NeurIPS 2025 Papers

Large language models (LLMs) have shown promise in transforming machine learning research, yet their capability to faithfully implement genuinely novel ideas from recent research papers—ideas unseen during pretraining—remains unclear. We introduce ResearchCodeBench, a benchmark that evaluates LLMs’ ability to translate cutting-edge ML contributions from top 2024-2025 research papers into executable code. We assessed 30+ proprietary and open-source LLMs, finding that even the best models correctly implement less than 40% of the code. We present empirical findings on performance comparison, contamination, and error patterns. By providing a rigorous evaluation platform, ResearchCodeBench enables continuous understanding and advancement of LLM-driven innovation in research code generation.

Spectral features are widely incorporated within Graph Neural Networks (GNNs) to improve their expressive power, or their ability to distinguish among non-isomorphic graphs. One popular example is the usage of graph Laplacian eigenvectors for positional encoding in MPNNs and Graph Transformers. The expressive power of such Spectrally-enhanced GNNs (SGNNs) is usually evaluated via the $k$-WL graph isomorphism test hierarchy and homomorphism counting. Yet, these frameworks align poorly with the graph spectra, yielding limited insight into SGNNs' expressive power. In this paper, we leverage a well-studied paradigm of classifying graphs by their largest eigenvalue multiplicity to introduce an expressivity hierarchy for SGNNs. We then prove that many SGNNs are incomplete even on graphs with distinct eigenvalues. To mitigate this deficiency, we adapt rotation equivariant neural networks to the graph spectra setting, yielding equiEPNN, a novel SGNN that provably improves upon contemporary SGNNs' expressivity on simple spectrum graphs. We then demonstrate that equiEPNN achieves perfect eigenvector canonicalization on ZINC, and performs favorably on image classification on MNIST-Superpixel and graph property regression on ZINC, compared to leading spectral methods.

Current transformer-based imitation learning approaches introduce discrete action representations and train an autoregressive transformer decoder on the resulting latent code. However, the initial quantization breaks the continuous structure of the action space thereby limiting the capabilities of the generative model. We propose a quantization-free method instead that leverages Generative Infinite-Vocabulary Transformers (GIVT) as a direct, continuous policy parametrization for autoregressive transformers. This simplifies the imitation learning pipeline while achieving state-of-the-art performance on a variety of popular simulated robotics tasks. We enhance our policy roll-outs by carefully studying sampling algorithms, further improving the results.

We present STARFlow, a scalable generative model based on normalizing flows that achieves strong performance on high-resolution image synthesis. STARFlow's main building block is Transformer Autoregressive Flow (TARFlow), which combines normalizing flows with Autoregressive Transformer architectures and has recently achieved impressive results in image modeling. In this work, we first establish the theoretical universality of TARFlow for modeling continuous distributions. Building on this foundation, we introduce a set of architectural and algorithmic innovations that significantly enhance the scalability: (1) a deep-shallow design where a deep Transformer block captures most of the model’s capacity, followed by a few shallow Transformer blocks that are computationally cheap yet contribute non-negligibly, (2) learning in the latent space of pretrained autoencoders, which proves far more effective than modeling pixels directly, and (3) a novel guidance algorithm that substantially improves sample quality. Crucially, our model remains a single, end-to-end normalizing flow, allowing exact maximum likelihood training in continuous space without discretization. STARFlow achieves competitive results in both class- and text-conditional image generation, with sample quality approaching that of state-of-the-art diffusion models. To our knowledge, this is the **first** successful demonstration of normalizing flows at this scale and resolution. Code and weights available at https://github.com/apple/ml-starflow.

Neural networks are powered by an implicit bias: a tendency of gradient descent to fit training data in a way that generalizes to unseen data. A recent class of neural network models gaining increasing popularity is structured state space models (SSMs). Prior work argued that the implicit bias of SSMs leads to generalization in a setting where data is generated by a low dimensional teacher. In this paper, we revisit the latter setting, and formally establish a phenomenon entirely undetected by prior work on the implicit bias of SSMs. Namely, we prove that while implicit bias leads to generalization under many choices of training data, there exist special examples whose inclusion in training completely distorts the implicit bias, to a point where generalization fails. This failure occurs despite the special training examples being labeled by the teacher, i.e., having clean labels! We empirically demonstrate the phenomenon, with SSMs trained independently and as part of non-linear neural networks. In the area of adversarial machine learning, disrupting generalization with cleanly labeled training examples is known as clean-label poisoning. Given the proliferation of SSMs, we believe that delineating their susceptibility to clean-label poisoning, and developing methods for overcoming this susceptibility, are critical research directions to pursue.

Modeling boiling---an inherently chaotic, multiphase process central to energy and thermal systems---remains a significant challenge for neural PDE surrogates. Existing models require future input (e.g., bubble positions) during inference because they fail to learn nucleation from past states, limiting their ability to autonomously forecast boiling dynamics. They also fail to model flow boiling velocity fields, where sharp interface–momentum coupling demands long-range and directional inductive biases. We introduce Bubbleformer, a transformer-based spatiotemporal model that forecasts stable and long-range boiling dynamics including nucleation, interface evolution, and heat transfer without dependence on simulation data during inference. Bubbleformer integrates factorized axial attention, frequency-aware scaling, and conditions on thermophysical parameters to generalize across fluids, geometries, and operating conditions.To evaluate physical fidelity in chaotic systems, we propose interpretable physics-based metrics that evaluate heat flux consistency, interface geometry, and mass conservation. We also release BubbleML 2.0, a high-fidelity dataset that spans diverse working fluids (cryogens, refrigerants, dielectrics), boiling configurations (pool and flow boiling), flow regimes (bubbly, slug, annular), and boundary conditions. Bubbleformer sets new benchmark results in both prediction and forecasting of two-phase boiling flows.

The open-source release of large language models (LLMs) enables malicious users to create unauthorized derivative models at low cost, posing significant threats to intellectual property (IP) and market stability. Existing IP protection methods either require access to model parameters or are vulnerable to fine-tuning attacks. To fill this gap, we propose ErrorTrace, a robust and black-box traceability mechanism for protecting LLM IP. Specifically, ErrorTrace leverages the unique error patterns of model families by mapping and analyzing their distinct error spaces, enabling robust and efficient IP protection without relying on internal parameters or specific query responses. Experimental results show that ErrorTrace achieves a traceability accuracy of 0.8518 for 27 base models when the suspect model is not included in ErrorTrace's training set, outperforming the baseline by 0.2593. Additionally,ErrorTrace successfully tracks 34 fine-tuned, pruned and merged models across various scenarios, demonstrating its broad applicability and robustness. In addition, ErrorTrace shows a certain level of resilience when subjected to adversarial attacks. Our code is available at: https://github.com/csdatazcc/ErrorTrace.

Multimodal representation learning, exemplified by multimodal contrastive learning (MMCL) using image-text pairs, aims to learn powerful representations by aligning cues across modalities. This approach relies on the core assumption that the exemplar image-text pairs constitute two representations of an identical concept. However, recent research has revealed that real-world datasets often exhibit cross-modal misalignment. There are two distinct viewpoints on how to address this issue: one suggests mitigating the misalignment, and the other leveraging it. We seek here to reconcile these seemingly opposing perspectives, and to provide a practical guide for practitioners. Using latent variable models we thus formalize cross-modal misalignment by introducing two specific mechanisms: Selection bias, where some semantic variables are absent in the text, and perturbation bias, where semantic variables are altered—both leading to misalignment in data pairs. Our theoretical analysis demonstrates that, under mild assumptions, the representations learned by MMCL capture exactly the information related to the subset of the semantic variables invariant to selection and perturbation biases. This provides a unified perspective for understanding misalignment. Based on this, we further offer actionable insights into how misalignment should inform the design of real-world ML systems. We validate our theoretical findings via extensive empirical studies on both synthetic data and real image-text datasets, shedding light on the nuanced impact of cross-modal misalignment on multimodal representation learning.

This paper presents a novel approach to addressing the long-sequence problem in high-resolution medical images for Vision Transformers (ViTs). Using smaller patches as tokens can enhance ViT performance, but quadratically increases computation and memory requirements. Therefore, the common practice for applying ViTs to high-resolution images is either to: (a) employ complex sub-quadratic attention schemes or (b) use large to medium-sized patches and rely on additional mechanisms within the model to capture the spatial hierarchy of details. We propose Symmetrical Hierarchical Forest (SHF), a lightweight approach that adaptively patches the input image to increase token information density and encode hierarchical spatial structures into the input embedding. We then apply a reverse depatching scheme to the output embeddings of the transformer encoder, eliminating the need for convolution-based decoders. Unlike previous methods that modify attention mechanisms \wahib{or use a complex hierarchy of interacting models}, SHF can be retrofitted to any ViT model to allow it to learn the hierarchical structure of details in high-resolution images without requiring architectural changes. Experimental results demonstrate significant gains in computational efficiency and performance: on the PAIP WSI dataset, we achieved a 3$\sim$32$\times$ speedup or a 2.95\% to 7.03\% increase in accuracy (measured by Dice score) at a $64K^2$ resolution with the same computational budget, compared to state-of-the-art production models. On the 3D medical datasets BTCV and KiTS, training was 6$\times$ faster, with accuracy gains of 6.93\% and 5.9\%, respectively, compared to models without SHF.

We introduce HBLLM, a wavelet-enhanced high-fidelity $1$-bit post-training quantization method for Large Language Models (LLMs). By leveraging Haar wavelet transforms to enhance expressive capacity through frequency decomposition, HBLLM significantly improves quantization fidelity while maintaining minimal overhead. This approach features two innovative structure-aware grouping strategies: (1) frequency-aware multi-parameter intra-row grouping and (2) $\ell_2$-norm-based saliency-driven column selection. For non-salient weights, a shared mean is employed across quantization groups within each frequency band to optimize storage efficiency. Experiments conducted on the OPT and LLaMA models demonstrate that HBLLM achieves state-of-the-art performance in $1$-bit quantization, attaining a perplexity of $6.71$ on LLaMA$2$-$13$B with an average weight storage of only $1.08$ bits. Code available at: https://github.com/Yeyke/HBLLM.

Solving time-dependent parametric partial differential equations (PDEs) remains a fundamental challenge for neural solvers, particularly when generalizing across a wide range of physical parameters and dynamics. When data is uncertain or incomplete—as is often the case—a natural approach is to turn to generative models. We introduce ENMA, a generative neural operator designed to model spatio-temporal dynamics arising from physical phenomena. ENMA predicts future dynamics in a compressed latent space using a generative masked autoregressive transformer trained with flow matching loss, enabling tokenwise generation. Irregularly sampled spatial observations are encoded into uniform latent representations via attention mechanisms and further compressed through a spatio-temporal convolutional encoder. This allows ENMA to perform in-context learning at inference time by conditioning on either past states of the target trajectory or auxiliary context trajectories with similar dynamics. The result is a robust and adaptable framework that generalizes to new PDE regimes and supports one-shot surrogate modeling of time-dependent parametric PDEs.

Recent advances in large language models (LLMs) have enabled automated dataset labeling with minimal human supervision. While majority voting across multiple LLMs can improve label reliability by mitigating individual model biases, it incurs high computational costs due to repeated querying. In this work, we propose a novel online framework, Cost-aware Majority Voting (CaMVo), for efficient and accurate LLM-based dataset annotation. CaMVo adaptively selects a subset of LLMs for each data instance based on contextual embeddings, balancing confidence and cost without requiring pre-training or ground-truth labels. Leveraging a LinUCB-based selection mechanism and a Bayesian estimator over confidence scores, CaMVo estimates a lower bound on labeling accuracy for each LLM and aggregates responses through weighted majority voting. Our empirical evaluation on the MMLU and IMDB Movie Review datasets demonstrates that CaMVo achieves comparable or superior accuracy to full majority voting while significantly reducing labeling costs. This establishes CaMVo as a practical and robust solution for cost-efficient annotation in dynamic labeling environments.

We consider the problem of estimating the correlation of two random variables $X$ and $Y$, where the pairs $(X,Y)$ are not observed together, but are instead separated co-ordinate-wise at two servers: server 1 contains all the $X$ observations, and server 2 contains the corresponding $Y$ observations. In this vertically distributed setting, we assume that each server has its own privacy constraints, owing to which they can only share suitably privatized statistics of their own component observations. We consider differing privacy budgets $(\varepsilon_1,\delta_1)$ and $(\varepsilon_2,\delta_2)$ for the two servers and determine the minimax optimal rates for correlation estimation allowing for both non-interactive and interactive mechanisms. We also provide correlation estimators that achieve these rates and further develop inference procedures, namely, confidence intervals, for the estimated correlations. Our results are characterized by an interesting rate in terms of the sample size $n$, $\varepsilon_1$, $\varepsilon_2$, which is strictly slower than the usual central privacy estimation rates. More interestingly, we find that the interactive mechanism is always better than its non-interactive counterpart whenever the two privacy budgets are different. Results from extensive numerical experiments support our theoretical findings.

Most image fusion methods are designed for ideal scenarios and struggle to handle noise. Existing noise-aware fusion methods are supervised and heavily rely on constructed paired data, limiting performance and generalization. This paper proposes a novel unsupervised noisy visible and infrared image fusion method, comprising two key modules. First, when only noisy source images are available, a convolutional low-rank optimization module decomposes clean components based on convolutional low-rank priors, guiding subsequent optimization. The unsupervised approach eliminates data dependency and enhances generalization across various and variable noise. Second, a unified network jointly realizes denoising and fusion. It consists of both intra-modal recovery and inter-modal recovery and fusion, also with a convolutional low-rankness loss for regularization. By exploiting the commonalities of denoising and fusion, the joint framework significantly reduces network complexity while expanding functionality. Extensive experiments validate the effectiveness and generalization of the proposed method for image fusion under various and variable noise conditions. The code is publicly available at https://github.com/hanna-xu/Deno-IF.

The paper proposes a novel machine learning-based approach to the pathfinding problem on extremely large graphs. This method leverages diffusion distance estimation via a neural network and uses beam search for pathfinding. We demonstrate its efficiency by finding solutions for 4x4x4 and 5x5x5 Rubik's cubes with unprecedentedly short solution lengths, outperforming all available solvers and introducing the first machine learning solver beyond the 3x3x3 case. In particular, it surpasses every single case of the combined best results in the Kaggle Santa 2023 challenge, which involved over 1,000 teams. For the 3x3x3 Rubik's cube, our approach achieves an optimality rate exceeding 98%, matching the performance of task-specific solvers and significantly outperforming prior solutions such as DeepCubeA (60.3%) and EfficientCube (69.6%). Our solution in its current implementation is approximately 25.6 times faster in solving 3x3x3 Rubik's cubes while requiring up to 8.5 times less model training time than the most efficient state-of-the-art competitor. Finally, it is demonstrated that even a single agent trained using a relatively small number of examples can robustly solve a broad range of puzzles represented by Cayley graphs of size up to $10^{145}$, confirming the generality of the proposed method.

The scale and quality of datasets are crucial for training robust perception models. However, obtaining large-scale annotated data is both costly and time-consuming. Generative models have emerged as a powerful tool for data augmentation by synthesizing samples that adhere to desired distributions. However, current generative approaches often rely on complex post-processing or extensive fine-tuning on massive datasets to achieve satisfactory results, and they remain prone to content–position mismatches and semantic leakage. To overcome these limitations, we introduce ReCon, a novel augmentation framework that enhances the capacity of structure-controllable generative models for object detection. ReCon integrates region-guided rectification into the diffusion sampling process, using feedback from a pre-trained perception model to rectify misgenerated regions within diffusion sampling process. We further propose region-aligned cross-attention to enforce spatial–semantic alignment between image regions and their textual cues, thereby improving both semantic consistency and overall image fidelity. Extensive experiments demonstrate that ReCon substantially improve the quality and trainability of generated data, achieving consistent performance gains across various datasets, backbone architectures, and data scales.

Designing protein sequences that fold into a target 3D structure—known as protein inverse folding—is a fundamental challenge in protein engineering. While recent deep learning methods have achieved impressive performance by recovering native sequences, they often overlook the one-to-many nature of the problem: multiple diverse sequences can fold into the same structure. This motivates the need for a generative model capable of designing diverse sequences while preserving structural consistency. To address this trade-off, we introduce ProtInvTree, the first reward-guided tree-search framework for protein inverse folding. ProtInvTree reformulates sequence generation as a deliberate, step-wise decision-making process, enabling the exploration of multiple design paths and exploitation of promising candidates through self-evaluation, lookahead, and backtracking. We propose a two-stage focus-and-grounding action mechanism that decouples position selection and residue generation. To efficiently evaluate intermediate states, we introduce a jumpy denoising strategy that avoids full rollouts. Built upon pretrained protein language models, ProtInvTree supports flexible test-time scaling by adjusting the search depth and breadth without retraining. Empirically, ProtInvTree outperforms state-of-the-art baselines across multiple benchmarks, generating structurally consistent yet diverse sequences, including those far from the native ground truth. The code is available at https://github.com/A4Bio/ProteinInvBench/.

Autoregressive models have transformed protein engineering by enabling the generation of novel protein sequences beyond those found in nature. However, their sequential inference introduces significant latency, limiting their utility in high-throughput protein screening. Speculative decoding accelerates generation by employing a lightweight draft model to sample tokens, which a larger target model then verifies and refines. Yet in protein sequence generation, draft models are typically agnostic to the structural and functional constraints of the target protein, leading to biologically implausible outputs and a shift in the likelihood distribution of generated sequences. We introduce SpecMER (Speculative Decoding via k-mer Guidance), a novel framework that incorporates biological, structural, and functional priors using k-mer motifs extracted from multiple sequence alignments. By scoring candidate sequences in parallel and selecting those most consistent with known biological patterns, SpecMER significantly improves sequence plausibility while retaining the efficiency of speculative decoding. SpecMER achieves 24–32% speedup over standard autoregressive decoding, along with higher acceptance rates and improved sequence likelihoods.

Recent long-form video-language understanding benchmarks have driven progress in video large multimodal models (Video-LMMs). However, the scarcity of well-annotated long videos has left the training of hour-long Video-LMMs underexplored. To close this gap, we present VideoMarathon, a large-scale hour-long video instruction-following dataset. This dataset includes around 9,700 hours of long videos sourced from diverse domains, ranging from 3 to 60 minutes per video. Specifically, it contains 3.3M high-quality QA pairs, spanning six fundamental topics: temporality, spatiality, object, action, scene, and event. Compared to existing video instruction datasets, VideoMarathon significantly extends training video durations up to 1 hour, and supports 22 diverse tasks requiring both short- and long-term video comprehension. Building on VideoMarathon, we propose Hour-LLaVA, a powerful and efficient Video-LMM for hour-scale video-language modeling. It enables hour-long video training and inference at 1-FPS sampling by leveraging a memory augmentation module, which adaptively integrates question-relevant and spatiotemporally informative semantics from the cached full video context. In our experiments, Hour-LLaVA achieves the best performance on multiple representative long video-language benchmarks, demonstrating the high quality of the VideoMarathon dataset and the superiority of the Hour-LLaVA model.