NeurIPS 2025 论文列表

Understanding when learning is possible is a fundamental task in the theory of machine learning. However, many characterizations known from the literature deal with abstract learning as a mathematical object and ignore the crucial question: when can learning be implemented as a computer program? We address this question for universal online learning, a generalist theoretical model of online binary classification, recently characterized by Bousquet et al. (STOC 2021). In this model, there is no hypothesis fixed in advance; instead, Adversary—playing the role of Nature—can change their mind as long as local consistency with the given class of hypotheses is maintained. We require Learner to achieve a finite number of mistakes while using a strategy that can be implemented as a computer program. We show that universal online learning does not imply computable universal online learning, even if the class of hypotheses is relatively easy from a computability-theoretic perspective. We then study the agnostic variant of computable universal online learning and provide an exact characterization of classes that are learnable in this sense. We also consider a variant of proper universal online learning and show exactly when it is possible. Together, our results give a more realistic perspective on the existing theory of online binary classification and the related problem of inductive inference.

We revisit the problem of constructing predictive confidence sets for which we wish to obtain some type of conditional validity. We provide new arguments showing how ``split conformal'' methods achieve near desired coverage levels with high probability, a guarantee conditional on the validation data rather than marginal over it. In addition, we directly consider (approximate) conditional coverage, where, e.g., conditional on a covariate $X$ belonging to some group of interest, we seek a guarantee that a predictive set covers the true outcome $Y$. We show that the natural method of performing quantile regression on a held-out (validation) dataset yields minimax optimal guarantees of coverage in these cases. Complementing these positive results, we also provide experimental evidence highlighting work that remains to develop computationally efficient valid predictive inference methods.

Boosting is a key method in statistical learning, allowing for converting weak learners into strong ones. While well studied in the realizable case, the statistical properties of weak-to-strong learning remain less understood in the agnostic setting, where there are no assumptions on the distribution of the labels. In this work, we propose a new agnostic boosting algorithm with substantially improved sample complexity compared to prior works under very general assumptions. Our approach is based on a reduction to the realizable case, followed by a margin-based filtering of high-quality hypotheses. Furthermore, we show a nearly-matching lower bound, settling the sample complexity of agnostic boosting up to logarithmic factors.

We say that a classifier is $\text{\emph{adversarially robust}}$ to perturbations of norm $r$ if, with high probability over a point $x$ drawn from the input distribution, there is no point within distance $\le r$ from $x$ that is classified differently. The $\text{\emph{boundary volume}}$ is the probability that a point falls within distance $r$ of a point with a different label. This work studies the task of learning a hypothesis with small boundary volume, where the input is distributed as a subgaussian isotropic log-concave distribution over $\mathbb{R}^d$. Linear threshold functions are adversarially robust; they have boundary volume proportional to $r$. Such concept classes are efficiently learnable by polynomial regression, which produces a polynomial threshold function (PTF), but PTFs in general may have boundary volume $\Omega(1)$, even for $r \ll 1$. We give an algorithm that agnostically learns linear threshold functions and returns a classfier with boundary volume $O(r+\varepsilon)$ at radius of perturbation $r$. The time and sample complexity of $d^{\tilde{O}(1/\varepsilon^2)}$ matches the complexity of polynomial regression. Our algorithm augments the classic approach of polynomial regression with three additional steps:\ $\quad$ a) performing the $\ell_1$-error regression under $\ell_1$ noise sensitivity constraints,\ $\quad$ b) a structured partitioning and rounding step that returns a Boolean classifier with error $\mathrm{opt} + O(\varepsilon)$ and noise sensitivity $O(r+\varepsilon)$ simultaneously, and \ $\quad c)$ a local corrector that ``smooths'' a function with low noise sensitivity into a function that is adversarially robust.

This work analyzes singular-value spectra of weight matrices in pretrained transformer models to understand how information is stored at both ends of the spectrum. Using Random Matrix Theory (RMT) as a zero information hypothesis, we associate agreement with RMT as evidence of randomness and deviations as evidence for learning. Surprisingly, we observe pronounced departures from RMT not only among the largest singular values - the usual outliers - but also among the smallest ones. A comparison of the associated singular vectors with the eigenvectors of the activation covariance matrices shows that there is considerable overlap wherever RMT is violated. Thus, significant directions in the data are captured by small singular values and their vectors as well as by the large ones. We confirm this empirically: zeroing out the singular values that deviate from RMT raises language-model perplexity far more than removing values from the bulk, and after fine-tuning the smallest decile can be the third most influential part of the spectrum. To explain how vectors linked to small singular values can carry more information than those linked to larger values, we propose a linear random-matrix model. Our findings highlight the overlooked importance of the low end of the spectrum and provide theoretical and practical guidance for SVD-based pruning and compression of large language models.

We propose Metric Automata Theory, an elegant generalisation of classic Automata Theory to continuous dynamical systems, that constitutes a unifying theory of all kinds of Recurrent Neural Networks (RNNs), including widely-adopted architectures such as xLSTM and State Space Models (SSMs). The theory allows one to analyse RNNs both in the finite and unbounded precision settings seamlessly, while utilising fundamental results of Automata Theory. It also provides a novel notion of robustness that guarantees numerical stability and contributes to stability of learning. We employ the theory to prove a comprehensive set of expressivity results for widely-adopted RNNs, with a focus on robustness and finite-precision. Notably, we contrast the capabilities of xLSTM and SSMs for robustly modelling all star-free regular languages—xLSTM can do so, while SSMs cannot robustly recognize the FLIP-FLOP language. Thus we give a novel perspective on the importance of non-linear recurrences, giving insight for why xLSTM shows superior performance to SSMs on several tasks. We provide an improved understanding of the capabilities of Mamba, a popular SSM model. We show that Mamba is not generally capable of recognising the star-free languages under finite-precision, which is seemingly in contrast with the existing theoretical and empirical results for SSMs. We clarify the picture, by showing that Mamba admits a piecewise-linearly separable state space that allows it to approximate star-free languages, with some length-generalisation abilities. At the same time, Mamba does not admit such state spaces for languages like Parity. This explains why empirically Mamba performs well on star-free languages, and fails on Parity.

Despite the promising results of disentangled representation learning in discovering latent patterns in graph-structured data, few studies have explored disentanglement for hypergraph-structured data. Integrating hyperedge disentanglement into hypergraph neural networks enables models to leverage hidden hyperedge semantics, such as unannotated relations between nodes, that are associated with labels. This paper presents an analysis of hyperedge disentanglement from a category-theoretical perspective and proposes a novel criterion for disentanglement derived from the naturality condition. Our proof-of-concept model experimentally showed the potential of the proposed criterion by successfully capturing functional relations of genes (nodes) in genetic pathways (hyperedges).

Recent works have introduced new equivariant neural networks, motivated by their improved generalization compared to traditional deep neural networks. While experiments support this advantage, the theoretical understanding of their generalization properties remains limited. In this paper, we analyze the generalization capabilities of Group Convolutional Neural Networks (GCNNs) with the ReLU activation function through the lens of Vapnik-Chervonenkis (VC) dimension theory. We investigate how architectural factors—such as the number of layers, weights, and input dimensions—affect the VC dimension. A key challenge in our analysis is proving a lower bound on the VC dimension, for which we introduce new techniques, establishing a novel connection between GCNNs and standard deep neural networks. Additionally, we compare our derived bounds to those known for fully connected neural networks. Our results extend previous findings on the VC dimension of continuous GCNNs with two layers, offering new insights into their generalization behavior, particularly their dependence on input resolution.

The quadratic complexity of self‑attention prevents transformers from scaling effectively to long input sequences. On the other hand, modern GPUs and other specialized hardware accelerators are well-optimized for processing small input sequences in transformers during both training and inference. A natural question arises: can we take advantage of the efficiency of small transformers to deal with long input sequences? In this paper, we show that transformers with long input sequences (large transformers) can be efficiently simulated by transformers that can only take short input sequences (small transformers). Specifically, we prove that any transformer with input length $N$ can be efficiently simulated by only $O((N/M)^2)$ transformers with input length $M \ll N$, and that this cannot be improved in the worst case. However, we then prove that in various natural scenarios including average-case inputs, sliding window masking and attention sinks, the optimal number $O(N/M)$ of small transformers suffice.

In recent years, there has been much interest in understanding the generalization behavior of interpolating predictors, which overfit on noisy training data. Whereas standard analyses are concerned with whether a method is consistent or not, recent observations have shown that even inconsistent predictors can generalize well. In this work, we revisit the classic interpolating Nadaraya-Watson (NW) estimator (also known as Shepard's method), and study its generalization capabilities through this modern viewpoint. In particular, by varying a single bandwidth-like hyperparameter, we prove the existence of multiple overfitting behaviors, ranging non-monotonically from catastrophic, through benign, to tempered. Our results highlight how even classical interpolating methods can exhibit intricate generalization behaviors. In addition, for the purpose of tuning the hyperparameter, the results suggest that over-estimating the intrinsic dimension of the data is less harmful than under-estimating it. Numerical experiments complement our theory, demonstrating the same phenomena.

We sharply characterize the optimal first-order query complexity of agnostic active learning for all concept classes, and propose a new general active learning algorithm which achieves it. Remarkably, the optimal query complexity admits a leading term which is always strictly smaller than the sample complexity of passive supervised learning (by a factor proportional to the best-in-class error rate). This was not previously known to be possible in the agnostic setting. For comparison, in all previous general analyses, the leading term exhibits an additional factor, such as the disagreement coefficient or related complexity measure, and therefore only provides improvements over passive learning in restricted cases. The present work completely removes such factors from the leading term, implying that $\textit{every}$ concept class benefits from active learning in the non-realizable case. The results established in this work resolve an important long-standing open question central to the past two decades of research on the theory of agnostic active learning.

Models such as Word2Vec and GloVe construct word embeddings based on the co-occurrence probability $P(i,j)$ of words $i$ and $j$ in text corpora. The resulting vectors $W_i$ not only group semantically similar words but also exhibit a striking linear analogy structure---for example, $W_{\text{king}} - W_{\text{man}} + W_{\text{woman}} \approx W_{\text{queen}}$---whose theoretical origin remains unclear. Previous observations indicate that this analogy structure: (i) already emerges in the top eigenvectors of the matrix $M(i,j) = P(i,j)/P(i)P(j)$, (ii) strengthens and then saturates as more eigenvectors of $M (i, j)$, which controls the dimension of the embeddings, are included, (iii) is enhanced when using $\log M(i,j)$ rather than $M(i,j)$, and (iv) persists even when all word pairs involved in a specific analogy relation (e.g., king--queen, man--woman) are removed from the corpus. To explain these phenomena, we introduce a theoretical generative model in which words are defined by binary semantic attributes, and co-occurrence probabilities are derived from attribute-based interactions. This model analytically reproduces the emergence of linear analogy structure and naturally accounts for properties (i)--(iv). It can be viewed as giving fine-grained resolution into the role of each additional embedding dimension. It is robust to various forms of noise and agrees well with co-occurrence statistics measured on Wikipedia and the analogy benchmark introduced by Mikolov et al.

Ensemble learning is a popular technique to improve the accuracy of machine learning models. It traditionally hinges on the rationale that aggregating multiple weak models can lead to better models with lower variance and hence higher stability, especially for discontinuous base learners. In this paper, we provide a new perspective on ensembling. By selecting the most frequently generated model from the base learner when repeatedly applied to subsamples, we can attain exponentially decaying tails for the excess risk, even if the base learner suffers from slow (i.e., polynomial) decay rates. This tail enhancement power of ensembling applies to base learners that have reasonable predictive power to begin with and is stronger than variance reduction in the sense of exhibiting rate improvement. We demonstrate how our ensemble methods can substantially improve out-of-sample performances in a range of numerical examples involving heavy-tailed data or intrinsically slow rates.

Curriculum learning has emerged as an effective strategy to enhance the training efficiency and generalization of machine learning models. However, its theoretical underpinnings remain relatively underexplored. In this work, we develop a theoretical framework for curriculum learning based on biased regularized empirical risk minimization (RERM), identifying conditions under which curriculum learning provably improves generalization. We introduce a sufficient condition that characterizes a "good" curriculum and analyze a multi-task curriculum framework, where solving a sequence of convex tasks can facilitate better generalization. We also demonstrate how these theoretical insights translate to practical benefits when using stochastic gradient descent (SGD) as an optimization method. Beyond convex settings, we explore the utility of curriculum learning for non-convex tasks. Empirical evaluations on synthetic datasets and MNIST validate our theoretical findings and highlight the practical efficacy of curriculum-based training.

Equivariant neural networks provide a principled framework for incorporating symmetry into learning architectures and have been extensively analyzed through the lens of their *separation power*, that is, the ability to distinguish inputs modulo symmetry. This notion plays a central role in settings such as graph learning, where it is often formalized via the Weisfeilern–Leman hierarchy. In contrast, the *universality* of equivariant models—their capacity to approximate target functions—remains comparatively underexplored. In this work, we investigate the approximation power of equivariant neural networks beyond separation constraints. We show that separation power does not fully capture expressivity: models with identical separation power may differ in their approximation ability. To demonstrate this, we characterize the universality classes of shallow invariant networks, providing a general framework for understanding which functions these architectures can approximate. Since equivariant models reduce to invariant ones under projection, this analysis yields sufficient conditions under which shallow equivariant networks fail to be universal. Conversely, we identify settings where shallow models do achieve separation-constrained universality. These positive results, however, depend critically on structural properties of the symmetry group, such as the existence of adequate normal subgroups, which may not hold in important cases like permutation symmetry.

This paper addresses the problem of inverse covariance (also known as precision matrix) estimation in high-dimensional settings. Specifically, we focus on two classes of estimators: linear shrinkage estimators with a target proportional to the identity matrix, and estimators derived from data augmentation (DA). Here, DA refers to the common practice of enriching a dataset with artificial samples—typically generated via a generative model or through random transformations of the original data—prior to model fitting. For both classes of estimators, we derive estimators and provide concentration bounds for their quadratic error. This allows for both method comparison and hyperparameter tuning, such as selecting the optimal proportion of artificial samples. On the technical side, our analysis relies on tools from random matrix theory. We introduce a novel deterministic equivalent for generalized resolvent matrices, accommodating dependent samples with specific structure. We support our theoretical results with numerical experiments.

We propose and analyze a new class of unbalanced weak optimal transport (OT) problems with total variation penalties, motivated by spatial resource allocation tasks. Unlike classical OT, our framework accommodates general unbalanced nonnegative measures and incorporates cost objectives that directly capture operational trade-offs between transport cost and supply–demand mismatch. In the general setting, we establish the existence of optimal solutions and a dual formulation. We then focus on the semi-discrete setting, where one measure is discrete and the other is absolutely continuous, a structure relevant to applications such as service area partitioning for facilities like schools or medical stations. Exploiting a tessellation-based structure, we derive the corresponding explicit optimality conditions. We further address a quantization problem that jointly optimizes the locations and weights of discrete support points, applicable to facility location tasks such as the cost-efficient deployment of battery swap stations or e-commerce warehouses, informed by demand-side data. The dual-tessellation structure also yields explicit gradient expressions, enabling efficient numerical optimization in finite dimensions.

Randomness is an unavoidable part of training deep learning models, yet something that traditional training data attribution algorithms fail to rigorously account for. They ignore the fact that, due to stochasticity in the initialisation and batching, training on the same dataset can yield different models. In this paper, we address this shortcoming through introducing _distributional_ training data attribution (d-TDA), the goal of which is to predict how the distribution of model outputs (over training runs) depends upon the dataset. Intriguingly, we find that _influence functions_ (IFs), a popular data attribution tool, are 'secretly distributional': they emerge from our framework as the limit to unrolled differentiation, without requiring restrictive convexity assumptions. This provides a new perspective on the effectiveness of IFs in deep learning. We demonstrate the practical utility of d-TDA in experiments, including improving data pruning for vision transformers and identifying influential examples with diffusion models.

Large language models are trained with tokenizers that map text to a fixed vocabulary, yet the resulting token distribution is highly imbalanced: a few words dominate the stream while most occur rarely. Recent practice favours ever-larger vocabularies, but it is unclear whether the benefit comes from better word segmentation or from amplifying this frequency skew. To this end, we perform a controlled study that scales the vocabulary of a constant-size Transformer from 24K to 196K symbols while holding data, compute and optimisation unchanged. Above 24K every common word is already a single token, so further growth only increases imbalance. Word-level loss decomposition shows that larger vocabularies reduce cross-entropy almost exclusively by lowering uncertainty on the ~$2,500$ most frequent words, even though loss on the rare tail rises. Same frequent words cover roughly $80\%$ of tokens in downstream benchmarks, this training advantage transfers intact. We further show that enlarging model parameters with a fixed tokenizer yields the same frequent-word benefit, revealing a shared mechanism behind vocabulary and model scaling. Our results recast “bigger vocabularies help” as “sharper frequency imbalance helps,” offering a simple, principled knob for tokenizer–model co-design and clarifying the loss dynamics that govern language-model scaling in pre-training.