ICML 2025 Papers

High-dimensional inference addresses scenarios where the dimension of the data approaches, or even surpasses, the sample size. In these settings, the regularized $M$-estimator is a common technique for inferring parameters. (Negahban et al., 2009) establish a unified framework for establishing convergence rates in the context of high-dimensional scaling, demonstrating that estimation errors are confined within a restricted set, and revealing fast convergence rates. The key assumption underlying their work is the convexity of the loss function. However, many loss functions in high-dimensional contexts are nonconvex. This leads to the question: if the loss function is nonconvex, do estimation errors still fall within a restricted set? If yes, can we recover convergence rates of the estimation error under nonconvex situations? This paper provides affirmative answers to these critical questions.

Social predictions do not passively describe the future; they actively shape it. They inform actions and change individual expectations in ways that influence the likelihood of the predicted outcome. Given these dynamics, to what extent can social events be predicted? This question was discussed throughout the 20th century by authors like Merton, Morgenstern, Simon, and others who considered it a central issue in social science methodology. In this work, we provide a modern answer to this old problem. Using recent ideas from performative prediction and outcome indistinguishability, we establish that one can always efficiently predict social events accurately, regardless of how predictions influence data. While achievable, we also show that these predictions are often undesirable, highlighting the limitations of previous desiderata. We end with a discussion of various avenues forward.

Continual learning has attracted increasing research attention in recent years due to its promising experimental results in real-world applications. In this paper, we study the issue of calibration in continual learning which reliably quantifies the uncertainty of model predictions. Conformal prediction (CP) provides a general framework for model calibration, which outputs prediction intervals or sets with a theoretical high coverage guarantee as long as the samples are exchangeable. However, the tasks in continual learning are learned in sequence, which violates the principle that data should be exchangeable. Meanwhile, the model learns the current task with limited or no access to data from previous tasks, which is not conducive to constructing the calibration set. To address these issues, we propose a CP-based method for model uncertainty quantification in continual learning (CPCL), which also reveals the connection between prediction interval length and forgetting. We analyze the oracle prediction interval in continual learning and theoretically prove the asymptotic coverage guarantee of CPCL. Finally, extensive experiments on simulated and real data empirically verify the validity of our proposed method.

The success of large generative models has driven a paradigm shift, leveraging massive multi-source data to enhance model capabilities. However, the interaction among these sources remains theoretically underexplored. This paper provides a first attempt to fill this gap by rigorously analyzing multi-source training in conditional generative modeling, where each condition represents a distinct data source. Specifically, we establish a general distribution estimation error bound in average total variation distance for conditional maximum likelihood estimation based on the bracketing number. Our result shows that when source distributions share certain similarities and the model is expressive enough, multi-source training guarantees a sharper bound than single-source training. We further instantiate the general theory on conditional Gaussian estimation and deep generative models including autoregressive and flexible energy-based models, by characterizing their bracketing numbers. The results highlight that the number of sources and similarity among source distributions improve the advantage of multi-source training. Simulations and real-world experiments validate our theory.

In this work, we develop a hypothesis testing framework to determine whether pairwise comparison data is generated by an underlying *generalized Thurstone model* $\mathcal{T}_F$ for a given choice function $F$. While prior work has predominantly focused on parameter estimation and uncertainty quantification for such models, we address the fundamental problem of minimax hypothesis testing for $\mathcal{T}_F$ models. We formulate this testing problem by introducing a notion of separation distance between general pairwise comparison models and the class of $\mathcal{T}_F$ models. We then derive upper and lower bounds on the critical threshold for testing that depend on the topology of the observation graph. For the special case of complete observation graphs, this threshold scales as $\Theta((nk)^{-1/2})$, where $n$ is the number of agents and $k$ is the number of comparisons per pair. Furthermore, we propose a hypothesis test based on our separation distance, construct confidence intervals, establish time-uniform bounds on the probabilities of type I and II errors using reverse martingale techniques, and derive minimax lower bounds using information-theoretic methods. Finally, we validate our results through experiments on synthetic and real-world datasets.

The field of *algorithms with predictions* incorporates machine learning advice in the design of online algorithms to improve real-world performance. A central consideration is the extent to which predictions can be trusted—while existing approaches often require users to specify an aggregate trust level, modern machine learning models can provide estimates of prediction-level uncertainty. In this paper, we propose *calibration* as a principled and practical tool to bridge this gap, demonstrating the benefits of calibrated advice through two case studies: the *ski rental* and *online job scheduling* problems. For ski rental, we design an algorithm that achieves near-optimal prediction-dependent performance and prove that, in high-variance settings, calibrated advice offers more effective guidance than alternative methods for uncertainty quantification. For job scheduling, we demonstrate that using a calibrated predictor leads to significant performance improvements over existing methods. Evaluations on real-world data validate our theoretical findings, highlighting the practical impact of calibration for algorithms with predictions.

Weak-to-Strong Generalization (Burns et al.,2024) is the phenomenon whereby a strong student, say GPT-4, learns a task from a weak teacher, say GPT-2, and ends up significantly outperforming the teacher. We show that this phenomenon does not require a complex and pretrained learner like GPT-4, can arise even in simple non-pretrained models, simply due to the size advantage of the student. But, we also show that there are inherint limits to the extent of such weak to strong generalization. We consider students and teachers that are random feature models, described by two-layer networks with a random and fixed bottom layer and trained top layer. A ‘weak’ teacher, with a small number of units (i.e. random features), is trained on the population, and a ‘strong’ student, with a much larger number of units (i.e. random features), is trained only on labels generated by the weak teacher. We demonstrate, prove, and understand how the student can outperform the teacher, even though trained only on data labeled by the teacher. We also explain how such weak-to-strong generalization is enabled by early stopping. We then show the quantitative limits of weak-to-strong generalization in this model, and in fact in a much broader class of models, for arbitrary teacher and student feature spaces and a broad class of learning rules, including when the student features are pre-trained or otherwise more informative. In particular, we show that in such models the student’s error can only approach zero if the teacher’s error approaches zero, and a strong student cannot “boost” a slightly-better-then-chance teacher to obtain a small error.

In this article, we explore the loss landscape of next-token prediction with transformers. Specifically, we focus on learning in-context n-gram language models with cross-entropy loss using a simplified two-layer transformer. We design a series of transformers that represent $k$-grams (for $k \leq n$) for which the gradient of the population loss approaches zero in the limit of both infinite sequence length and infinite parameter norm. This construction reveals a key property of the loss landscape: \emph{$k$-grams are stationary points of the population cross-entropy loss}, offering theoretical insights for widely observed empirical phenomena such as stage-wise learning dynamics and emergent phase transitions. These insights are further supported by comprehensive numerical experiments that illustrate the dynamics of learning $n$-grams, characterized by jumps between stationary points.

In this manuscript, we study the learning of deep attention neural networks, defined as the composition of multiple self-attention layers, with tied and low-rank weights. We first establish a mapping of such models to sequence multi-index models, a generalization of the widely studied multi-index model to sequential covariates, for which we establish a number of general results. In the context of Bayes-optimal learning, in the limit of large dimension $D$ and proportionally large number of samples $N$, we derive a sharp asymptotic characterization of the optimal performance as well as the performance of the best-known polynomial-time algorithm for this setting --namely approximate message-passing--, and characterize sharp thresholds on the minimal sample complexity required for better-than-random prediction performance. Our analysis uncovers, in particular, how the different layers are learned sequentially. Finally, we discuss how this sequential learning can also be observed in a realistic setup.

We introduce a new tool, Transductive Local Complexity (TLC), to analyze the generalization performance of transductive learning methods and motivate new transductive learning algorithms. Our work extends the idea of the popular Local Rademacher Complexity (LRC) to the transductive setting with considerable and novel changes compared to the analysis of typical LRC methods in the inductive setting. While LRC has been widely used as a powerful tool in the analysis of inductive models with sharp generalization bounds for classification and minimax rates for nonparametric regression, it remains an open problem whether a localized version of Rademacher complexity based tool can be designed and applied to transductive learning and gain sharp bound for transductive learning which is consistent with the inductive excess risk bound by (LRC). We give a confirmative answer to this open problem by TLC. Similar to the development of LRC, we build TLC by first establishing a novel and sharp concentration inequality for supremum of empirical processes for the gap between test and training loss in the setting of sampling uniformly without replacement. Then a peeling strategy and a new surrogate variance operator are used to derive the following excess risk bound in the transductive setting, which is consistent with that of the classical LRC based excess risk bound in the inductive setting. As an application of TLC, we use the new TLC tool to analyze the Transductive Kernel Learning (TKL) model, and derive sharper excess risk bound than that by the current state-of-the-art. As a result of independent interest, the concentration inequality for the test-train process is used to derive a sharp concentration inequality for the general supremum of empirical process involving random variables in the setting of sampling uniformly without replacement, with comparison to current concentration inequalities.

Existing works on the expressive power of neural networks typically assume real-valued parameters and exact mathematical operations during the evaluation of networks. However, neural networks run on actual computers can take parameters only from a small subset of the reals and perform inexact mathematical operations with round-off errors and overflows. In this work, we study the expressive power of floating-point neural networks, i.e., networks with floating-point parameters and operations. We first observe that for floating-point neural networks to represent all functions from floating-point vectors to floating-point vectors, it is necessary to distinguish different inputs: the first layer of a network should be able to generate different outputs for different inputs. We also prove that such distinguishability is sufficient, along with mild conditions on activation functions. Our result shows that with practical activation functions, floating-point neural networks can represent floating-point functions from a wide domain to all finite or infinite floats. For example, the domain is all finite floats for Sigmoid and tanh, and it is all finite floats of magnitude less than 1/8 times the largest float for ReLU, ELU, SeLU, GELU, Swish, Mish, and sin.

Personalized federated learning (PFL) offers a flexible framework for aggregating information across distributed clients with heterogeneous data. This work considers a personalized federated learning setting that simultaneously learns global and local models. While purely local training has no communication cost, collaborative learning among the clients can leverage shared knowledge to improve statistical accuracy, presenting an accuracy-communication trade-off in personalized federated learning. However, the theoretical analysis of how personalization quantitatively influences sample and algorithmic efficiency and their inherent trade-off is largely unexplored. This paper makes a contribution towards filling this gap, by providing a quantitative characterization of the personalization degree on the tradeoff. The results further offer theoretical insights for choosing the personalization degree. As a side contribution, we establish the minimax optimality in terms of statistical accuracy for a widely studied PFL formulation. The theoretical result is validated on both synthetic and real-world datasets and its generalizability is verified in a non-convex setting.

We study the algorithmic problem of robust mean estimation of an identity covariance Gaussian in the presence of mean-shift contamination. In this contamination model, we are given a set of points in $\mathbb{R}^d$ generated i.i.d. via the following process. For a parameter $\alpha<1/2$, the $i$-th sample $x_i$ is obtained as follows: with probability $1-\alpha$, $x_i$ is drawn from $\mathcal{N}(\mu, I)$, where $\mu \in \mathbb{R}^d$ is the target mean; and with probability $\alpha$, $x_i$ is drawn from $\mathcal{N}(z_i, I)$, where $z_i$ is unknown and potentially arbitrary. Prior work characterized the information-theoretic limits of this task. Specifically, it was shown that— in contrast to Huber contamination— in the presence of mean-shift contamination consistent estimation is possible. On the other hand, all known robust estimators in the mean-shift model have running times exponential in the dimension. Here we give the first computationally efficient algorithm for high-dimensional robust mean estimation with mean-shift contamination that can tolerate a constant fraction of outliers. In particular, our algorithm has near-optimal sample complexity, runs in sample-polynomial time, and approximates the target mean to any desired accuracy. Conceptually, our result contributes to a growing body of work that studies inference with respect to natural noise models lying in between fully adversarial and random settings.

Estimating the unknown reward functions driving agents' behavior is a central challenge in inverse games and reinforcement learning. This paper introduces a unified framework for reward function recovery in two-player zero-sum matrix games and Markov games with entropy regularization. Given observed player strategies and actions, we aim to reconstruct the underlying reward functions. This task is challenging due to the inherent ambiguity of inverse problems, the non-uniqueness of feasible rewards, and limited observational data coverage. To address these challenges, we establish reward function identifiability using the quantal response equilibrium (QRE) under linear assumptions. Building on this theoretical foundation, we propose an algorithm to learn reward from observed actions, designed to capture all plausible reward parameters by constructing confidence sets. Our algorithm works in both static and dynamic settings and is adaptable to incorporate other methods, such as Maximum Likelihood Estimation (MLE). We provide strong theoretical guarantees for the reliability and sample-efficiency of our algorithm. Empirical results demonstrate the framework’s effectiveness in accurately recovering reward functions across various scenarios, offering new insights into decision-making in competitive environments.

Neural network training is inherently sensitive to initialization and the randomness induced by stochastic gradient descent. However, it is unclear to what extent such effects lead to meaningfully different networks, either in terms of the models' weights or the underlying functions that were learned. In this work, we show that during the initial "chaotic" phase of training, even extremely small perturbations reliably causes otherwise identical training trajectories to diverge-an effect that diminishes rapidly over training time. We quantify this divergence through (i) $L^2$ distance between parameters, (ii) the loss barrier when interpolating between networks, (iii) $L^2$ and barrier between parameters after permutation alignment, and (iv) representational similarity between intermediate activations; revealing how perturbations across different hyperparameter or fine-tuning settings drive training trajectories toward distinct loss minima. Our findings provide insights into neural network training stability, with practical implications for fine-tuning, model merging, and diversity of model ensembles.

This paper explores how theory can guide and enhance practical algorithms, using Low-Rank Adaptation (LoRA) (Hu et al., 2022) in large language models as a case study. We rigorously prove that, under gradient descent, LoRA adapters align with specific singular subspaces of the one-step full fine-tuning gradient. This result suggests that, by properly initializing the adapters using the one-step full gradient, subspace alignment can be achieved immediately—applicable to both linear and nonlinear models. Building on our theory, we propose a theory-driven algorithm, LoRA-One, where the linear convergence (as well as generalization) is built and incorporating preconditioners theoretically helps mitigate the effects of ill-conditioning. Besides, our theory reveals connections between LoRA-One and other gradient-alignment-based methods, helping to clarify misconceptions in the design of such algorithms. LoRA-One achieves significant empirical improvements over LoRA and its variants across benchmarks in natural language understanding, mathematical reasoning, and code generation. Code is available at: https://github.com/YuanheZ/LoRA-One.

This paper investigates decentralized unlearning, aiming to eliminate the impact of a specific client on the whole decentralized system. However, decentralized communication characterizations pose new challenges for effective unlearning: the indirect connections make it difficult to trace the specific client's impact, while the dynamic topology limits the scalability of retraining-based unlearning methods. In this paper, we propose the first **P**rovable **D**ecentralized **U**nlearning algorithm under **D**ynamic **T**opologies called PDUDT. It allows clients to eliminate the influence of a specific client without additional communication or retraining. We provide rigorous theoretical guarantees for PDUDT, showing it is statistically indistinguishable from perturbed retraining. Additionally, it achieves an efficient convergence rate of $\mathcal{O}(\frac{1}{T})$ in subsequent learning, where $T$ is the total communication rounds. This rate matches state-of-the-art results. Experimental results show that compared with the Retrain method, PDUDT saves more than 99\% of unlearning time while achieving comparable unlearning performance.

Graph-based data structures have become powerful and ubiquitous tools for scalable approximate nearest-neighbor (ANN) search over the past decade. In spite of their apparent practical performance, there has only recently been progress on the **worst-case** performance of these data structures. Indeed, the influential work of Indyx and Xu (2023) introduced the key concept of $\alpha$-reachable graphs, showing that graphs constructed by the DiskANN algorithm (Subramanya, et. al. 2023) produce an $\left(\frac{\alpha+1}{\alpha-1}\right)$-approximate solution with a simple best-first search that runs in poly-logarithmic query time. In our work, we improve and generalize this analysis as follows: - We introduce **sorted** $\alpha$-reachable graphs, and use this notion to obtain a stronger approximation factor of $\frac{\alpha}{\alpha-1}$ for the DiskANN algorithm on Euclidean metrics. - We present the **first** worst-case theoretical analysis for the popular **beam-search** algorithm, which is used in practice to search these graphs for $k > 1$ candidate nearest neighbors. We also present empirical results validating the significance of sorted $\alpha$-reachable graphs, which aligns with our theoretical findings.

Markov processes serve as universal models for many real-world random processes. This paper presents a data-driven approach to learning these models through the spectral decomposition of the infinitesimal generator (IG) of the Markov semigroup. Its unbounded nature complicates traditional methods such as vector-valued regression and Hilbert-Schmidt operator analysis. Existing techniques, including physics-informed kernel regression, are computationally expensive and limited in scope, with no recovery guarantees for transfer operator methods when the time-lag is small. We propose a novel method leveraging the IG's resolvent, characterized by the Laplace transform of transfer operators. This approach is robust to time-lag variations, ensuring accurate eigenvalue learning even for small time-lags. Our statistical analysis applies to a broader class of Markov processes than current methods while reducing computational complexity from quadratic to linear in the state dimension. Finally, we demonstrate our theoretical findings in several experiments.