NeurIPS 2025 论文列表

Synthetically-generated data plays an increasingly larger role in training large language models. However, while synthetic data has been found to be useful, studies have also shown that without proper curation it can cause LLM performance to plateau, or even "collapse", after many training iterations. In this paper, we formalize this question and develop a theoretical framework to investigate how much curation is needed in order to ensure that LLM performance continually improves. Our analysis is inspired by boosting, a classic machine learning technique that leverages a very weak learning algorithm to produce an arbitrarily good classifier. The approach we analyze subsumes many recently proposed methods for training LLMs on synthetic data, and thus our analysis sheds light on why they are successful, and also suggests opportunities for future improvement. We present experiments that validate our theory, and show that dynamically focusing labeling resources on the most challenging examples --- in much the same way that boosting focuses the efforts of the weak learner --- leads to improved performance.

Transformers have revolutionized nucleotide sequence analysis, yet capturing long‑range dependencies remains challenging. Recent studies show that autoregressive transformers often exhibit Markovian behavior by relying on fixed-length context windows for next-token prediction. However, standard self-attention mechanisms are computationally inefficient for long sequences due to their quadratic complexity and do not explicitly enforce global transition consistency. We introduce CARMANIA (Context-Aware Regularization with Markovian Integration for Attention-Based Nucleotide Analysis), a self-supervised pretraining framework that augments next-token (NT) prediction with a transition-matrix (TM) loss. The TM loss aligns predicted token transitions with empirically derived n-gram statistics from each input sequence, encouraging the model to capture higher-order dependencies beyond local context. This integration enables CARMANIA to learn organism-specific sequence structures that reflect both evolutionary constraints and functional organization. We evaluate CARMANIA across diverse genomic tasks, including regulatory element prediction, functional gene classification, taxonomic inference, antimicrobial resistance detection, and biosynthetic gene cluster classification. CARMANIA outperforms the previous best long-context model by at least 7\%, matches state-of-the-art on shorter sequences (exceeding prior results on 20/40 tasks while running $\sim$2.5$\times$ faster), and shows particularly strong improvements on enhancer and housekeeping gene classification tasks—including up to a 34\% absolute gain in Matthews correlation coefficient (MCC) for enhancer prediction. The TM loss boosts accuracy in 33 of 40 tasks, especially where local motifs or regulatory patterns drive prediction. This enables more effective modeling of sequence-dependent biological features while maintaining robustness across non-coding and low-signal regions. Code available at https://github.com/EESI/carmania.

Large-scale AI models, such as Large Language Models (LLMs) and Diffusion Models (DMs), have grown rapidly in size, creating significant challenges for efficient deployment on resource-constrained hardware. In this paper, we introduce Dynamic-Length Float (DFloat11), a lossless compression framework that reduces LLM and DM size by 30\% while preserving outputs that are bit-for-bit identical to the original model. DFloat11 is motivated by the low entropy in the BFloat16 weight representation of LLMs, which reveals significant inefficiency in the existing storage format. By applying entropy coding, DFloat11 assigns dynamic-length encodings to weights based on frequency, achieving near information-optimal compression without any loss of precision. To facilitate efficient inference with dynamic-length encodings, we develop a custom GPU kernel for fast online decompression. Our design incorporates the following: (i) compact, hierarchical lookup tables (LUTs) that fit within GPU SRAM for efficient decoding, (ii) a two-phase GPU kernel for coordinating thread read/write positions using lightweight auxiliary variables, and (iii) transformer-block-level decompression to minimize latency. Experiments on Llama 3.3, Qwen 3, Mistral 3, FLUX.1, and others validate our hypothesis that DFloat11 achieves around 30\% model size reduction while preserving bit-for-bit identical outputs. Compared to a potential alternative of offloading parts of an uncompressed model to the CPU to meet memory constraints, DFloat11 achieves 2.3--46.2$\times$ higher throughput in token generation. With a fixed GPU memory budget, DFloat11 enables 5.7--14.9$\times$ longer generation lengths than uncompressed models. Notably, our method enables lossless inference of Llama 3.1 405B, an 810GB model, on a single node equipped with 8$\times$80GB GPUs.

The modeling of complicated time-evolving physical dynamics from partial observations is a long-standing challenge. Particularly, observations can be sparsely distributed in a seemingly random or unstructured manner, making it difficult to capture highly nonlinear features in a variety of scientific and engineering problems. However, existing data-driven approaches are often constrained by fixed spatial and temporal discretization. While some researchers attempt to achieve spatio-temporal continuity by designing novel strategies, they either overly rely on traditional numerical methods or fail to truly overcome the limitations imposed by discretization. To address these, we propose CoPS, a purely data-driven methods, to effectively model continuous physics simulation from partial observations. Specifically, we employ multiplicative filter network to fuse and encode spatial information with the corresponding observations. Then we customize geometric grids and use message-passing mechanism to map features from original spatial domain to the customized grids. Subsequently, CoPS models continuous-time dynamics by designing multi-scale graph ODEs, while introducing a Markov-based neural auto-correction module to assist and constrain the continuous extrapolations. Comprehensive experiments demonstrate that CoPS advances the state-of-the-art methods in space-time continuous modeling across various scenarios. The source code is available at~\url{https://github.com/Sunxkissed/CoPS}.

Relational databases underpin critical infrastructure across a wide range of domains, yet the design of generalizable pre-training strategies for learning from relational databases remains an open challenge due to task heterogeneity. Specifically, there exist many possible downstream tasks, as tasks are defined based on relational schema graphs, temporal dependencies, and SQL-defined label logics. An effective pre-training framework is desired to take these factors into account in order to obtain task-aware representations. By incorporating knowledge of the underlying distribution that drives label generation, downstream tasks can benefit from relevant side-channel information. To bridge this gap, we introduce Task Vector Estimation (TVE), a novel pre-training framework that constructs predictive supervisory signals via set-based aggregation over schema traversal graphs, explicitly modeling next-window relational dynamics. We formalize our approach through an information-theoretic lens, demonstrating that task-informed representations retain more relevant signals than those obtained without task priors. Extensive experiments on the RelBench benchmark show that TVE consistently outperforms traditional pre-training baselines. Our findings advocate for pre-training objectives that encode task heterogeneity and temporal structure as design principles for predictive modeling on relational databases. Our code is publicly available at https://github.com/quang-truong/task-vector-estimation.

Multivariate time series (MTS) classification has attracted increasing attention across various domains. Existing methods either decompose MTS into separate univariate series, ignoring inter-variable dependencies, or jointly model all variables, which may lead to over-smoothing and loss of semantic structure. These limitations become particularly pronounced when dealing with complex and heterogeneous variable types. To address these challenges, we propose SwinGroupNet (SGN), which explores a novel perspective for constructing variable interaction and temporal dependency. Specifically, SGN processes multi-scale time series using (1) Variable Group Embedding (VGE), which partitions variables into groups and performs independent group-wise embedding; (2) Multi-Scale Group Window Mixing (MGWM), which reconstructs variable interactions by modeling both intra-group and inter-group dependencies while extracting multi-scale temporal features; and (3) Periodic Window Shifting and Merging (PWSM), which exploits inherent periodic patterns to enable hierarchical temporal interaction and feature aggregation. Extensive experiments on diverse benchmark datasets from multiple domains demonstrate that SGN consistently achieves state-of-the-art performance, with an average improvement of 4.2% over existing methods. We release the source code at https://anonymous.4open.science/r/SGN.

Few-shot node classification on hypergraphs requires models that generalize from scarce labels while capturing high-order structures. Existing hypergraph neural networks (HNNs) effectively encode such structures but often suffer from overfitting and scalability issues due to complex, black-box architectures. In this work, we propose ZEN (Zero-Parameter Hypergraph Neural Network), a fully linear and parameter-free model that achieves both expressiveness and efficiency. Built upon a unified formulation of linearized HNNs, ZEN introduces a tractable closed-form solution for the weight matrix and a redundancy-aware propagation scheme to avoid iterative training and to eliminate redundant self-information. On 11 real-world hypergraph benchmarks, ZEN consistently outperforms eight baseline models in classification accuracy while achieving up to 696x speedups over the fastest competitor. Moreover, the decision process of ZEN is fully interpretable, providing insights into the characteristic of a dataset. Our code and datasets are fully available at https://github.com/chaewoonbae/ZEN.

Activation steering is an efficient technique for aligning the behavior of large language models (LLMs) by injecting steering vectors directly into a model’s residual stream during inference. A pivotal challenge in this approach lies in choosing the right layers to intervene, as inappropriate selection can undermine behavioral alignment and even impair the model’s language fluency and other core capabilities. While single-layer steering allows straightforward evaluation on held-out data to identify the "best" layer, it offers only limited alignment improvements. Multi-layer steering promises stronger control but faces a combinatorial explosion of possible layer subsets, making exhaustive search impractical. To address these challenges, we propose LayerNavigator, which provides a principled and promising layer selection strategy. The core innovation of LayerNavigator lies in its novel, quantifiable criterion that evaluates each layer's steerability by jointly considering two key aspects: discriminability and consistency. By reusing the activations computed during steering vector generation, LayerNavigator requires no extra data and adds negligible overhead. Comprehensive experiments show that LayerNavigator achieves not only superior alignment but also greater scalability and interpretability compared to existing strategies. Our code is available at https://github.com/Bryson-Arrot/LayerNavigator

Calibrating blackbox machine learning models to achieve risk control is crucial to ensure reliable decision-making. A rich line of literature has been studying how to calibrate a model so that its predictions satisfy explicit finite-sample statistical guarantees under a _fixed_, _static_, and unknown data-generating distribution. However, prediction-supported decisions may influence the outcome they aim to predict, a phenomenon named _performativity_ of predictions, which is commonly seen in social science and economics. In this paper, we introduce _Performative Risk Control_, a framework to calibrate models to achieve risk control under performativity with provable theoretical guarantees. Specifically, we provide an iteratively refined calibration process, where we ensure the predictions are improved and risk-controlled throughout the process. We also study different types of risk measures and choices of tail bounds. Lastly, we demonstrate the effectiveness of our framework by numerical experiments on the task of predicting credit default risk. To the best of our knowledge, this work is the first one to study statistically rigorous risk control under performativity, which will serve as an important safeguard against a wide range of strategic manipulation in decision-making processes.

Mean-field variational inference (VI), despite its scalability, is limited by the independence assumption, making it unsuitable for scenarios with correlated data instances. Existing structured VI methods either focus on correlations among latent dimensions which lack scalability for modeling instance-level correlations, or are restricted to simple first-order dependencies, limiting their expressiveness. In this paper, we propose High-order Tree-structured Variational Inference (HoT-VI), that explicitly models $k$-order instance-level correlations among latent variables. By expressing the global posterior through overlapping $k$-dimensional local marginals, our method enables efficient parameterized sampling via a sequential procedure. To ensure the validity of these marginals, we introduce a conditional correlation parameterization method that guarantees positive definiteness of their correlation matrices. We further extend our method with a tree-structured backbone to capture more flexible dependency patterns. Extensive experiments on time-series and graph-structured datasets demonstrate that modeling higher-order correlations leads to significantly improved posterior approximations and better performance across various downstream tasks.

Safe deployment of machine learning (ML) models in safety-critical domains such as medical imaging requires detecting inputs with characteristics not seen during training, known as out-of-distribution (OOD) detection, to prevent unreliable predictions. Effective OOD detection after deployment could benefit from access to the training data, enabling direct comparison between test samples and the training data distribution to identify differences. State-of-the-art OOD detection methods, however, either discard the training data after deployment or assume that test samples and training data are centrally stored together, an assumption that rarely holds in real-world settings. This is because shipping the training data with the deployed model is usually impossible due to the size of training databases, as well as proprietary or privacy constraints. We introduce the Isolation Network, an OOD detection framework that quantifies the difficulty of separating a target test sample from the training data by solving a binary classification task. We then propose Decentralized Isolation Networks (DIsoN), which enables the comparison of training and test data when data-sharing is impossible, by exchanging only model parameters between the remote computational nodes of training and deployment. We further extend DIsoN with class-conditioning, comparing a target sample solely with training data of its predicted class. We evaluate DIsoN on four medical imaging datasets (dermatology, chest X-ray, breast ultrasound, histopathology) across 12 OOD detection tasks. DIsoN performs favorably against existing methods while respecting data-privacy. This decentralized OOD detection framework opens the way for a new type of service that ML developers could provide along with their models: providing remote, secure utilization of their training data for OOD detection services. Code available at: https://github.com/FelixWag/DIsoN

Class Incremental Learning (CIL) aims to continuously learn new categories while retaining the knowledge of old ones. Pre-trained models (PTMs) show promising capabilities in CIL. However, existing approaches that apply lightweight fine-tuning to backbones still induce parameter drift, thereby compromising the generalization capability of pre-trained models. Parameter drift can be conceptualized as a form of noise that obscures critical patterns learned for previous tasks. However, recent researches have shown that noise is not always harmful. For example, the large number of visual patterns learned from pre-training can be easily abused by a single task, and introducing appropriate noise can suppress some low-correlation features, thus leaving a margin for future tasks. To this end, we propose learning beneficial noise for CIL guided by information theory and propose Mixture of Noise (MiN), aiming to mitigate the degradation of backbone generalization from adapting new tasks. Specifically, task-specific noise is learned from high-dimension features of new tasks. Then, a set of weights is adjusted dynamically for optimal mixture of different task noise. Finally, MiN embeds the beneficial noise into the intermediate features to mask the response of inefficient patterns. Extensive experiments on six benchmark datasets demonstrate that MiN achieves state-of-the-art performance in most incremental settings, with particularly outstanding results in 50-steps incremental settings. This shows the significant potential for beneficial noise in continual learning. Code is available at https://github.com/ASCIIJK/MiN-NeurIPS2025.

We prove the first generalization bound for large-margin halfspaces that is asymptotically tight in the tradeoff between the margin, the fraction of training points with the given margin, the failure probability and the number of training points.

The rapid proliferation of domain-specialized machine learning models presents a challenge: while individual models excel in specific domains, their performance varies significantly across diverse applications. This makes selecting the optimal model when faced with an unknown mixture of tasks, especially with limited or no data to estimate the mixture, a difficult problem. We address this challenge by formulating it as a multiple-source domain adaptation (MSA) problem. We introduce a novel, scalable algorithm that effectively routes each input to the best-suited model from a pool of available models. Our approach provides a strong performance guarantee: remarkably, for any mixture domain, the accuracy achieved by the best source model is maintained. This guarantee is established through a theoretical bound on the regret for new domains, expressed as a convex combination of the best regrets in the source domains, plus a concentration term that diminishes as the amount of source data increases. While our primary contributions are theoretical and algorithmic, we also present empirical results demonstrating the effectiveness of our approach.

Reasoning-enhanced large language models (RLLMs), whether explicitly trained for reasoning or prompted via chain-of-thought (CoT), have achieved state-of-the-art performance on many complex reasoning tasks. However, we uncover a surprising and previously overlooked phenomenon: explicit CoT reasoning can significantly degrade instruction-following accuracy. Evaluating 20+ models on two benchmarks: IFEval (with simple, rule-verifiable constraints) and ComplexBench (with complex, compositional constraints), we consistently observe performance drops when CoT prompting is applied. Through large-scale case studies and an attention-based analysis, we identify common patterns where reasoning either helps (e.g., with formatting or lexical precision) or hurts (e.g., by neglecting simple constraints or introducing unnecessary content). We propose a metric, constraint attention, to quantify model focus during generation and show that CoT reasoning often diverts attention away from instruction-relevant tokens. To mitigate these effects, we introduce and evaluate four strategies: in-context learning, self-reflection, self-selective reasoning, and classifier-selective reasoning. Our results demonstrate that selective reasoning strategies, particularly classifier-selective reasoning, can substantially recover lost performance. To our knowledge, this is the first work to systematically expose reasoning-induced failures in instruction-following and offer practical mitigation strategies.

We study *gradient descent* (GD) with a constant stepsize for $\ell_2$-regularized logistic regression with linearly separable data. Classical theory suggests small stepsizes to ensure monotonic reduction of the optimization objective, achieving exponential convergence in $\widetilde{\mathcal{O}}(\kappa)$ steps with $\kappa$ being the condition number. Surprisingly, we show that this can be *accelerated* to $\widetilde{\mathcal{O}}(\sqrt{\kappa})$ by simply using a large stepsize---for which the objective evolves *nonmonotonically*. The acceleration brought by large stepsizes extends to minimizing the population risk for separable distributions, improving on the best-known upper bounds on the number of steps to reach a near-optimum. Finally, we characterize the largest stepsize for the local convergence of GD, which also determines the global convergence in special scenarios. Our results extend the analysis of Wu et al. (2024) from convex settings with minimizers at infinity to strongly convex cases with finite minimizers.

The ability to extract oriented edges from visual input is a core computation across animal vision systems. Orientation maps, long associated with the layered architecture of the mammalian visual cortex, systematically organise neurons by their preferred edge orientation. Despite lacking cortical structures, the *Drosophila melanogaster* brain contains feature-selective neurons and exhibits complex visual detection capacity, raising the question of whether map-like vision representations can emerge without cortical infrastructure. We integrate a complete fruit fly brain connectome with biologically grounded spiking neuron models to simulate neuroprocessing in the fly visual system. By driving the network with oriented stimuli and analysing downstream responses, we show that coherent orientation maps can emerge from purely connectome-constrained dynamics. These results suggest that species of independent origin could evolve similar visual structures.

Discrete diffusion models, like continuous diffusion models, generate high-quality samples by gradually undoing noise applied to datapoints with a Markov process. Gradual generation in theory comes with many conceptual benefits; for example, inductive biases can be incorporated into the noising Markov process. In practice, however, the consistently best performing discrete diffusion model is masking diffusion, which does not denoise gradually. Here we explain the superior performance of masking diffusion by noting that it makes use of a fundamental difference between continuous and discrete Markov processes: discrete Markov processes evolve by discontinuous jumps at a fixed rate and, unlike other discrete diffusion models, masking diffusion builds in the known distribution of jump times and only learns where to jump to. We show that we can similarly bake in the known distribution of jump times into any discrete diffusion model. The resulting models -- schedule-conditioned diffusion (SCUD) -- generalize classical discrete diffusion and masking diffusion. By applying SCUD to models with noising processes that incorporate inductive biases on images, text, and protein data, we build diffusion models that outperform masking.

We study value-iteration (VI) algorithms for solving general (a.k.a. multichain) Markov decision processes (MDPs) under the average-reward criterion, a fundamental but theoretically challenging setting. Beyond the difficulties inherent to all average-reward problems posed by the lack of contractivity and non-uniqueness of solutions to the Bellman operator, in the multichain setting an optimal policy must solve the navigation subproblem of steering towards the best connected component, in addition to optimizing long-run performance within each component. We develop algorithms which better solve this navigational subproblem in order to achieve faster convergence for multichain MDPs, obtaining improved rates of convergence and sharper measures of complexity relative to prior work. Many key components of our results are of potential independent interest, including novel connections between average-reward and discounted problems, optimal fixed-point methods for discounted VI which extend to general Banach spaces, new sublinear convergence rates for the discounted value error, and refined suboptimality decompositions for multichain MDPs. Overall our results yield faster convergence rates for discounted and average-reward problems and expand the theoretical foundations of VI approaches.