NeurIPS 2025 Papers

With the significant progress of large reasoning models in complex coding and reasoning tasks, existing benchmarks, like LiveCodeBench and CodeElo, are insufficient to evaluate the coding capabilities of large language models (LLMs) in real competition environments. Moreover, current evaluation metrics such as Pass@K fail to capture the reflective abilities of reasoning models. To address these challenges, we propose ICPC-Eval, a top-level competitive coding benchmark designed to probing the frontiers of LLM reasoning. ICPC-Eval includes 118 carefully curated problems from 11 recent ICPC contests held in various regions of the world, offering three key contributions: 1) A challenging realistic ICPC competition scenario, featuring a problem type and difficulty distribution consistent with actual contests. 2) A robust test case generation method and a corresponding local evaluation toolkit, enabling efficient and accurate local evaluation. 3) An effective test-time scaling evaluation metric, Refine@K, which allows iterative repair of solutions based on execution feedback. The results underscore the significant challenge in evaluating complex reasoning abilities: top-tier reasoning models like DeepSeek-R1 often rely on multi-turn code feedback to fully unlock their in-context reasoning potential when compared to non-reasoning counterparts. Furthermore, despite recent advancements in code generation, these models still lag behind top-performing human teams. We release the benchmark at: https://github.com/RUCAIBox/ICPC-Eval

Diffusion models have demonstrated remarkable capabilities in generating high-quality images. Recent advancements in Layout-to-Image (L2I) generation have leveraged positional conditions and textual descriptions to facilitate precise and controllable image synthesis. Despite overall progress, current L2I methods still exhibit suboptimal performance. Therefore, we propose InstanceAssemble, a novel architecture that incorporates layout conditions via instance-assembling attention, enabling position control with bounding boxes (bbox) and multimodal content control including texts and additional visual content. Our method achieves flexible adaption to existing DiT-based T2I models through light-weighted LoRA modules. Additionally, we propose a Layout-to-Image benchmark, Denselayout, a comprehensive benchmark for layout-to-image generation, containing 5k images with 90k instances in total. We further introduce Layout Grounding Score (LGS), an interpretable evaluation metric to more precisely assess the accuracy of L2I generation. Experiments demonstrate that our InstanceAssemble method achieves state-of-the-art performance under complex layout conditions, while exhibiting strong compatibility with diverse style LoRA modules. The code and pretrained models are publicly available at \url{https://github.com/FireRedTeam/InstanceAssemble}.

We present Eve3D, a novel framework for dense 3D reconstruction based on 3D Gaussian Splatting (3DGS). While most existing methods rely on imperfect priors derived from pre-trained vision models, Eve3D fully leverages these priors by jointly optimizing both them and the 3DGS backbone. This joint optimization creates a mutually reinforcing cycle: the priors enhance the quality of 3DGS, which in turn refines the priors, further improving the reconstruction. Additionally, Eve3D introduces a novel optimization step based on bundle adjustment, overcoming the limitations of the highly local supervision in standard 3DGS pipelines. Eve3D achieves state-of-the-art results in surface reconstruction and novel view synthesis on the Tanks & Temples, DTU, and Mip-NeRF360 datasets. while retaining fast convergence, highlighting an unprecedented trade-off between accuracy and speed.

We introduce EvaLearn, a pioneering benchmark designed to evaluate large language models (LLMs) on their learning capability and efficiency in challenging tasks, a critical, yet underexplored aspect of model potential. EvaLearn contains 648 challenging problems across six task types, grouped into 182 sequences, each sequence dedicated to one task type. Diverging from most existing benchmarks that evaluate models in parallel, EvaLearn requires models to solve problems sequentially, allowing them to leverage the experience gained from previous solutions. EvaLearn provides five comprehensive automated metrics to evaluate models and quantify their learning capability and efficiency. We extensively benchmark nine frontier models and observe varied performance profiles: some models, such as Claude-3.7-sonnet, start with moderate initial performance but exhibit strong learning ability, while some models struggle to benefit from experience and may even show negative transfer. Moreover, we investigate model performance under two learning settings and find that instance-level rubrics and teacher-model feedback further facilitate model learning. Importantly, we observe that current LLMs with stronger static abilities do not show a clear advantage in learning capability across all tasks, highlighting that EvaLearn evaluates a new dimension of model performance. We hope EvaLearn provides a novel evaluation perspective for assessing LLM potential and understanding the gap between models and human capabilities, promoting the development of deeper and more dynamic evaluation approaches. All datasets, the automatic evaluation framework, and the results studied in this paper are available in the supplementary materials.

We present the B-spline Encoded Action Sequence Tokenizer (BEAST), a novel action tokenizer that encodes action sequences into compact discrete or continuous tokens using B-splines. In contrast to existing action tokenizers based on vector quantization or byte pair encoding, BEAST requires no separate tokenizer training and consistently produces tokens of uniform length, enabling fast action sequence generation via parallel decoding. Leveraging our B-spline formulation, BEAST inherently ensures generating smooth trajectories without discontinuities between adjacent segments. We extensively evaluate BEAST by integrating it with three distinct model architectures: a Variational Autoencoder (VAE) with continuous tokens, a decoder-only Transformer with discrete tokens, and Florence-2, a pretrained Vision-Language Model with an encoder-decoder architecture, demonstrating BEAST's compatibility and scalability with large pretrained models. We evaluate BEAST across three established benchmarks consisting of 166 simulated tasks and on three distinct robot settings with a total of 8 real-world tasks. Experimental results demonstrate that BEAST (i) significantly reduces both training and inference computational costs, and (ii) consistently generates smooth, high-frequency control signals suitable for continuous control tasks while (iii) reliably achieves competitive task success rates compared to state-of-the-art methods.

The development of Reasoning Large Language Models (RLLMs) has significantly improved multi-step reasoning capabilities, but it has also made hallucination problems more frequent and harder to eliminate. While existing approaches address hallucination through external knowledge integration, model parameter analysis, or self-verification mechanisms, they fail to provide a comprehensive insight into how hallucinations **emerge** and **evolve** throughout the reasoning chain. In this work, we investigate hallucination causality under constrained knowledge domains by auditing the Chain-of-Thought (CoT) trajectory and assessing the model's cognitive confidence in potentially erroneous or biased claims. Analysis reveals that in long-CoT settings, RLLMs may iteratively reinforce biases and errors through flawed reflective processes, ultimately inducing hallucinated reasoning paths. Counterintuitively, even with interventions at hallucination origins, reasoning chains display pronounced ''chain disloyalty'', resisting correction and sustaining flawed trajectories. We further point out that existing hallucination detection methods are *less reliable and interpretable than previously assumed*, especially in complex multi-step reasoning contexts. Unlike Anthropic's circuit tracing that requires access to model parameters, our auditing **enables more interpretable long-chain hallucination attribution in black-box settings**, demonstrating stronger generalizability and practical utility. Our code is available at [this link](https://github.com/Winnie-Lian/AHa_Meta_Cognitive).

Physics-informed neural networks (PINNs) have emerged as a promising framework to develop complex scientific surrogate models, yet their scalability and accuracy often degrade in non-canonical geometries, such as non-rectangular domains or three-dimensional (3D) domains with high aspect ratios. These limitations hinder the broader adoption of vanilla PINNs in real-world, practical systems. In this work, we introduce a multi-domain PINN (MDPINN) framework designed to address the scalability and generalization challenges inherent in 3D non-rectangular domains governed by nonlinear fluid dynamics. The target domain consists of intersecting 3D fluid channels with a high aspect ratio, inducing complex flow features such as deflections, mixing, and recirculations. Our approach is grounded in two key innovations: 1) domain decomposition, which partitions the channel volumes into multiple cubic-like subdomains, each modeled by an individual PINN, 2) enforcement of global dynamics (MDPINN-GD), which ensures that the total mass flow rate entering the domain equals that exiting. These innovations reduce the complexity of the problem imposed on individual PINNs and guide effective network optimization toward physically consistent solutions throughout the domain. We demonstrate that our method achieves: 1) 74.8\% accuracy improvement over a single-network PINN, and 2) 52.9\% accuracy improvement over MDPINN that do not enforce global mass conservation. Furthermore, the MDPINN-GD framework exhibits accurate prediction even in highly complex regions-such as the channel intersecting zone and the outlet zone characterized by intense flow mixing and large velocity gradients-achieving maximum normalized mean absolute errors below 14.9\% for velocity predictions compared to simulation results. This work establishes a path towards scalable, physically grounded surrogate modeling approach that is extensible to multiphysics and high-dimensional scientific problems.

Numerous well-established studies have demonstrated the superhuman capabilities of modern Vision-Language Models (VLMs) across a wide range of tasks. However, growing is the doubt about the continuing availability of reliable high-quality labeling (supervision) from human annotators, leading to stagnation of the model's performance. To address this challenge, ``superalignment'' employs the so-called weak-to-strong generalization paradigm, where the supervision from a weak model can provide generalizable knowledge for a strong model. While effective in aligning knowledge for clean samples between the strong and weak models, the standard weak-to-strong approach typically fails to capture adversarial robustness, exposing strong VLMs to adversarial attacks. This inability to transfer adversarial robustness is because adversarial samples are normally missing in the superalignment stage. To this end, we are the first to propose the weak-to-strong (adversarial) robustness generalization method to elicit zero-shot robustness in large-scale models by an unsupervised scheme, mitigating the unreliable information source for alignment from two perspectives: alignment re-weighting and source guidance refinement. We analyze settings under which robustness generalization is possible. Extensive experiments across various vision-language benchmarks validate the effectiveness of our method in numerous scenarios, demonstrating its plug-and-play applicability to large-scale VLMs.

Recent advances in deep learning have driven rapid progress in time series forecasting, yet many state-of-the-art models continue to struggle with robust performance in real-world applications, even when they achieve strong results on standard benchmark datasets. This persistent gap can be attributed to the black-box nature of deep learning architectures and the inherent limitations of current evaluation frameworks, which frequently lack the capacity to provide clear, quantitative insights into the specific strengths and weaknesses of different models, thereby complicating the selection of appropriate models for particular forecasting scenarios. To address these issues, we propose a synthetic data-driven evaluation paradigm, SynTSBench, that systematically assesses fundamental modeling capabilities of time series forecasting models through programmable feature configuration. Our framework isolates confounding factors and establishes an interpretable evaluation system with three core analytical dimensions: (1) temporal feature decomposition and capability mapping, which enables systematic evaluation of model capacities to learn specific pattern types; (2) robustness analysis under data irregularities, which quantifies noise tolerance thresholds and anomaly recovery capabilities; and (3) theoretical optimum benchmarking, which establishes performance boundaries for each pattern type—enabling direct comparison between model predictions and mathematical optima.Our experiments show that current deep learning models do not universally approach optimal baselines across all types of temporal features.

The expanding computational costs and limited resources underscore the critical need for budgeted-iteration training, which aims to achieve optimal learning within predetermined iteration budgets. While learning rate schedules fundamentally govern the performance of different networks and tasks, particularly in budgeted-iteration scenarios, their design remains largely heuristic, lacking theoretical foundations. In addition, the optimal learning rate schedule requires extensive trial-and-error selection, making the training process inefficient. In this work, we propose the Unified Budget-Aware (UBA) schedule, a theoretically grounded learning rate schedule that consistently outperforms commonly-used schedules among diverse architectures and tasks under different constrained training budgets. First, we bridge the gap by constructing a novel training budget-aware optimization framework, which explicitly accounts for the robustness to landscape curvature variations. From this framework, we derive the UBA schedule, controlled by a single hyper-parameter $\varphi$ that provides a trade-off between flexibility and simplicity, eliminating the need for per-network numerical optimization. Moreover, we establish a theoretical connection between $\varphi$ and the condition number, adding interpretation and justification to our approach. Besides, we prove the convergence for different values of $\varphi$. We offer practical guidelines for $\varphi$ selection via theoretical analysis and empirical results. Extensive experimental results show that UBA $\textit{consistently surpasses}$ the commonly-used schedules across diverse vision and language tasks, spanning network architectures (e.g., ResNet, OLMo) and scales, under different training-iteration budgets.

Multimodal federated learning in real-world settings often encounters incomplete and heterogeneous data across clients. This results in misaligned local feature representations that limit the effectiveness of model aggregation. Unlike prior work that assumes either differing modality sets without missing input features or a shared modality set with missing features across clients, we consider a more general and realistic setting where each client observes a different subset of modalities and might also have missing input features within each modality. To address the resulting misalignment in learned representations, we propose a new federated learning framework featuring locally adaptive representations based on learnable client-side embedding controls that encode each client’s data-missing patterns. These embeddings serve as reconfiguration signals that align the globally aggregated representation with each client's local context, enabling more effective use of shared information. Furthermore, the embedding controls can be algorithmically aggregated across clients with similar data-missing patterns to enhance the robustness of reconfiguration signals in adapting the global representation. Empirical results on multiple federated multimodal benchmarks with diverse data-missing patterns across clients demonstrate the efficacy of the proposed method, achieving up to 36.45\% performance improvement under severe data incompleteness. The method is also supported by a theoretical analysis with an explicit performance bound that matches our empirical observations. Our source codes are provided at [https://github.com/nmduonggg/PEPSY](https://github.com/nmduonggg/PEPSY)

Sharpness‐aware minimization (SAM) has emerged as a highly effective technique for improving model generalization, but its underlying principles are not fully understood. We investigated the phenomenon known as m-sharpness, where the performance of SAM improves monotonically as the micro-batch size for computing perturbations decreases. In practice, the empirical m-sharpness effect underpins the deployment of SAM in distributed training, yet a rigorous theoretical account has remained lacking. To provide a theoretical explanation for m-sharpness, we leverage an extended Stochastic Differential Equation (SDE) framework and analyze the structure of stochastic gradient noise (SGN) to characterize the dynamics of various SAM variants, including n-SAM and m-SAM. Our findings reveal that the stochastic noise introduced during SAM perturbations inherently induces a variance-based sharpness regularization effect. Motivated by our theoretical insights, we introduce Reweighted SAM (RW-SAM), which employs sharpness-weighted sampling to mimic the generalization benefits of m-SAM while remaining parallelizable. Comprehensive experiments validate the effectiveness of our theoretical analysis and proposed method.

Computational methods for learning to sample from the Boltzmann distribution—where the target distribution is known only up to an unnormalized energy function—have advanced significantly recently. Due to the lack of explicit target samples, however, prior diffusion-based methods, known as _diffusion samplers_, often require importance-weighted estimation or complicated learning processes. Both trade off scalability with extensive evaluations of the energy and model, thereby limiting their practical usage. In this work, we propose **Adjoint Schrödinger Bridge Sampler (ASBS)**, a new diffusion sampler that employs simple and scalable matching-based objectives yet without the need to estimate target samples during training. ASBS is grounded on a mathematical model—the Schrödinger Bridge—which enhances sampling efficiency via kinetic-optimal transportation. Through a new lens of stochastic optimal control theory, we demonstrate how SB-based diffusion samplers can be learned at scale via Adjoint Matching and prove convergence to the global solution. Notably, ASBS generalizes the recent Adjoint Sampling (Havens et al., 2025) to arbitrary source distributions by relaxing the so-called memoryless condition that largely restricts the design space. Through extensive experiments, we demonstrate the effectiveness of ASBS on sampling from classical energy functions, amortized conformer generation, and molecular Boltzmann distributions. Codes are available at https://github.com/facebookresearch/adjoint_samplers

This work introduces a hybrid non-Euclidean optimization method which generalizes gradient norm clipping by combining steepest descent and conditional gradient approaches. The method achieves the best of both worlds by establishing a descent property under a generalized notion of ($L_0$,$L_1$)-smoothness. Weight decay is incorporated in a principled manner by identifying a connection to the Frank-Wolfe short step. In the stochastic case, we show an order optimal $O(n^{-1/4})$ convergence rate by leveraging a momentum based gradient estimator. We discuss how to instantiate the algorithms for deep learning, which we dub Clipped Scion, and demonstrate their properties on image classification and language modeling. The code is available at https://github.com/LIONS-EPFL/ClippedScion.

Novel view synthesis is an important problem with many applications, including AR/VR, gaming, and robotic simulations. With the recent rapid development of Neural Radiance Fields (NeRFs) and 3D Gaussian Splatting (3DGS) methods, it is becoming difficult to keep track of the current state of the art (SoTA) due to methods using different evaluation protocols, codebases being difficult to install and use, and methods not generalizing well to novel 3D scenes. In our experiments, we show that even tiny differences in the evaluation protocols of various methods can artificially boost the performance of these methods. This raises questions about the validity of quantitative comparisons performed in the literature. To address these questions, we propose NerfBaselines, an evaluation framework which provides consistent benchmarking tools, ensures reproducibility, and simplifies the installation and use of various methods. We validate our implementation experimentally by reproducing the numbers reported in the original papers. For improved accessibility, we release a web platform that compares commonly used methods on standard benchmarks. We strongly believe NerfBaselines is a valuable contribution to the community as it ensures that quantitative results are comparable and thus truly measure progress in the field of novel view synthesis.

Despite remarkable progress in driving world models, their potential for autonomous systems remains largely untapped: the world models are mostly learned for world simulation and decoupled from trajectory planning. While recent efforts aim to unify world modeling and planning in a single framework, the synergistic facilitation mechanism of world modeling for planning still requires further exploration. In this work, we introduce a new driving paradigm named Policy World Model (PWM), which not only integrates world modeling and trajectory planning within a unified architecture, but is also able to benefit planning using the learned world knowledge through the proposed action-free future state forecasting scheme. Through collaborative state-action prediction, PWM can mimic the human-like anticipatory perception, yielding more reliable planning performance. To facilitate the efficiency of video forecasting, we further introduce a parallel token generation mechanism, equipped with a context-guided tokenizer and an adaptive dynamic focal loss. Despite utilizing only front camera input, our method matches or exceeds state-of-the-art approaches that rely on multi-view and multi-modal inputs. Code will be released at https://github.com/6550Zhao/Policy-World-Model.

Predictions in the form of sets of probability distributions, so-called credal sets, provide a suitable means to represent a learner's epistemic uncertainty. In this paper, we propose a theoretically grounded approach to credal prediction based on the statistical notion of relative likelihood: The target of prediction is the set of all (conditional) probability distributions produced by the collection of plausible models, namely those models whose relative likelihood exceeds a specified threshold. This threshold has an intuitive interpretation and allows for controlling the trade-off between correctness and precision of credal predictions. We tackle the problem of approximating credal sets defined in this way by means of suitably modified ensemble learning techniques. To validate our approach, we illustrate its effectiveness by experiments on benchmark datasets demonstrating superior uncertainty representation without compromising predictive performance. We also compare our method against several state-of-the-art baselines in credal prediction.

Diffusion models (DMs) excel in image generation, but suffer from slow inference and the training-inference discrepancies. Although gradient-based solvers like DPM-Solver accelerate the denoising inference, they lack theoretical foundations in information transmission efficiency. In this work, we introduce an information-theoretic perspective on the inference processes of DMs, revealing that successful denoising fundamentally reduces conditional entropy in reverse transitions. This principle leads to our key insights into the inference processes: (1) data prediction parameterization outperforms its noise counterpart, and (2) optimizing conditional variance offers *a reference-free way* to minimize both transition and reconstruction errors. Based on these insights, we propose an entropy-aware variance optimized method for the generative process of DMs, called *EVODiff*, which systematically reduces uncertainty by optimizing conditional entropy during denoising. Extensive experiments on DMs validate our insights and demonstrate that our method significantly and consistently outperforms state-of-the-art (SOTA) gradient-based solvers. For example, compared to the DPM-Solver++, EVODiff reduces the reconstruction error by up to *45.5\%* (FID improves from 5.10 to 2.78) at 10 function evaluations (NFE) on CIFAR-10, cuts the NFE cost by *25\%* (from 20 to 15 NFE) for high-quality samples on ImageNet-256, and improves text-to-image generation while reducing artifacts. Code is available at https://github.com/ShiguiLi/EVODiff.

This paper presents **Pixel-Perfect Depth**, a monocular depth estimation model based on pixel-space diffusion generation that produces high-quality, flying-pixel-free point clouds from estimated depth maps. Current generative depth estimation models fine-tune Stable Diffusion and achieve impressive performance. However, they require a VAE to compress depth maps into the latent space, which inevitably introduces flying pixels at edges and details. Our model addresses this challenge by directly performing diffusion generation in the pixel space, avoiding VAE-induced artifacts. To overcome the high complexity associated with pixel-space generation, we introduce two novel designs: 1) **Semantics-Prompted Diffusion Transformers** (**SP-DiT**), which incorporate semantic representations from vision foundation models into DiT to prompt the diffusion process, thereby preserving global semantic consistency while enhancing fine-grained visual details; and 2) **Cascade DiT Design** that progressively increases the number of tokens to further enhance efficiency and accuracy. Our model achieves the best performance among all published generative models across five benchmarks, and significantly outperforms all other models in edge-aware point cloud evaluation. Project page: https://pixel-perfect-depth.github.io/.