQ). With OLMo (1B-7B) and Pythia (160M-12B) models, we uncover a consistent non-monotonic sequence of three geometric phases during autoregressive pretraining. The initial “warmup” phase exhibits rapid representational collapse. This is followed by an “entropy-seeking” phase, where the manifold’s dimensionality expands substantially, coinciding with peak n-gram memorization. Subsequently, a “compression-seeking” phase imposes anisotropic consolidation, selectively preserving variance along dominant eigendirections while contracting others, a transition marked with significant improvement in downstream task performance. We show these phases can emerge from a fundamental interplay of cross-entropy optimization under skewed token frequencies and representational bottlenecks (d ≪ |V|). Post-training further transforms geometry: SFT and DPO drive “entropy-seeking” dynamics to integrate specific instructional or preferential data, improving in-distribution performance while degrading out-of-distribution robustness. Conversely, RLVR induces “compression-seeking” , enhancing reward alignment but reducing generation diversity.

Designing effective black‑box optimizers is hampered by limited problem-specific knowledge and manual control that spans months for almost every detail. In this paper, we present DesignX, the first automated algorithm design framework that generates an effective optimizer specific to a given black-box optimization problem within seconds. Rooted in the first principles, we identify two key sub-tasks: 1) algorithm structure generation and 2) hyperparameter control. To enable systematic construction, a comprehensive modular algorithmic space is first built, embracing hundreds of algorithm components collected from decades of research. We then introduce a dual-agent reinforcement learning system that collaborates on structural and parametric design through a novel cooperative training objective, enabling large-scale meta-training across 10k diverse instances. Remarkably, through days of autonomous learning, the DesignX-generated optimizers continuously surpass human-crafted optimizers by orders of magnitude, either on synthetic testbed or on realistic optimization scenarios such as Protein-docking, AutoML and UAV path planning. Further in-depth analysis reveals DesignX's capability to discover non-trivial algorithm patterns beyond expert intuition, which, conversely, provides valuable design insights for the optimization community. We provide DesignX's Python project at~\url{https://github.com/MetaEvo/DesignX}.

Reinforcement learning has achieved significant milestones, but sample efficiency remains a bottleneck for real-world applications. Recently, CrossQ has demonstrated state-of-the-art sample efficiency with a low update-to-data (UTD) ratio of 1. In this work, we explore CrossQ's scaling behavior with higher UTD ratios. We identify challenges in the training dynamics, which are emphasized by higher UTD ratios. To address these, we integrate weight normalization into the CrossQ framework, a solution that stabilizes training, has been shown to prevent potential loss of plasticity, and keeps the effective learning rate constant. Our proposed approach reliably scales with increasing UTD ratios, achieving competitive performance across 25 challenging continuous control tasks on the DeepMind Control Suite and Myosuite benchmarks, notably the complex dog and humanoid environments. This work eliminates the need for drastic interventions, such as network resets, and offers a simple yet robust pathway for improving sample efficiency and scalability in model-free reinforcement learning.

Pretraining a policy on offline data followed by fine-tuning through online interactions, known as Offline-to-Online Reinforcement Learning (O2O RL), has emerged as a promising paradigm for real-world RL deployment. However, both offline datasets and online interactions in practical environments are often noisy or even maliciously corrupted, severely degrading the performance of O2O RL. Existing works primarily focus on mitigating the conservatism of offline policies via online exploration, while the robustness of O2O RL under data corruption, including states, actions, rewards, and dynamics, is still unexplored. In this work, we observe that data corruption induces heavy-tailed behavior in the policy, thereby substantially degrading the efficiency of online exploration. To address this issue, we incorporate Inverse Probability Weighted (IPW) into the online exploration policy to alleviate heavy-tailedness, and propose a novel, simple yet effective method termed $\textbf{RPEX}$: $\textbf{R}$obust $\textbf{P}$olicy $\textbf{EX}$pansion. Extensive experimental results on D4RL datasets demonstrate that RPEX achieves SOTA O2O performance across a wide range of data corruption scenarios.

We present the first finite-sample analysis of policy evaluation in robust average-reward Markov Decision Processes (MDPs). Prior work in this setting have established only asymptotic convergence guarantees, leaving open the question of sample complexity. In this work, we address this gap by showing that the robust Bellman operator is a contraction under a carefully constructed semi-norm, and developing a stochastic approximation framework with controlled bias. Our approach builds upon Multi-Level Monte Carlo (MLMC) techniques to estimate the robust Bellman operator efficiently. To overcome the infinite expected sample complexity inherent in standard MLMC, we introduce a truncation mechanism based on a geometric distribution, ensuring a finite expected sample complexity while maintaining a small bias that decays exponentially with the truncation level. Our method achieves the order-optimal sample complexity of $\tilde{\mathcal{O}}(\epsilon^{-2})$ for robust policy evaluation and robust average reward estimation, marking a significant advancement in robust reinforcement learning theory.

Decision tree, despite its unmatched interpretability and lightweight structure, faces two key issues that limit its broader applicability: non-differentiability and low testing accuracy. This study addresses these issues by developing a differentiable oblique tree that optimizes the entire tree using gradient-based optimization. We propose an exact reformulation of hard-split trees based on "ReLU+Argmin" mechanism, and then cast the reformulated tree training as an unconstrained optimization task. The ReLU-based sample branching, expressed as exact-zero or non-zero values, preserve a unique decision path, in contrast to soft decision trees with probabilistic routing. The subsequent Argmin operation identifies the unique zero-violation path, enabling deterministic predictions. For effective gradient flow, we approximate Argmin behaviors by scaling softmin function. To ameliorate numerical instability, we propose a warm-start annealing scheme that solves multiple optimization tasks with increasingly accurate approximations. This reformulation alongside distributed GPU parallelism offers strong scalability, supporting 12-depth tree even on million-scale datasets where most baselines fail. Extensive experiments demonstrate that our optimized tree achieves a superior testing accuracy against 14 baselines, including an average improvement of 7.54\% over CART.

Many \textit{Deep Reinforcement Learning} (DRL) algorithms are sensitive to time discretization, which reduces their performance in real-world scenarios. We propose Continuous Soft Actor-Critic, an off-policy actor-critic DRL algorithm in continuous time and space. It is robust to environment time discretization. We also extend the framework to multi-agent scenarios. This \textit{Multi-Agent Reinforcement Learning} (MARL) algorithm is suitable for both competitive and cooperative settings. Policy evaluation employs stochastic control theory, with loss functions derived from martingale orthogonality conditions. We establish scaling principles for hyperparameters of the algorithm as the environment time discretization $\delta t$ changes ($\delta t \rightarrow 0$). We provide theoretical proofs for the relevant theorems. To validate the algorithm's effectiveness, we conduct comparative experiments between the proposed algorithm and other mainstream methods across multiple tasks in \textit{Virtual Multi-Agent System} (VMAS). Experimental results demonstrate that the proposed algorithm achieves robust performance across various environments with different time discretization parameter settings, outperforming other methods.

Training multi-modal large language models (MLLMs) that align with human intentions is a long-term challenge. Traditional score-only reward models for alignment suffer from low accuracy, weak generalization, and poor interpretability, blocking the progress of alignment methods, \textit{e.g.,} reinforcement learning from human feedback (RLHF). Generative reward models (GRMs) leverage MLLMs' intrinsic reasoning capabilities to discriminate pair-wise responses, but their pair-wise paradigm makes it hard to generalize to learnable rewards. We introduce Generative RLHF-V, a novel alignment framework that integrates GRMs with multi-modal RLHF. We propose a two-stage pipeline: \textbf{multi-modal generative reward modeling from RL}, where RL guides GRMs to actively capture human intention, then predict the correct pair-wise scores; and \textbf{RL optimization from grouped comparison}, which enhances multi-modal RL scoring precision by grouped responses comparison. Experimental results demonstrate that, besides out-of-distribution generalization of RM discrimination, our framework improves 4 MLLMs' performance across 7 benchmarks by 18.1\%, while the baseline RLHF is only 5.3\%. We further validate that Generative RLHF-V achieves a near-linear improvement with an increasing number of candidate responses.

Learning from human involvement aims to incorporate the human subject to monitor and correct agent behavior errors. Although most interactive imitation learning methods focus on correcting the agent’s action at the current state, they do not adjust its actions in future states, which may be potentially more hazardous. To address this, we introduce Predictive Preference Learning from Human Interventions (PPL), which leverages the implicit preference signals contained in human interventions to inform predictions of future rollouts. The key idea of PPL is to bootstrap each human intervention into L future time steps, called the preference horizon, with the assumption that the agent follows the same action and the human makes the same intervention in the preference horizon. By applying preference optimization on these future states, expert corrections are propagated into the safety-critical regions where the agent is expected to explore, significantly improving learning efficiency and reducing human demonstrations needed. We evaluate our approach with experiments on both autonomous driving and robotic manipulation benchmarks and demonstrate its efficiency and generality. Our theoretical analysis further shows that selecting an appropriate preference horizon L balances coverage of risky states with label correctness, thereby bounding the algorithmic optimality gap. Demo and code are available at: https://metadriverse.github.io/ppl.

We study how to fine-tune LLMs using user-edit deployment data consisting of a set of context, an agent's response, and user edits. This deployment data is naturally generated by users in applications such as LLMs-based writing assistants and coding agents. The _natural_ origin of user edits makes it a desired source for adapting and personalizing of LLMs. In this setup, there emerges a unification of various feedback types namely preferences, supervised labels, and cost that are typically studied separately in the literature. In this paper, we initiate the theoretical investigation of learning from user edits. We first derive bounds for learning algorithms that learn from each of these feedback types. We prove that these algorithms have different trade-offs depending upon the user, data distribution, and model class. We then propose a simple ensembling procedure to jointly learn from these feedback types. On two domains from Gao et al. 2024, we show our ensembling procedure outperforms these methods that learn from individual feedback. Further, we show that our proposed procedure can robustly adapt to different user-edit distributions at test time.

Despite recent progress in large-scale reinforcement learning (RL) for reasoning, the training recipe for building high-performing reasoning models remains elusive. Key implementation details of frontier models, such as DeepSeek-R1, including data curation strategies and RL training recipe, are often omitted. Moreover, recent research indicates distillation remains more effective than RL for smaller models. In this work, we demonstrate that large-scale RL can significantly enhance the reasoning capabilities of strong, small- and mid-sized models, achieving results that surpass those of state-of-the-art distillation-based models. We systematically study the RL training process through extensive ablations and propose a simple yet effective approach: first training on math-only prompts, then on code-only prompts. Notably, we find that math-only RL not only significantly enhances the performance of strong distilled models on math benchmarks (e.g., +14.6\% / +17.2\% on AIME 2025 for the 7B / 14B models), but also code reasoning tasks (e.g., +6.8\% / +5.8\% on LiveCodeBench for the 7B / 14B models). In addition, extended code-only RL further improves code benchmark performance while causing minimal degradation in math results. We develop a robust data curation pipeline to collect challenging prompts with high-quality, verifiable answers and test cases to enable verification-based RL across both domains. The dataset will be released to support open research. Finally, we identify key experimental insights, including curriculum learning with progressively increasing response lengths and the stabilizing effect of on-policy parameter updates. We find that RL not only elicits the foundational reasoning capabilities acquired during pretraining and supervised fine-tuning (SFT), but also pushes the limits of the model’s reasoning ability, enabling it to solve problems that were previously unsolvable.

The rapid spread of multimodal misinformation on social media has raised growing concerns, while research on video misinformation detection remains limited due to the lack of large-scale, diverse datasets. Existing methods often overfit to rigid templates and lack deep reasoning over deceptive content. To address these challenges, we introduce FakeVV, a large-scale benchmark comprising over 100,000 video-text pairs with fine-grained, interpretable annotations. In addition, we further propose Fact-R1, a novel framework that integrates deep reasoning with collaborative rule-based reinforcement learning. Fact-R1 is trained through a three-stage process: (1) misinformation long-Chain-of-Thought (CoT) instruction tuning, (2) preference alignment via Direct Preference Optimization (DPO), and (3) Group Relative Policy Optimization (GRPO) using a novel verifiable reward function. This enables Fact-R1 to exhibit emergent reasoning behaviors comparable to those observed in advanced text-based reinforcement learning systems, but in the more complex multimodal misinformation setting. Our work establishes a new paradigm for misinformation detection, bridging large-scale video understanding, reasoning-guided alignment, and interpretable verification.

Effectively retrieving, reasoning and understanding visually rich information remains a challenge for traditional Retrieval-Augmented Generation (RAG) methods. On the one hand, traditional text-based methods cannot handle visual-related information. On the other hand, current vision-based RAG approaches are often limited by fixed pipelines and frequently struggle to reason effectively due to the insufficient activation of the fundamental capabilities of models. As reinforcement learning (RL) has been proven to be beneficial for model reasoning, we introduce VRAG-RL, a novel RL framework tailored for complex reasoning across visually rich information. With this framework, VLMs interact with search engines, autonomously sampling single-turn or multi-turn reasoning trajectories with the help of visual perception tokens and undergoing continual optimization based on these samples. Our approach highlights key limitations of RL in RAG domains: (i) Prior Multi-modal RAG approaches tend to merely incorporate images into the context, leading to insufficient reasoning token allocation and neglecting visual-specific perception; and (ii) When models interact with search engines, their queries often fail to retrieve relevant information due to the inability to articulate requirements, thereby leading to suboptimal performance. To address these challenges, we define an action space tailored for visually rich inputs, with actions including cropping and scaling, allowing the model to gather information from a coarse-to-fine perspective. Furthermore, to bridge the gap between users' original inquiries and the retriever, we employ a simple yet effective reward that integrates query rewriting and retrieval performance with a model-based reward. Our VRAG-RL optimizes VLMs for RAG tasks using specially designed RL strategies, aligning the model with real-world applications. Extensive experiments on diverse and challenging benchmarks show that our VRAG-RL outperforms existing methods by 20\% (Qwen2.5-VL-7B) and 30\% (Qwen2.5-VL-3B), demonstrating the effectiveness of our approach. The code is available at https://github.com/Alibaba-NLP/VRAG.

Recent work on neural scaling laws demonstrates that model performance scales predictably with compute budget, model size, and dataset size. In this work, we develop scaling laws based on problem complexity. We analyze two fundamental complexity measures: solution space size and representation space size. Using the Traveling Salesman Problem (TSP) as a case study, we show that combinatorial optimization promotes smooth cost trends, and therefore meaningful scaling laws can be obtained even in the absence of an interpretable loss. We then show that suboptimality grows predictably for fixed-size models when scaling the number of TSP nodes or spatial dimensions, independent of whether the model was trained with reinforcement learning or supervised fine-tuning on a static dataset. We conclude with an analogy to problem complexity scaling in local search, showing that a much simpler gradient descent of the cost landscape produces similar trends.

Data assimilation (DA) aims to estimate the full state of a dynamical system by combining partial and noisy observations with a prior model forecast, commonly referred to as the background. In atmospheric applications, this problem is fundamentally ill-posed due to the sparsity of observations relative to the high-dimensional state space. Traditional methods address this challenge by simplifying background priors to regularize the solution, which are empirical and require continual tuning for application. Inspired by alignment techniques in text-to-image diffusion models, we propose Align-DA, which formulates DA as a generative process and uses reward signals to guide background priors—replacing manual tuning with data-driven alignment. Specifically, we train a score-based model in the latent space to approximate the background-conditioned prior, and align it using three complementary reward signals for DA: (1) assimilation accuracy, (2) forecast skill initialized from the assimilated state, and (3) physical adherence of the analysis fields. Experiments with multiple reward signals demonstrate consistent improvements in analysis quality across different evaluation metrics and observation-guidance strategies. These results show that preference alignment, implemented as a soft constraint, can automatically adapt complex background priors tailored to DA, offering a promising new direction for advancing the field.

Modern reinforcement learning (RL) algorithms have found success by using powerful probabilistic models, such as transformers, energy-based models, and diffusion/flow-based models. To this end, RL researchers often choose to pay the price of accommodating these models into their algorithms -- diffusion models are expressive, but are computationally intensive due to their reliance on solving differential equations, while autoregressive transformer models are scalable but typically require learning discrete representations. Normalizing flows (NFs), by contrast, seem to provide an appealing alternative, as they enable likelihoods and sampling without solving differential equations or autoregressive architectures. However, their potential in RL has received limited attention, partly due to the prevailing belief that normalizing flows lack sufficient expressivity. We show that this is not the case. Building on recent work in NFs, we propose a single NF architecture which integrates seamlessly into RL algorithms, serving as a policy, Q-function, and occupancy measure. Our approach leads to much simpler algorithms, and achieves higher performance in imitation learning, offline, goal conditioned RL and unsupervised RL.

Synthetic data is crucial for advancing autonomous driving (AD) systems, yet current state-of-the-art video generation models, despite their visual realism, suffer from subtle geometric distortions that limit their utility for downstream perception tasks. We identify and quantify this critical issue, demonstrating a significant performance gap in 3D object detection when using synthetic versus real data. To address this, we introduce Reinforcement Learning with Geometric Feedback (RLGF), RLGF uniquely refines video diffusion models by incorporating rewards from specialized latent-space AD perception models. Its core components include an efficient Latent-Space Windowing Optimization technique for targeted feedback during diffusion, and a Hierarchical Geometric Reward (HGR) system providing multi-level rewards for point-line-plane alignment, and scene occupancy coherence. To quantify these distortions, we propose GeoScores. Applied to models like DiVE on nuScenes, RLGF substantially reduces geometric errors (e.g., VP error by 21\%, Depth error by 57\%) and dramatically improves 3D object detection mAP by 12.7\%, narrowing the gap to real-data performance. RLGF offers a plug-and-play solution for generating geometrically sound and reliable synthetic videos for AD development.

Preference-based reinforcement learning (PbRL) bypasses complex reward engineering by learning from human feedback. However, due to the high cost of obtaining feedback, PbRL typically relies on a limited set of preference-labeled samples. This data scarcity introduces two key inefficiencies: (1) the reward model overfits to the limited feedback, leading to poor generalization to unseen samples, and (2) the agent exploits the learned reward model, exacerbating overestimation of action values in temporal difference (TD) learning. To address these issues, we propose STAR, an efficient PbRL method that integrates preference margin regularization and policy regularization. Preference margin regularization mitigates overfitting by introducing a bounded margin in reward optimization, preventing excessive bias toward specific feedback. Policy regularization bootstraps a conservative estimate $\widehat{Q}$ from well-supported state-action pairs in the replay memory, reducing overestimation during policy learning. Experimental results show that STAR improves feedback efficiency, achieving 34.8\% higher performance in online settings and 29.7\% in offline settings compared to state-of-the-art methods. Ablation studies confirm that STAR facilitates more robust reward and value function learning. The videos of this project are released at https://sites.google.com/view/pbrl-star.

Recent endeavors in Multimodal Large Language Models (MLLMs) aim to unify visual comprehension and generation. However, these two capabilities remain largely independent, as if they are two separate functions encapsulated within the same model. Consequently, visual comprehension does not enhance visual generation, and the reasoning mechanisms of LLMs have not been fully integrated to revolutionize image generation. In this paper, we propose to enable the collaborative co-evolution of visual comprehension and generation, advancing image generation into an iterative introspective process. We introduce a two-stage training approach: supervised fine-tuning teaches the MLLM with the foundational ability to generate genuine CoT for visual generation, while reinforcement learning activates its full potential via an exploration-exploitation trade-off. Ultimately, we unlock the Aha moment in visual generation, advancing MLLMs from text-to-image tasks to unified image generation. Extensive experiments demonstrate that our model not only excels in text-to-image generation and image editing, but also functions as a superior image semantic evaluator with enhanced visual comprehension capabilities. Project Page: \url{https://janus-pro-r1.github.io}.