As large language models are increasingly trained and fine-tuned, practitioners need methods to identify which training data drive specific behaviors, particularly unintended ones. Training Data Attribution (TDA) methods address this by estimating datapoint influence. Existing approaches like influence functions are both computationally expensive and attribute based on single test examples, which can bias results toward syntactic rather than semantic similarity. To address these issues of scalability and influence to abstract behavior, we leverage interpretable structures within the model during the attribution. First, we introduce Concept Influence which attribute model behavior to semantic directions (such as linear probes or sparse autoencoder features) rather than individual test examples. Second, we show that simple probe-based attribution methods are first-order approximations of Concept Influence that achieve comparable performance while being over an order-of-magnitude faster. We empirically validate Concept Influence and approximations across emergent misalignment benchmarks and real post-training datasets, and demonstrate they achieve comparable performance to classical influence functions while being substantially more scalable. More broadly, we show that incorporating interpretable structure within traditional TDA pipelines can enable more scalable, explainable, and better control of model behavior through data.
We study two log-concave sampling problems: constrained sampling and composite sampling. First, we consider sampling from a distribution proportional to $e^{-f}$ on a convex set $K\subset\mathbb R^d$, where $f$ is convex. Using an epigraph lifting, we reduce this problem to uniform sampling up to a one-dimensional exponential tilt over a convex set in $\mathbb R^{d+1}$. We then apply the proximal sampler to the lifted distribution. We develop an unbiased implementation of its restricted Gaussian oracle by using the cutting-plane method and rejection sampling, requiring only a separation oracle for $K$ and evaluation and subgradient oracles for $f$, without relying on projections or barrier functions. Second, we study composite sampling from $e^{-f-h}$, where $f$ and $h$ are closed and convex. This structure arises naturally in Bayesian inference, where $f$ represents data fidelity and $h$ encodes prior information. By applying an epigraph lifting to $h$, we reduce composite sampling to constrained sampling in $\mathbb R^{d+1}$, allowing us to build on the approach developed in the first part. This leads to a double lifting formulation in $\mathbb R^{d+2}$, for which we implement the proximal sampler using different combinations of subgradient and proximal oracles for $f$ and $h$. We establish mixing time bounds in R\'enyi and $\chi^2$ divergences for both problems. Finally, we revisit the convergence analysis of the proximal sampler from a variational perspective. Starting from variational formulations of its two half-steps as optimization problems over probability measures, we use a dynamic formulation of entropy-regularized optimal transport to recast them as optimal control problems. This allows us to derive their governing PDEs directly through the lens of optimization, without first analyzing particle-level SDEs.
Batch size scheduling (BSS) plays a critical role in large-scale deep learning training, influencing both optimization dynamics and computational efficiency. Yet, its theoretical foundations remain poorly understood. In this work, we show that the functional scaling law (FSL) framework introduced in Li et al. (2025a) provides a principled lens for analyzing BSS. Specifically, we characterize the optimal BSS under a fixed data budget and show that its structure depends sharply on task difficulty. For easy tasks, optimal schedules keep increasing batch size throughout. In contrast, for hard tasks, the optimal schedule maintains small batch sizes for most of training and switches to large batches only in a late stage. To explain the emergence of late switching, we uncover a dynamical mechanism -- the fast catch-up effect -- which also manifests in large language model (LLM) pretraining. After switching from small to large batches, the loss rapidly aligns with the constant large-batch trajectory. Using FSL, we show that this effect stems from rapid forgetting of accumulated gradient noise, with the catch-up speed determined by task difficulty. Crucially, this effect implies that large batches can be safely deferred to late training without sacrificing performance, while substantially reducing data consumption. Finally, extensive LLM pretraining experiments -- covering both Dense and MoE architectures with up to 1.1B parameters and 1T tokens -- validate our theoretical predictions. Across all settings, late-switch schedules consistently outperform constant-batch and early-switch baselines.
Classical two-sample tests aim to determine whether two distributions are equal or not based on samples from each one. Even in an distribution-free setting, well-known methods such as the permutation test can provide statistical evidence that the two distributions differ. However, failure to reject equality does not necessarily establish that they are similar. In this work, we study the complementary problem of certifying distributional similarity with finite-sample confidence, without imposing assumptions on the underlying distributions. We show, however, that for continuous distributions, any distribution-free upper confidence bound on the total variation (TV) distance must be trivial with high probability, even if the distributions are identical. Thus, certifying similarity in TV with a distribution-free guarantee is practically impossible. We therefore examine the blurred TV distance, a relaxation of TV distance that enables us to quantify similarity after applying a common blurring operation while retaining the interpretation of TV. We construct distribution-free upper and lower confidence bounds on blurred TV, establish their finite-sample validity, and examine its properties in high dimensions.
In this work, we provide a sharp theory of scaling laws for two-layer neural networks trained on a class of hierarchical multi-index targets, in a genuinely representation-limited regime. We derive exact information-theoretic scaling laws for subspace recovery and prediction error, revealing how the hierarchical features of the target are sequentially learned through a cascade of phase transitions. We further show that these optimal rates are achieved by a simple, target-agnostic spectral estimator, which can be interpreted as the small learning-rate limit of gradient descent on the first-layer weights. Once an adapted representation is identified, the readout can be learned statistically optimally, using an efficient procedure. As a consequence, we provide a unified and rigorous explanation of scaling laws, plateau phenomena, and spectral structure in shallow neural networks trained on such hierarchical targets.
Recent progress has rapidly advanced our understanding of the mechanisms underlying in-context learning in modern attention-based neural networks. However, existing results focus exclusively on unimodal data; in contrast, the theoretical underpinnings of in-context learning for multi-modal data remain poorly understood. We introduce a mathematically tractable framework for studying multi-modal learning and explore when transformer-like architectures can recover Bayes-optimal performance in-context. To model multi-modal problems, we assume the observed data arises from a latent factor model. Our first result comprises a negative take on expressibility: we prove that single-layer self-attention models fail to recover the Bayes-optimal predictor uniformly over the task distribution. To address this limitation, we introduce a novel, linearized cross-attention mechanism, which we study in the regime where both the number of cross-attention layers and the context length are large. We show that this cross-attention mechanism is provably Bayes optimal when optimized using gradient flow. Our results underscore the benefits of depth for in-context learning and establish the provable utility of cross-attention for multi-modal distributions.
A common approach to perform PCA on probability measures is to embed them into a Hilbert space where standard functional PCA techniques apply. While convergence rates for estimating the embedding of a single measure from $m$ samples are well understood, the literature has not addressed the setting involving multiple measures. In this paper, we study PCA in a double asymptotic regime where $n$ probability measures are observed, each through $m$ samples. We derive convergence rates of the form $n^{-1/2} + m^{-\alpha}$ for the empirical covariance operator and the PCA excess risk, where $\alpha>0$ depends on the chosen embedding. This characterizes the relationship between the number $n$ of measures and the number $m$ of samples per measure, revealing a sparse (small $m$) to dense (large $m$) transition in the convergence behavior. Moreover, we prove that the dense-regime rate is minimax optimal for the empirical covariance error. Our numerical experiments validate these theoretical rates and demonstrate that appropriate subsampling preserves PCA accuracy while reducing computational cost.
Constructing accurate and computationally efficient surrogate models for predicting dynamical system responses is critical in many engineering domains, yet remains challenging due to the strongly nonlinear and high-dimensional mapping from external excitations and system parameters to system responses. This work introduces a novel Function-on-Function Nonlinear AutoRegressive model with eXogenous inputs (F2NARX), which reformulates the recently proposed $\mathcal{F}$-NARX method from a function-on-function regression perspective. The proposed framework substantially improves predictive efficiency while maintaining high accuracy. By combining principal component analysis with Gaussian process regression, F2NARX further enables probabilistic predictions of dynamical responses via the unscented transform in an autoregressive manner. The effectiveness of the method is demonstrated through case studies of varying complexity. Results show that F2NARX outperforms state-of-the-art NARX model by orders of magnitude in efficiency while achieving higher accuracy in general. The probabilistic prediction capability of F2NARX is further demonstrated by combining it with active learning for first-passage probability estimation, showing its potential for reliability-oriented applications.
Self-attention has greatly contributed to the success of the widely used Transformer architecture by enabling learning from data with long-range dependencies. In an effort to improve performance, a gated attention model that leverages a gating mechanism within the multi-head self-attention has recently been proposed as a promising alternative. Gated attention has been empirically demonstrated to increase the expressiveness of low-rank mapping in standard attention and even to eliminate the attention sink phenomenon. Despite its efficacy, a clear theoretical understanding of gated attention's benefits remains lacking in the literature. To close this gap, we rigorously show that each entry in a gated attention matrix or a multi-head self-attention matrix can be written as a hierarchical mixture of experts. By recasting learning as an expert estimation problem, we demonstrate that gated attention is more sample-efficient than multi-head self-attention. In particular, while the former needs only a polynomial number of data points to recover the model parameters, the latter requires exponentially many data points to achieve the same estimation error. Furthermore, our analysis also provides a theoretical justification for why gated attention is more effective at mitigating the attention sink phenomenon and the low-rank bottleneck when a gate is placed at the output of the scaled dot product attention or the value map rather than at other positions in the multi-head self-attention architecture.
We establish dual representation and stability results for multimarginal optimal transport (MOT) with bounded continuous costs. Our approach uses mutually $c$-conjugate potentials: each component is the largest function compatible with the dual constraint when the others are fixed. A single sweep of successive component updates produces such a family from any real-valued feasible family while preserving contact points. The resulting potentials are bounded and upper semicontinuous. On fixed compact metric spaces, normalization at fixed reference points yields a compact class of continuous potentials in the product uniform norm. This class is independent of the marginals and contains a dual maximizer for every marginal tuple. We prove its Lipschitz dependence on the cost in Hausdorff distance. Under uniform cost convergence and weak marginal convergence, transport values converge. The corresponding normalized optimizer sets are upper stable in the product uniform norm, without uniqueness assumptions. We also obtain value bounds in total variation and, for coordinatewise Lipschitz costs, Wasserstein distance. For coordinatewise Lipschitz reference costs, a power-growth condition on the dual objective gives one-sided H\"older estimates for optimizer sets under joint perturbations. On finite spaces, polyhedral structure yields linear growth and one-sided Lipschitz estimates with reference-dependent constants. On Polish spaces, total boundedness of the coordinate cost profiles in the uniform norm yields a compact normalized representation. For a fixed cost satisfying this condition, value continuity and upper stability of normalized conjugate optimizer sets hold under weak marginal convergence.
Instrumental variable estimation has emerged as a standard approach to mitigating confounding bias in the social sciences and epidemiology, where conducting randomized experiments can be too costly or infeasible. However, justifying the validity of the instrument is frequently challenging. We highlight a problem generally neglected in arguments for instrumental variable validity: the presence of an "aggregate treatment variable", where the treatment (e.g., education, GDP, caloric intake) is composed of finer-grained, unobserved components that each may have a different effect on the outcome. While the aggregation problem itself is general, our focus is on instrumental variable estimation in a linear setting, the regime underlying much of applied instrumental variable practice. We show that the causal effect of an aggregate treatment is generally ambiguous, as it depends on how an intervention on the aggregate is instantiated at the component level. We formalize this relation using the aggregate-constrained component intervention distribution (ACID). We then identify two key conditions under which standard instrumental variable estimators identify the aggregate effect. The contrived nature of these conditions implies major limitations on the interpretation of instrumental variable estimates based on aggregate treatments and highlights the need for a broader justificatory base for the exclusion restriction in such settings.
Source inversion problems under sparse measurements for parabolic partial differential equations (PDEs) are particularly challenging to solve using physics-informed neural networks (PINNs), due to their severe ill-posedness and the multiple unknowns involved including the source function and the PDE parameters. Although the neural tangent kernel (NTK) of PINNs has been widely used in forward problems involving a single neural network, its extension to inverse problems involving multiple neural networks remains less explored. In this work, we propose a weighted adaptive approach based on the NTK of PINNs including multiple networks representing the solution, the unknown source, and the PDE parameters. The key idea behind our methodology is to simultaneously solve the joint recovery of these unknowns thereby using the underlying PDE as a constraint that couples multiple unknowns functional parameters, leading to more efficient use of the limited information in the measurements. Furthermore, we provide a theoretical analysis of the approximation error of PINNs in a source inverse problem governed by the advection-diffusion equation (ADE) with known velocity and diffusion coefficients. In particular, we derive an error bound for the estimated source function in terms of the PINN's training error, the noise level and the discretization error. This analysis provides theoretical insight into the factors influencing the approximation accuracy. We present numerical experiments on 2D and 3D ADE using different types of measurements data reflecting practical engineering systems. Our results demonstrate the successful performance of PINNs in solving our source inversion problem and its robustness to noise in the measurements.
U-statistics are a fundamental class of estimators that generalize the sample mean and underpin much of nonparametric statistics. Although extensively studied in both statistics and probability, key challenges remain: their high computational cost - addressed partly through incomplete U-statistics - and their non-standard asymptotic behavior in the degenerate case, which typically requires resampling methods for hypothesis testing. This paper presents a novel perspective on U-statistics, grounded in hypergraph theory and combinatorial designs. We identify the dependency graph of an incomplete U-statistic with the line graph of its design hypergraph and use Stein's method to derive Berry-Esseen bounds for deterministic designs under full degeneracy. We further extend these results to studentized statistics using a variance estimator tailored to the fully degenerate setting. The resulting bounds establish Gaussian limits for superlinear design sizes with fixed kernel order and accommodate diverging order under separate conditions. We also develop efficient algorithms for constructing equireplicate designs, a subclass of deterministic designs, which yield minimum-variance incomplete U-statistics in the second-order case. Beyond its theoretical contributions, our framework provides permutation-free alternatives to tests based on fully degenerate U-statistics, as demonstrated in experiments with the two-sample Maximum Mean Discrepancy test and a one-sample mean test based on a higher-order U-statistic.
Machine learning (ML) is increasingly central to scientific discovery, but translating ML predictions into rigorous scientific claims still requires statistical inference on gold-standard labeled data. A setting that has emerged across the sciences provides the analyst with a small labeled sample, a much larger unlabeled sample, and a pre-trained predictor, with the goal of constructing a confidence interval for a quantity of interest, as was formalized in Angelopoulos et al. (2023). In this ML-driven scientific discovery literature, labeled data is typically the binding constraint on precision while computation is comparatively unconstrained, motivating the design question of how additional computation can extract more inferential value from a fixed labeled set, a fixed unlabeled set, and a fixed predictor. To address this question, we propose the Prediction-Augmented Residual Tree (PART) estimator. PART takes the same inputs as the original PPI estimator and replaces its single global rectifier with locally computed rectifiers obtained from an adaptive partition of the feature space. We show that PART outperforms existing methods by producing tighter confidence intervals across real-world datasets from ecology, astronomy, and census reports, among other domains. We then describe and analyze PAQ, an estimator based on numerical quadrature techniques. Under appropriate assumptions on the input data, we show that the variance of PAQ shrinks at a rate of $O(N^{-1} + n^{-4})$, showing that there are settings where the rate of $O(N^{-1}+n^{-1})$ of existing methods can be provably broken.
We study the statistical accuracy of Schr\"odinger bridge models based on an Ornstein-Uhlenbeck reference process. Given a known, compactly supported initial distribution and independent samples from an unknown target distribution, we estimate a right Schr\"odinger log-potential by minimizing empirical Kullback-Leibler (KL) risk over a prescribed function class. For bounded sub-Gaussian target densities, potentially with unbounded support, we establish a non-asymptotic high-probability oracle inequality for the KL-divergence between the target and estimated terminal marginals. Under additional two-sided Gaussian bounds and smoothness assumptions on the target log-density, we derive approximation guarantees for sparse ReLU networks and establish a KL convergence rate of $\widetilde{O}(n^{-2s / (2s + d)})$ for exact empirical risk minimization. This matches, up to logarithmic factors, the classical nonparametric KL-risk convergence exponent, while allowing target densities on $\mathbb R^d$ whose tails are controlled by Gaussian envelopes.
Benign overfitting occurs when a model fits noisy training data exactly and still generalizes well. We establish this behavior for finite-width two-layer networks with the standard ReLU activation, trained by gradient flow in fixed dimension. Our data-independent initialization combines Gaussian weights with neurons initially active only on shrinking spherical caps. For uniform spherical inputs and any fixed bounded conditional response law, training and excess risks vanish in probability uniformly over all sufficiently late times. The parameters also converge, with probability tending to one, to an interpolating limit whose excess risk vanishes in probability. We use one schedule depending only on sample size and dimension. Our proof connects the network to a kernel flow whose limit combines ridge regression with localized spikes that fit the training residuals. We give explicit sample-size, width and training-time bounds.
This paper introduces the Difference-in-Differences Bayesian Causal Forest (DiD-BCF), a novel non-parametric model addressing key challenges in DiD estimation, such as staggered adoption and heterogeneous treatment effects. DiD-BCF provides a unified framework for estimating Average (ATE), Group-Average (GATE), and Conditional Average Treatment Effects (CATE). By structuring separate regression forests for cohort-specific levels, common untreated trends and treatment effects, the model directly embeds conditional parallel trends and no-anticipation restrictions into Bayesian nonparametrics. We establish non-parametric identification and integrated treatment effect consistency of the exact treatment effect posterior distribution. Extensive simulations demonstrate DiD-BCF's superior performance over established benchmarks, particularly under non-linearity, selection biases, and effect heterogeneity. Applied to U.S. minimum wage policy, the model uncovers significant conditional treatment effect heterogeneity related to county population, insights obscured by traditional methods. DiD-BCF offers a robust and versatile tool for more nuanced causal inference in modern DiD applications.
Transformers consist of diverse building blocks, such as embedding layers, normalization layers, self-attention mechanisms, and point-wise feedforward networks. Thus, understanding the differences and interactions among these blocks is important. In this paper, we uncover a clear Sharpness Disparity across these blocks, which emerges early in training and intriguingly persists throughout the training process. Motivated by this finding, we propose Blockwise Learning Rate (LR), a strategy that tailors the LR to each block's sharpness, accelerating large language model (LLM) pre-training. By integrating Blockwise LR into AdamW, we consistently achieve lower terminal loss and nearly $2\times$ speedup compared to vanilla AdamW. We demonstrate this acceleration across GPT-2 and LLaMA, with model sizes ranging from 0.12B to 2B and datasets of OpenWebText, MiniPile, and C4. Finally, we incorporate Blockwise LR into Adam-mini (Zhang et al., 2024), a recently proposed memory-efficient variant of Adam, achieving a combined $2\times$ speedup and $2\times$ memory saving. These results underscore the potential of exploiting the sharpness disparity to improve LLM training.
We present a new method for generating confidence sets within the split conformal prediction framework. Our method performs a trainable transformation of any given conformity score to improve conditional coverage while ensuring exact marginal coverage. The transformation is based on an estimate of the conditional quantile of conformity scores. The resulting method is particularly beneficial for constructing adaptive confidence sets in multi-output problems where standard conformal quantile regression approaches have limited applicability. We develop a theoretical bound that captures the influence of the accuracy of the quantile estimate on the approximate conditional validity, unlike classical bounds for conformal prediction methods that only offer marginal coverage. We experimentally show that our method is highly adaptive to the local data structure and outperforms existing methods in terms of conditional coverage, improving the reliability of statistical inference in various applications.
A seller running an assortment experiment must earn revenue while collecting data to compare future offers. For a regular finite-dimensional parametric choice model and a finite assortment family, we characterize the oracle frontier between expected revenue loss and the largest local first-order variance constant across prespecified revenue comparisons. The oracle benchmark remains finite at zero cost exactly when revenue-maximizing assortments provide information in every target-gradient direction. When costly offers can supply the missing information, we derive the exact leading coefficient of the inverse-budget variance law and the limiting optimal spending shares. A pilot-based procedure reserves reference offers to preserve information and tightens the estimated budget constraint to account for uncertainty. We give conditions for the implemented design's variance criterion to converge to the stabilized oracle value at a fixed budget and reference share. Simulations with synthetic instances and Expedia-calibrated markets show that allocating offers across the full assortment family can improve the precision of revenue comparisons under a common opportunity-cost budget. Additional experiments assess pilot spending, budget protection, and sensitivity to choice-model misspecification.