NeurIPS 2025 Papers

The rapid advancement of native multi-modal models and omni-models, exemplified by GPT-4o, Gemini and o3 with their capability to process and generate content across modalities such as text and images, marks a significant milestone in the evolution of intelligence. Systematic evaluation of their multi-modal output capabilities in visual thinking process (a.k.a., multi-modal chain of thought, M-CoT) becomes critically important. However, existing benchmarks for evaluating multi-modal models primarily focus on assessing multi-modal inputs and text-only reasoning process while neglecting the importance of reasoning through multi-modal outputs. In this paper, we present a benchmark, dubbed as RBench-V, designed to assess models’ vision-indispensable reasoning. To conduct RBench-V, we carefully hand-pick 803 questions covering math, physics, counting and games. Unlike problems in previous benchmarks, which typically specify certain input modalities, RBench-V presents problems centered on multi-modal outputs, which require image manipulation, such as generating novel images and constructing auxiliary lines to support reasoning process. We evaluate numerous open- and closed-source models on RBench-V, including o3, Gemini 2.5 pro, Qwen2.5-VL, etc. Even the best-performing model, o3, achieves only 25.8% accuracy on RBench-V, far below the human score of 82.3%, which shows current models struggle to leverage multi-modal reasoning. Data and code are available at https://evalmodels.github.io/rbenchv.

Urban computing harnesses big data to decode complex urban dynamics and revolutionize location-based services. Traditional approaches have treated geospatial prediction tasks (e.g., estimating socio-economic indicators) and retrieval tasks (e.g., querying geographic objects) as isolated challenges, necessitating separate models with distinct training objectives. This fragmentation imposes significant computational burdens and limits cross-task synergy, despite advances in representation learning and multi-task foundation models. We present UrbanSparse, a pioneering framework that unifies geospatial prediction and retrieval through a novel sparse-dense representation architecture. By synergistically combining these tasks, UrbanSparse eliminates redundant systems while amplifying their mutual strengths. Our approach introduces two innovations: (1) Bloom filter-based sparse encodings that compress high-sparsity geographic queries and fine-grained text terms for retrieval effectiveness, and (2) a dense semantic codebook that captures granular urban features to boost prediction accuracy. A two-view contrastive learning mechanism further bridges urban objects, regions, and contexts. Experiments on real-world datasets demonstrate 25.16% gains in prediction accuracy and 20.76% improvements in retrieval precision over state-of-the-art baselines, alongside 65.97% faster training. These advantages position UrbanSparse as a scalable solution for large urban datasets. To our knowledge, this is the first unified framework bridging geospatial prediction and retrieval, opening new frontiers in data-driven urban intelligence.

Targeted maximum likelihood estimation (TMLE) is a widely used debiasing algorithm for plug-in estimation. While its statistical guarantees, such as double robustness and asymptotic efficiency, are well-studied, the convergence properties of TMLE as an iterative optimization scheme have remained underexplored. To bridge this gap, we study TMLE’s iterative updates through an optimization-theoretic lens, establishing global convergence under standard assumptions and regularity conditions. We begin by providing the first complete characterization of different stopping criteria and their relationship to convergence in TMLE. Next, we provide geometric insights. We show that each submodel induces a smooth, non-selfintersecting path (homotopy) through the probability simplex. We then analyze the solution space of the estimating equation and loss landscape. We show that all valid solutions form a submanifold of the statistical model, with the difference in dimension (i.e., codimension) exactly matching the dimension of the target parameter. Building on these geometric insights, we deliver the first strict proof of TMLE’s convergence from an optimization viewpoint, as well as explicit sufficient criteria under which TMLE terminates in a single update. As a by-product, we discover an unidentified overshooting phenomenon wherein the algorithm can surpass feasible roots to the estimating equation along a homotopy path, highlighting a promising avenue for designing enhanced debias algorithms.

Linear and quadratic optimization are crucial in numerous real-world applications, ranging from training machine learning models to solving integer linear programs. Recently, learning-to-optimize methods (L2O) for linear (LPs) or quadratic programs (QPs) using message-passing graph neural networks (MPNNs) have gained traction, promising lightweight, data-driven proxies for solving such optimization problems. For example, they replace the costly computation of strong branching scores in branch-and-bound solvers, thereby reducing the need to solve many such optimization problems. However, robust L2O MPNNs remain challenging in data-scarce settings, especially when addressing complex optimization problems such as QPs. This work introduces a principled approach to data augmentation tailored for QPs via MPNNs. Our method leverages theoretically justified data augmentation techniques to generate diverse yet optimality-preserving instances. Furthermore, we integrate these augmentations into a self-supervised contrastive learning framework, thereby pretraining MPNNs for improved performance on L2O tasks. Extensive experiments demonstrate that our approach improves generalization in supervised scenarios and facilitates effective transfer learning to related optimization problems.

The advancement of image style transfer has been fundamentally constrained by the absence of large-scale, high-quality datasets with explicit content-style-stylized supervision. Existing methods predominantly adopt training-free paradigms (e.g., image inversion), which limit controllability and generalization due to the lack of structured triplet data. To bridge this gap, we design a scalable and automated pipeline that constructs and purifies high-fidelity content-style-stylized image triplets. Leveraging this pipeline, we introduce IMAGStyle—the first large-scale dataset of its kind, containing 210K diverse and precisely aligned triplets for style transfer research. Empowered by IMAGStyle, we propose CSGO, a unified, end-to-end trainable framework that decouples content and style representations via independent feature injection. CSGO jointly supports image-driven style transfer, text-driven stylized generation, and text-editing-driven stylized synthesis within a single architecture. Extensive experiments show that CSGO achieves state-of-the-art controllability and fidelity, demonstrating the critical role of structured synthetic data in unlocking robust and generalizable style transfer. Source code: \url{https://github.com/instantX-research/CSGO}

LLM-based agent applications have shown increasingly remarkable capabilities in complex workflows but incur substantial costs and latency due to extensive planning and reasoning requirements. Existing LLM caching techniques (like context caching and semantic caching), primarily designed for serving chatbots, are insufficient for agent applications where outputs depend on external data and environmental contexts. We propose **Agentic Plan Caching (APC)**, a novel **test-time memory** that extracts, stores, adapts, and reuses structured plan templates from planning stages of agent applications across semantically similar tasks to reduce the cost and latency of serving. Unlike traditional semantic caching, our system extracts plan templates from completed agent executions at test-time, employs keyword extraction to match new requests against cached plans, and utilizes lightweight models to adapt these templates to task-specific plans with contexts. Evaluation across multiple real-world agent applications shows that our system can reduce costs by 50.31\% and latency by 27.28\% on average while maintaining performance, offering a more efficient solution for serving LLM-based agents that complements existing LLM serving infrastructures.

Virtual footwear try-on (VFTON), a critical yet underexplored area in virtual try-on (VTON), aims to synthesize faithful try-on results given diverse footwear and model images while maintaining 3D consistency and texture authenticity. Unlike conventional garment-focused VTON methods, VFTON presents unique challenges due to (1) Data Scarcity, which arises from the difficulty of perfectly matching product shoes with models wearing the identical ones, (2) Viewpoint Misalignment, where the target foot pose and source shoe views are always misaligned, leading to incomplete texture information and detail distortion, and (3) Background-induced Color Distortion, where complex material of footwear interacts with environmental lighting, causing unintended color contamination. To address these challenges, we introduce MVShoes, a multi-view shoe try-on dataset consisting of 7305 well-annotated image triplets, covering diverse footwear categories and challenging try-on scenarios. Furthermore, we propose a dual-stream DiT architecture, ShoeFit, designed to mitigate viewpoint misalignment through Multi-View Conditioning with 3D Rotary Position Embedding, and alleviate background-induced distortion using the LayeredRefAttention which leverages background features to modulate footwear latents. The proposed framework effectively decouples shoe appearance from environmental interferences while preserving high-quality texture detail through decoupled denoising and conditioning branches. Extensive quantitative and qualitative experiments demonstrate that our method substantially improves rendering fidelity and robustness under challenging real-world product shoes, establishing a new benchmark in high-fidelity footwear try-on synthesis. The dataset and benchmark will be publicly available upon acceptance of the paper.

This work presents Orient Anything V2, an enhanced foundation model for unified understanding of object 3D orientation and rotation from single or paired images. Building upon Orient Anything V1, which defines orientation via a single unique front face, V2 extends this capability to handle objects with diverse rotational symmetries and directly estimate relative rotations. These improvements are enabled by four key innovations: 1) Scalable 3D assets synthesized by generative models, ensuring broad category coverage and balanced data distribution; 2) An efficient, model-in-the-loop annotation system that robustly identifies 0 to N valid front faces for each object; 3) A symmetry-aware, periodic distribution fitting objective that captures all plausible front-facing orientations, effectively modeling object rotational symmetry; 4) A multi-frame architecture that directly predicts relative object rotations. Extensive experiments show that Orient Anything V2 achieves state-of-the-art zero-shot performance on orientation estimation, 6DoF pose estimation, and object symmetry recognition across 11 widely used benchmarks. The model demonstrates strong generalization, significantly broadening the applicability of orientation estimation in diverse downstream tasks.

It has been recently observed in much of the literature that neural networks exhibit a bottleneck rank property: for larger depths, the activation and weights of neural networks trained with gradient-based methods tend to be of approximately low rank. In fact, the rank of the activations of each layer converges to a fixed value referred to as the ``bottleneck rank", which is the minimum rank required to represent the training data. This perspective is in line with the observation that regularizing linear networks (without activations) with weight decay is equivalent to minimizing the Schatten $p$ quasi norm of the neural network. In this paper we investigate the implications of this phenomenon for generalization. More specifically, we prove generalization bounds for neural networks which exploit the approximate low rank structure of the weight matrices if present. The final results rely on the Schatten $p$ quasi norms of the weight matrices: for small p, the bounds exhibit a sample complexity $ \widetilde{O}(WrL^2)$ where $W$ and $L$ are the width and depth of the neural network respectively and where $r$ is the rank of the weight matrices. As $p$ increases, the bound behaves more like a norm-based bound instead. The proof techniques involve a careful interpolation between the parametric and norm based regimes. We also demonstrate in experiments that this bound outperforms both classic parameter counting and norm based bounds in the typical overparametrized regime.

Inverse design aims to design the input variables of a physical system to optimize a specified objective function, typically formulated as a search or optimization problem. However, in 3D domains, the design space grows exponentially, rendering exhaustive grid-based searches infeasible. Recent advances in deep learning have accelerated inverse design by providing powerful generative priors and differentiable surrogate models. Nevertheless, current methods tend to approximate the 3D design space using 2D projections or fine-tune existing 3D shapes. These approaches sacrifice volumetric detail and constrain design exploration, preventing true 3D design from scratch. In this paper, we propose a 3D Inverse Design (3DID) framework that directly navigates the 3D design space by coupling a continuous latent representation with a physics-aware optimization strategy. We first learn a unified physics–geometry embedding that compactly captures shape and physical field data in a continuous latent space. Then, we introduce a two-stage strategy to perform physics-aware optimization. In the first stage, a gradient-guided diffusion sampler explores the global latent manifold. In the second stage, an objectivedriven, topology-preserving refinement further sculpts each candidate toward the target objective. This enables 3DID to generate high-fidelity 3D geometries, outperforming existing methods in both solution quality and design versatility.

Predicting disease risk from DNA presents an unprecedented emerging challenge as biobanks approach population scale sizes ($N>10^6$ individuals) with ultra-high-dimensional features ($L>10^5$ genotypes). Current methods, often linear and reliant on summary statistics, fail to capture complex genetic interactions and discard valuable individual-level information. We introduce **PRSformer**, a scalable deep learning architecture designed for end-to-end, multitask disease prediction directly from million-scale individual genotypes. PRSformer employs neighborhood attention, achieving linear $O(L)$ complexity per layer, making Transformers tractable for genome-scale inputs. Crucially, PRSformer utilizes a stacking of these efficient attention layers, progressively increasing the effective receptive field to model local dependencies (e.g., within linkage disequilibrium blocks) before integrating information across wider genomic regions. This design, tailored for genomics, allows PRSformer to learn complex, potentially non-linear and long-range interactions directly from raw genotypes. We demonstrate PRSformer's effectiveness using a unique large private cohort ($N \approx 5$M) for predicting 18 autoimmune and inflammatory conditions using $L \approx 140$k variants. PRSformer significantly outperforms highly optimized linear models trained on the *same individual-level data* and state-of-the-art summary-statistic-based methods (LDPred2) derived from the *same cohort*, quantifying the benefits of non-linear modeling and multitask learning at scale. Furthermore, experiments reveal that the advantage of non-linearity emerges primarily at large sample sizes ($N > 1$M), and that a multi-ancestry trained model improves generalization, establishing PRSformer as a new framework for deep learning in population-scale genomics.

Graph Neural Networks (GNNs) have achieved strong performance across a range of graph representation learning tasks, yet their adversarial robustness in graph classification remains underexplored compared to node classification. While most existing defenses focus on the message-passing component, this work investigates the overlooked role of pooling operations in shaping robustness. We present a theoretical analysis of standard flat pooling methods (sum, average and max), deriving upper bounds on their adversarial risk and identifying their vulnerabilities under different attack scenarios and graph structures. Motivated by these insights, we propose Robust Singular Pooling (RS-Pool), a novel pooling strategy that leverages the dominant singular vector of the node embedding matrix to construct a robust graph-level representation. We theoretically investigate the robustness of RS-Pool and interpret the resulting bound leading to improved understanding of our proposed pooling operator. While our analysis centers on Graph Convolutional Networks (GCNs), RS-Pool is model-agnostic and can be implemented efficiently via power iteration. Empirical results on real-world benchmarks show that RS-Pool provides better robustness than the considered pooling methods when subject to state-of-the-art adversarial attacks while maintaining competitive clean accuracy.

We consider matroid allocation problems under \textit{opportunity fairness} constraints: resources need to be allocated to a set of agents under matroid constraints (which includes classical problems such as bipartite matching). Agents are divided into $C$ groups according to a sensitive attribute, and an allocation is opportunity-fair if each group receives the same share proportional to the maximum feasible allocation it could achieve in isolation. We study the Price of Fairness (PoF), i.e., the ratio between maximum size allocations and maximum size opportunity-fair allocations. We first provide a characterization of the PoF leveraging the underlying polymatroid structure of the allocation problem. Based on this characterization, we prove bounds on the PoF in various settings from fully adversarial (worst-case) to fully random. Notably, one of our main results considers an arbitrary matroid structure with agents randomly divided into groups. In this setting, we prove a PoF bound as a function of the (relative) size of the largest group. Our result implies that, as long as there is no dominant group (i.e., the largest group is not too large), opportunity fairness constraints do not induce any loss of social welfare (defined as the allocation size). Overall, our results give insights into which aspects of the problem's structure affect the trade-off between opportunity fairness and social welfare.

Transformers can implement both generalizable algorithms (e.g., induction heads) and simple positional shortcuts (e.g., memorizing fixed output positions). In this work, we study how the choice of pretraining data distribution steers a shallow transformer toward one behavior or the other. Focusing on a minimal trigger-output prediction task -- copying the token immediately following a special trigger upon its second occurrence -- we present a rigorous analysis of gradient-based training of a single-layer transformer. In both the infinite and finite sample regimes, we prove a transition in the learned mechanism: if input sequences exhibit sufficient diversity, measured by a low “max-sum” ratio of trigger-to-trigger distances, the trained model implements an induction head and generalizes to unseen contexts; by contrast, when this ratio is large, the model resorts to a positional shortcut and fails to generalize out-of-distribution (OOD). We also reveal a trade-off between the pretraining context length and OOD generalization, and derive the optimal pretraining distribution that minimizes computational cost per sample. Finally, we validate our theoretical predictions with controlled synthetic experiments, demonstrating that broadening context distributions robustly induces induction heads and enables OOD generalization. Our results shed light on the algorithmic biases of pretrained transformers and offer conceptual guidelines for data-driven control of their learned behaviors.

Recent advances in multimodal training have significantly improved the integration of image understanding and generation within a unified model. This study investigates how vision-language models (VLMs) handle image-understanding tasks, focusing on how visual information is processed and transferred to the textual domain. We compare *native multimodal VLMs*, models trained from scratch on multimodal data to generate both text and images, and *non-native multimodal VLMs*, models adapted from pre-trained large language models or capable of generating only text, highlighting key differences in information flow. We find that in native multimodal VLMs, image and text embeddings are more separated within the residual stream. Moreover, VLMs differ in how visual information reaches text: non-native multimodal VLMs exhibit a distributed communication pattern, where information is exchanged through multiple image tokens, whereas models trained natively for joint image and text generation tend to rely on a single post-image token that acts as a *narrow gate* for visual information. We show that ablating this single token significantly deteriorates image-understanding performance, whereas targeted, token-level interventions reliably steer image semantics and downstream text with fine-grained control.

With the increasing adoption of Retrieval-Augmented Generation (RAG) systems for knowledge-intensive tasks, ensuring the adequacy of retrieved documents has become critically important for generation quality. Traditional reranking approaches face three significant challenges: substantial computational overhead that scales with document length, dependency on plain text that limits application in sensitive scenarios, and insufficient assessment of document value beyond simple relevance metrics. We propose EAReranker, an efficient embedding-based adequacy assessment framework that evaluates document utility for RAG systems without requiring access to original text content. The framework quantifies document adequacy through a comprehensive scoring methodology considering verifiability, coverage, completeness and structural aspects, providing interpretable adequacy classifications for downstream applications. EAReranker employs a Decoder-Only Transformer architecture that introduces embedding dimension expansion method and bin-aware weighted loss, designed specifically to predict adequacy directly from embedding vectors. Our comprehensive evaluation across four public benchmarks demonstrates that EAReranker achieves competitive performance with state-of-the-art plaintext rerankers while maintaining constant memory usage ($\sim$550MB) regardless of input length and processing 2-3x faster than traditional approaches. The semantic bin adequacy prediction accuracy of 92.85\% LACC@10 and 86.12\% LACC@25 demonstrates its capability to effectively filter out inadequate documents that could potentially mislead or adversely impact RAG system performance, thereby ensuring only high-utility information serves as generation context. These results establish EAReranker as an efficient and practical solution for enhancing RAG system performance through improved context selection while addressing the computational and privacy challenges of existing methods.

Reinforcement learning (RL) has emerged as a powerful post-training technique to incentivize the reasoning ability of large language models (LLMs). However, LLMs can respond very inconsistently to RL finetuning: some show substantial performance gains, while others plateau or even degrade. To understand this divergence, we analyze the per-step influence of the RL objective and identify two key conditions for effective post-training: (1) RL-informative rollout accuracy, and (2) strong data co-influence, which quantifies how much the training data affects performance on other samples. Guided by these insights, we propose behavior injection, a task-agnostic data augmentation scheme applied prior to RL. Behavior injection enriches the supervised finetuning (SFT) data by seeding exploratory and exploitative behaviors, effectively making the model more RL-ready. We evaluate our method across two reasoning benchmarks with multiple base models. The results demonstrate that our theoretically motivated augmentation can significantly increase the performance gain from RL over the pre-RL model.

The modeling of samples of distributions is a major challenge since distributions do not form a vector space. While various approaches exist for univariate distributions, including transformations to a Hilbert space, far less is known about the multivariate case. We utilize a transformation approach to map multivariate distributions to a Hilbert space via a Wasserstein slicing method that is invertible. This approach combines functional data analysis tools, such as functional principal component analysis and modes of variation, with the facility to map back to interpretable distributions. We also provide convergence guarantees for the Hilbert space representations under a broad class of such transforms. The method is illustrated using joint systolic and diastolic blood pressure data.

While large language models (LLMs) with Chain-of-Thought (CoT) reasoning excel in mathematics and coding, their potential for systematic reasoning in chemistry, a domain demanding rigorous structural analysis for real-world tasks like drug design and reaction engineering, remains untapped. Current benchmarks focus on simple knowledge retrieval, neglecting step-by-step reasoning required for complex tasks such as molecular optimization and reaction prediction. To address this, we introduce ChemCoTBench, a reasoning framework that bridges molecular structure understanding with arithmetic-inspired operations, including addition, deletion, and substitution, to formalize chemical problem-solving into transparent, step-by-step workflows. By treating molecular transformations as modular "chemical operations", the framework enables slow-thinking reasoning, mirroring the logic of mathematical proofs while grounding solutions in real-world chemical constraints. We evaluate models on two high-impact tasks: Molecular Property Optimization and Chemical Reaction Prediction. These tasks mirror real-world challenges while providing structured evaluability. We further provide ChemCoTDataset, a pioneering 22,000-instance chemical reasoning dataset with expert-annotated chains of thought to facilitate LLM fine-tuning. By providing annotated trainable datasets, a reasoning taxonomy, and baseline evaluations, our work bridges the gap between abstract reasoning methods and practical chemical discovery, establishing a foundation for advancing LLMs as tools for AI-driven scientific innovation.