Skip to content

📐 Learning Theory

🧠 NeurIPS2026 · 202 paper notes

📌 Same area in other venues: 🎞️ ECCV2026 (1) · 🔬 ICLR2026 (293) · 🧪 ICML2026 (45) · 🤖 AAAI2026 (3) · 🧠 NeurIPS2025 (25) · 🧪 ICML2025 (16)

🔥 Top topics: Diffusion Models ×6 · Adversarial Robustness ×5 · Compression ×5 · Layout & Composition ×3 · LLM ×2

Sponsored Questions and How to Auction Them

This paper introduces a formal model for auctioning sponsored clarifying questions and subsequent ad slots in conversational AI platforms, proving that an end-to-end VCG mechanism is dominant-strategy truthful and socially efficient, whereas standard modular VCG-per-stage auctions suffer from an unbounded Price of Anarchy.

A Local Geometric Analysis of Maximal Coding Rate Reduction via Error Bounds

This paper proves that the non-convex maximal coding rate reduction (MCR²) objective satisfies a local error-bound property around its local maximizers, guaranteeing that the distance to the non-isolated maximizer set is linearly controlled by the gradient norm, and rigorously establishing local linear convergence for first-order optimization methods.

A Mathematical Theory of Balancing Relational Generalization and Memorization

This paper introduces the transitive inference with exceptions task paradigm, analytically characterizes how kernel ridge regression balances rule generalization and exception memorization across representational geometries, and validates predicted systematic errors in fine-tuned large language models.

A multi-scale information geometry reveals the structure of mutual information in neural populations

By formalizing representational distance contraction under isotropic Gaussian coarse-graining from first principles, this paper derives a unique multi-scale Fisher information metric whose trace rigorously equals population mutual information, and introduces a conditional diffusion estimator that extracts robust, biologically grounded representational geometries from primate visual cortex.

A Theoretical Analysis of Why Masked Diffusion Models Mitigate the Reversal Curse

This paper theoretically establishes that masked diffusion models (MDMs) mitigate the reversal curse through a parameter-level coupling between position-invariant evidence storage and attention-based routing under relative positional encodings, where forward masked gradient steps monotonically decrease the reverse query loss to first order.

A Topological Sorting Criterion for Random Causal Directed Acyclic Graphs

Reveals that popular random DAG generators (Erdős-Rényi and Scale-Free) naturally encode monotonically increasing numbers of "relatives" along the causal order, a topological artifact that leaks causal ordering, collapses Markov equivalence classes (MECs) into singletons, and can be systematically eliminated by time-unrolled DAG benchmarks.

Adaptive inference for functionals of M-estimands

This paper extends inference for smooth functionals of nonparametric M-estimands to adaptive data with known sampling policies, combining one-step bias correction with per-round conditional variance stabilization, or self-normalization when variance converges to a random nonzero limit, to construct asymptotically valid confidence intervals; dynamic pricing simulations show undercoverage for ordinary GLMs and GLMs using inverse propensity weighting alone.

Aligning Inductive Bias for Data-Efficient Generalization in State Space Models

By formalizing linear time-invariant State Space Models (SSMs) as kernel feature maps, this paper proves that their inductive bias is governed by the frequency response and introduces Task-Dependent Initialization (TDI) to align initial filter spectra with task statistics, dramatically improving data-efficient generalization under constrained data budgets.

Ambient Discrete Diffusion: Using the Wrong Data at the Right Time for Data Efficient Learning

Under severe data scarcity, RefineMix introduces out-of-distribution (OOD) data selectively at low diffusion noise levels while keeping high noise levels strictly in-domain, exploiting the fundamental insight that discrete token masking preserves domain separability to eliminate sampling bias and nearly double functional protein yield from only 197 examples.

Amortized Vine Copulas for High-Dimensional Density and Information Estimation

Vine Denoising Copula (VDC) amortizes bivariate edge fitting across high-dimensional regular vines using a single pre-trained U-Net edge denoiser followed by fast Sinkhorn/IPFP marginal projection, turning costly \(O(d^2)\) pair-copula optimizations into millisecond GPU forward passes while preserving exact vine likelihoods and edge-wise mutual information decomposition.

Anchor PCA

Anchor PCA introduces a multi-domain dimensionality reduction framework that balances pooled explained variance against cross-domain subspace agreement, establishing exact recovery guarantees for the maximal invariant subspace and achieving minimax-optimal reconstruction robustness under bounded covariance inflations.

Announced Breaks Separate Conformal Reliability from Frequency Calibration

This paper demonstrates that when data streams encounter publicly announced structural breaks, per-time conformal reliability and window-averaged frequency calibration inherently decouple, and proves that post-boundary empirical quantile resets and state-resetting adaptive recursions respectively achieve their optimal convergence rates under the two audit regimes.

Asymptotically Log-Optimal Bayes-Assisted Confidence Sequences for Bounded Means

This paper introduces a Bayes-assisted framework for constructing bounded-mean confidence sequences, adaptively selecting martingale betting stakes via Bayesian or pseudo-Bayesian predictive distributions to achieve asymptotic log-optimality while guaranteeing non-asymptotic anytime-valid coverage under arbitrary prior misspecification.

Asymptotically Optimal Best Arm Identification with Fixed-Budget under Differential Privacy

Addressing pure \(\epsilon\)-differentially private best-arm identification under a fixed budget where the entire interactive transcript is protected, this paper establishes the optimal error exponent governed by an asymmetric alternative-to-true transportation divergence and proposes AO-Pri-BAI, an adaptive algorithm combining Laplace-tree continual release and AdaHedge min-max game optimization to achieve matching asymptotic optimality.

Attention as In-Context Empirical Bayes: A Two-Stage View via Particle Dynamics

This paper formulates multilayer attention-only transformers under all-token corruption as a two-stage in-context empirical Bayes procedure, proving that depth refines a non-parametric prior via particle dynamics mimicking reverse diffusion, while long-range input residuals perform posterior mean inference.

Attention Sinks Induce Gradient Sinks: Massive Activations as Gradient Regulators in Transformers

This paper investigates the interplay between attention sinks and massive activations from the perspective of backpropagation, showing that causal attention sinks induce localized gradient concentrations (gradient sinks) on the initial token, and that massive activations emerge adaptively via RMSNorm's inverse-norm Jacobian scaling to regulate this gradient pressure.

Augmented Lagrangian Predictive Coding

By framing predictive coding within the augmented Lagrangian method of multipliers and accumulating prediction errors into layer-local dual variables, PC-ALM enables ballistic credit propagation that matches backpropagation across all depth-width regimes under a matched inference budget.

Autoregressive Learning in Joint KL: Sharp Oracle Bounds and Lower Bounds

This paper establishes the first sharp, matching upper and lower bounds for misspecified autoregressive learning under the natural joint KL divergence objective, proving that the approximation error admits a horizon-free \(\mathcal{O}(1)\) factor while the \(\Omega(H)\) linear horizon dependence in estimation error is an unavoidable information-theoretic barrier.

Bayes-Optimal BER and AUC: Estimation and Evaluation of Estimators

This paper introduces soft-label-based estimators for Bayes-optimal balanced error rate (BER) and area under the ROC curve (AUC), applies isotonic regression to eliminate unknown order-preserving corruption and additive noise, and extends an unsupervised noise-injection framework to evaluate estimators without ground-truth optima.

Bayesian Additive Distribution Regression

This paper introduces DistBART, a Bayesian nonparametric distribution regression framework that places a Bayesian Additive Regression Trees prior on the Riesz representer of a linear functional, leveraging the inductive bias of shallow tree ensembles toward sparse low-dimensional marginals and enabling scalable inference via random-feature approximations with near-minimax posterior contraction guarantees.

Bentkus-Type Asymptotic E-Values

To overcome the conservatism and the "missing factor" of existing exponential asymptotic e-values under unknown variance bounds, this paper introduces a family of Bentkus-type asymptotic e-values derived from near-optimal concentration inequalities, theoretically closing the gap to Gaussian quantiles and substantially improving power in post-hoc inference and e-BH multiple testing.

Beyond Augmented-Action Surrogates for Multi-Expert Learning-to-Defer

Addressing optimization pathologies such as target distortion, gradient amplification, and expert starvation in augmented-action multi-expert learning-to-defer (L2D), this paper departs from the augmented-action paradigm and proposes a decoupled surrogate that pairs a softmax classifier head with independent sigmoid expert heads, achieving coordinatewise updates and an excess-risk consistency bound invariant to expert pool size.

Beyond Kemeny Medians: Consensus Ranking Distributions - Definition, Properties and Statistical Learning

Going beyond crude single Kemeny medians for multimodal ranking data, this paper introduces the Consensus Ranking Distribution (CRD) as a sparse mixture of Dirac masses at local medians, proves distortion bounds under the Kendall \(\tau\) Wasserstein distance, and develops a top-down tree-structured partitioning algorithm COAST for efficient statistical recovery.

Beyond Pessimism: Offline Learning in KL-regularized Games

This paper breaks away from the conventional reliance on pessimistic value estimation in offline game learning by exploiting the smoothness of KL-regularized best responses and the skew-symmetry of zero-sum game operators, establishing the first pessimism-free offline learning framework that improves sample complexity from \(\widetilde{\mathcal{O}}(1/\sqrt{n})\) to a fast \(\widetilde{\mathcal{O}}(1/n)\) rate under unilateral coverage alongside an efficient self-play algorithm with last-iterate convergence guarantees.

Beyond the Half Approximation: Fair and Efficient Online Class Matching

Introduces the first threshold-based, two-phase framework for online class matching, breaking the long-standing 1/2 utilitarian social welfare efficiency ceiling under class envy-freeness, supported by a near-tight price-of-fairness upper bound.

Beyond Worst-Case Coreset Bounds for k-Clustering via Determinantal Sampling

Addressing the long-standing worst-case \(\Omega(\varepsilon^{-2})\) lower bound for \((k,z)\)-clustering coresets, this paper introduces a negative-dependence determinantal point process (DPP) sampling framework that provably achieves an improved coreset size of \(\tilde{O}_{d,z}(k^{\frac{d}{d+1}}\varepsilon^{-\frac{2d}{d+1}})\) under mild sub-Gaussian assumptions, achieving substantial size reductions on benchmark datasets.

Block Optimism for Nonstationary Bandits with Latent Linear Dynamics

For endogenous nonstationary bandits where actions drive an unobserved latent linear state and yield bilinear rewards, this paper introduces a cyclic block approximation and an optimistic block-level algorithm (Block-OFUL) that improves the open-loop regret bound from \(\widetilde{\mathcal{O}}(T^{2/3})\) to \(\widetilde{\mathcal{O}}(\sqrt{T})\).

BOOST: Power-Optimal Strong-FWER Testing for Block-Structured Multiplicity

Targeting the block-structured multiplicity prevalent in clinical trials, eQTL mapping, and bundled A/B testing, this paper presents BOOST, achieving finite-sample strong FWER control and power-optimal allocation for size-3 blocks (\(K=3\)) within the block-separable class at \(\mathcal{O}(K)\) cost.

Calibration Without Ground Truth

This paper proposes a label-free post-processing framework that improves a strong but miscalibrated model using a weaker yet better-calibrated reference (such as a pre-trained base model) via functional Bregman projection onto a reference-compatible set, theoretically guaranteeing strict performance gains under any strictly proper loss without human supervision.

Calibration Without Labels in Multiple Testing

Addressing the fundamental challenge that ground-truth labels are never revealed in large-scale multiple testing, this paper constructs asymptotically unbiased pseudo-labels from ordered \(p\)-value spacings, proves that isotonic calibration coincides with nonparametric empirical Bayes estimation, provides finite-sample Brier regret bounds without requiring true monotonicity, and enables principled calibration diagnostics for individual hypotheses.

Can Metadata Fix the Gauge? Calibration Turns Sparse Multi-Source Learning from Sparse PCA into Sparse Mean Recovery

In a sparse Gaussian multi-source model with source-specific sign gauges, relying solely on source labels collapses the experiment into sparse PCA, whereas task-independent relative-gauge calibration metadata enables spectral synchronization to linearize the problem into sparse mean recovery, establishing fundamental statistical and computational separations in the micro-source regime.

Can the Parts Fool the Test? Counterfactual Pairing Cycles for Relational OOD

The paper proposes counterfactual pairing cycles as an evaluation and training primitive that swaps image-caption pairings across a \(2\times 2\) block to algebraically cancel all additive unimodal marginal shortcuts in finite samples, revealing that frozen CLIP representations capture geometric size relations but remain close to random chance on strict directional left/right role binding.

Causal pieces: analysing and improving spiking neural networks piece by piece

Analogous to linear pieces in ReLU artificial neural networks, this paper introduces the concept of "causal pieces" for feedforward spiking neural networks with single-spike coding, proves that the approximation error is lower bounded inversely by the number of pieces, and demonstrates that the initial count of causal pieces strongly correlates with training success.

Cheap and Powerful Tests for Supervised Subspaces: Per-Component Inference for PLS

The paper turns significance testing for partial least squares (PLS) supervised subspaces into tests of held-out predictive correlation, supplies an NB-corrected fast approximation with an explicit empirical scope and a finite-sample valid permutation test under the omnibus independence null, and separates ordered inference on original components from interpretation of rotated coordinates.

Class Adaptive Conformal Training

To resolve the issues of inflated prediction sets and severe class-conditional coverage disparities in end-to-end conformal training caused by rigid global penalty weights under class imbalance and large label spaces, this paper introduces Class Adaptive Conformal Training (CaCT), which formulates class-wise set size constraints into an augmented Lagrangian optimization framework that dynamically tunes per-class dual multipliers and penalty parameters without prior distributional assumptions.

Classification Fields: Arbitrarily Fine Recursive Hierarchical Clustering From Few Examples

This paper introduces classification fields, which model hierarchical clustering as an infinite-depth geometric object recursively generated by a local parent-to-child refinement rule, enabling finite neural predictors to roll out arbitrarily deep cluster hierarchies while preserving genealogy, geometry, and path metrics.

Closing the Approximation Gap in Simulation-Free Latent SDEs

Addressing the approximation truncation gap caused by the non-uniqueness of path laws from one-time marginals in simulation-free latent SDEs, this paper proposes Helmholtz-SDE, which introduces an analytical divergence-free correction projected onto an orthogonal Hermite polynomial basis to close the posterior KL gap while maintaining high simulation-free efficiency.

Clustered Randomized Smoothing for Stochastic Prediction Functions

Addressing mode collapse in conventional randomized smoothing for multimodal regression in stochastic neural networks, this paper proposes clustered \(\alpha\)-smoothing—a framework that clusters outputs, performs localized trimmed smoothing per cluster, and recombines predictions into a mixture distribution with provable probabilistic certificates.

Coherence Mechanisms for Provable Self-Improvement

This paper introduces a geometric framework for provable self-improvement via output coherence under task-preserving transformations, establishing that Bregman projections onto coherent model sets guarantee monotonic divergence reduction toward an ideal model and proving that any mechanism offering universal improvement guarantees must rigidly conform to a coherence projection structure.

Complexity of Differentially Private Selection via Federated APIs

The paper proves an inherent \(\Omega(\sqrt{d}/\varepsilon)\) sample complexity lower bound for differentially private selection via standard noisy aggregation APIs, and introduces a federated Sparse Vector Technique (SVT) API that achieves central-optimal \(O(\log(dn)/\varepsilon)\) additive error using a near-linear \(O(d\log^* d)\) expected number of queries.

Computationally Efficient Replicable Learning of Parities and Applications

This paper introduces the first computationally efficient, replicable algorithm for learning parity functions under arbitrary distributions, breaking the barrier that restricted efficient replicable learning to SQ-learnable problems, while establishing that converting replicability into pure differential privacy fundamentally incurs a sample complexity loss.

Computationally Sufficient Statistics for Ising Models

This paper investigates learning discrete Gibbs distributions under restricted observational power, proving that local polynomial moment statistics up to order \(O(\gamma)\) are computationally sufficient to learn the structure and parameters of Ising models with \(\ell_1\)-width bounded by \(\gamma\) in polynomial time and near-optimal sample complexity.

Computing All Optimal Partial p-Wasserstein Matchings on the Line

To break the quadratic barrier in computing one-dimensional partial \(p\)-Wasserstein optimal transport profiles for \(p>1\), this paper introduces an FFT-accelerated balanced-interval matching data structure based on monotone alignment, achieving the first \(O(p\,n\log^2 n)\) near-linear time algorithm and proving a conditional quadratic lower bound for \(p=\infty\) under the \((\min,+)\)-convolution hypothesis.

Concise and Logically Consistent Conformal Sets for Neuro-Symbolic Concept-Based Models

COCOCO separately calibrates the concepts and task labels of neuro-symbolic concept-based models, then removes unsupported candidates through one deduction–abduction intersection step, producing smaller, mutually consistent sets with explicit coverage-loss bounds and supporting adaptive size budgets through COCOCO*.

Consistent Geometric Deep Learning via Hilbert Bundles and Cellular Sheaves

This paper extends geometric deep learning from scalar and tangent fields to infinite-dimensional signals supported on Riemannian manifolds by proposing HilbNets—a convolutional framework based on Hilbert bundle connection Laplacians and the Borel functional calculus—and proves through a two-stage sampling procedure that the induced Hilbert cellular sheaf Laplacians converge in probability to continuous connection Laplacians with guaranteed cross-sampling transferability.

Constraint-Aware Influence Estimation in Deep Constrained Learning

To resolve the fundamental failure of classical influence functions in deep constrained learning caused by feasibility violation and directional non-differentiability, this paper proposes the Directional Influence Function (DIF), which formulates optimality as a variational inequality, derives first-order sensitivity via critical cone geometry, and solves an equivalent quadratic program for constraint-preserving attribution.

Continuity Laws for Sequential Models

This paper establishes temporal continuity as a crucial inductive bias for sequential models, formalizes it via convergence under temporal refinement alongside a data-driven task continuity metric, reveals that S4 exhibits stable continuity while S6 suffers from numerical discontinuity under large amplitudes, and develops a stage-wise temporal subsampling training strategy that accelerates training and improves final accuracy.

Contrastive Identification and Generation in the Limit

This paper extends Gold's classical identification in the limit and Kleinberg–Mullainathan's generation in the limit to undirected XOR contrastive pair streams, establishes an exact geometric characterization and a combinatorial closure dimension via common crossing graphs, reveals a strict diamond hierarchy, and demonstrates a sharp robustness reversal under adversarial corruption where contrastive data outperform positive examples.

Conveyance: A Versatile Framework for Learning in Structured Class Spaces

Conveyance tackles the fundamental limitation of structure-agnostic cross-entropy by introducing a lightweight loss function that encodes class topological relationships via a simple boolean plausibility matrix, optimizing dual margins in logit space at standard cross-entropy complexity to match or outperform specialized methods across multiple structured classification paradigms.

Cost-Aware Best-LLM Identification using Dueling Feedback

The paper formulates fixed-confidence identification of the best model under pairwise preferences as a heterogeneous-cost dueling bandit, allocates comparisons using rejection evidence per unit cost, tracks sampling targets, and stops through best-arm hypothesis testing, with an almost-sure asymptotic cost guarantee; however, the confidence-interval variant DCTAC costs less than the main algorithm DCTAS in the finite experiments.

DASS: A Solver-Adaptable Dynamic Auxiliary Search Strategy for Symbolic Regression

DASS introduces a solver-adaptable dynamic auxiliary search strategy that models the symbolic regression solution space as a functional subspace spanned by high-potential basis function terms, calibrating term potentials via multi-environment evaluation, stability filtering, and Gibbs posterior updates while reshaping the basis library through solver feedback, doubling downstream accuracy with only a 24.14% runtime increase.

Diffusion Model's Generalization Can Be Characterized by Inductive Biases toward a Data-Dependent Ridge Manifold

By constructing a time-dependent family of log-density ridge manifolds from the smoothed empirical training distribution, this paper proves that non-memorizing reverse-time inference in diffusion models follows a three-stage "reach–align–slide" geometric mechanism, and quantitatively links normal and tangential training errors (decomposed into finite-width architectural bias and finite-step gradient descent optimization bias) to the spatial geometry of generated samples.

Dimension-Uniform Discretization Analysis of Preconditioned Annealed Langevin Dynamics for Multimodal Gaussian Mixtures

For sampling high- and infinite-dimensional multimodal Gaussian mixtures, this paper proves that Euler–Maruyama discretization suffers from severe high-frequency step-size stability constraints caused by forward-Euler integration of stiff linear drifts, and establishes a dimension-uniform Kullback–Leibler convergence bound for an exact-linear-part (ELP) exponential-integrator scheme.

Disentangling Continuous-Time Latent Dynamics: Identifiability of Latent SDEs via Diffusion Shifts

This paper proves that under unknown nonlinear diffeomorphic observations, two continuous-time additive-noise latent stochastic differential equation (SDE) environments with shared drift and diagonal diffusion covariances exhibiting coordinate-wise distinct variance ratios identify the latent coordinates up to a monomial affine transformation (permutation, coordinate-wise scaling, and constant shift), enabling exact recovery of the instantaneous drift-Jacobian causal graph via a two-stage estimator without drift sparsity assumptions.

Distribution Shift in Missing Data Imputation: A Risk-Based Perspective and Importance-Weighted Correction under MAR

Starting from an MSE risk minimization perspective, this paper proves that a natural covariate shift exists between observed training data and full evaluation data under Missing at Random (MAR), and proposes a weighted iterative imputation algorithm combining density ratio estimation and variance tempering to consistently improve imputation accuracy and distributional fidelity.

Distribution Shift in Missing Data Imputation: A Risk-Based Perspective and Importance-Weighted Correction under MAR

This paper rigorously formulates missing data imputation as a mean squared error (MSE) risk minimization problem, proves that Missing At Random (MAR) mechanisms induce a systematic covariate shift between observed training data and the target full data distribution, and proposes an importance-weighted round-robin imputation algorithm that significantly reduces imputation error and improves marginal Wasserstein distributional fidelity.

Distributionally-Robust Policy Learning with Observational Data

Addressing test-time covariate distribution shifts in policy learning from observational data, this paper proposes a distributionally robust estimator using cross-fitted doubly robust scores and dual sample-splitting bias correction, proving the first fast rates under Tsybakov-style wide-margin conditions that interpolate between \(\widetilde{O}(n^{-1/2})\) and \(\widetilde{O}(n^{-1})\).

Do Deep Networks Forget Initialization? A Forgetting-Time View of Practical Inductive Bias

This paper introduces and measures "initialization memory" (the dependence of validation-selected predictors on random initialization scale), demonstrating that low-learning-rate SGD interpolates while preserving strong initial bias, whereas adaptive methods and explicit/implicit regularizers rapidly erase it across characteristic dynamical clocks.

Do Heavy Tails Help Diffusion? On the Subtle Trade-off Between Initialization and Training

Revisiting the recent paradigm of incorporating heavy-tailed noise into diffusion models, this paper establishes non-asymptotic total variation error bounds proving that while heavy-tailed noise yields exponentially faster prior mixing, it causes statistical training errors to accumulate linearly along the reverse trajectory, leading standard Gaussian light-tailed diffusion to consistently achieve superior sample fidelity and tail coverage under finite samples.

Do More Modalities Always Help? A Geometric Perspective on Missing-Modality Robustness

Reveals the counterintuitive deployment-time degradation where full-modality models underperform unimodal baselines under missing-modality evaluation, attributes it to pathological rotations of principal input subspaces in cross-modal interaction layers, and proposes Geodesic Unlearning (GU) for lightweight, provably optimal post-training subspace correction.

Does Compression Imply Generalization? A Minimum Description Length Perspective

This paper approaches neural network generalization through the minimum description length (MDL) principle, proving that prequential regret computable entirely from the training trajectory upper-bounds parameter description length in the nearly-interpolating regime and yields rigorous PAC-MDL generalization bounds with Pearson correlations exceeding 0.9 on CIFAR.

does sparse connectivity improve generalization convolutional networks below the

Under the edge-of-stability (BEoS) regime of gradient descent, this paper proves that sparse connectivity shifts the Hessian curvature constraint from the high-dimensional ambient space to the low-dimensional patch multiset, yielding non-vacuous generalization bounds on spherical data where fully connected networks provably fail.

Does Weight Decay Enhance Training Stability?

By formalizing gradient descent with weight decay as an underdamped harmonic oscillator, this paper demonstrates that weight decay dampens curvature oscillations at the Edge of Stability (EoS) and triggers a limit-cycle collapse phase transition driven by global parameter alignment, thereby bounding the empirical NTK spectrum and stabilizing function-space training.

Does Your Neural Network Extrapolate? Feature Engineering as Identifiability Bias for OOD Generalization

Starting from data-generating process (DGP) identifiability theory, this paper proves that Empirical Risk Minimization (ERM) on a single training window cannot distinguish observationally equivalent candidate processes, demonstrating that the joint selection of feature map, label map, and model class \((\varphi,\psi,\mathcal{M})\) serves as an essential structural commitment that turns divergent extrapolation into manifold interpolation.

Drifting Fields are not Conservative

This paper demonstrates that the vector-valued drift fields in drifting generative models are generally non-conservative due to an order mismatch in local density normalization, and shows that substituting the sharp kernel restores full conservatism as a score-difference flow optimized via a simple scalar log-KDE loss.

DTA-GT: Direction- and Topology-Aware Graph Transformer for Neural Network Representation Learning

This paper proposes DTA-GT, a graph Transformer built on encoder-wide directional consistency that preserves directed computational semantics across node initialization, Magnetic Laplacian spectral encoding, direction-dual structural attention, and a direction-sensitive MoE-FFN, consistently outperforming representative GNN and Transformer predictors in low-budget NAS accuracy and latency estimation.

Early Prediction of Future Behavioral Strategy from Process Traces

Addressing the challenge where adaptive systems struggle to anticipate an individual's behavioral strategy in a future task from limited early evidence, this paper proposes the Process-Level Latent Variable Model (PLVM), which encodes partial process traces from multiple source tasks via self-supervised causal transformers and fuses them into a shared person-level latent representation, successfully predicting future navigation styles (Zone Planner vs. Zone Hopper) in held-out target tasks across naturalistic gameplay and controlled cognitive simulations.

Efficient Test-time Adaptation through Candidate Verification and Divergence Shifts

Addressing target-domain distribution shifts in vision-language models during inference, this paper proposes Test-Time Correction (TTC), a training-free framework that reframes test-time adaptation from traditional prediction-side adjustments into a hypothesis-driven candidate verification process using reconstruction-induced divergence shifts over candidate memory subspaces.

EM-NeSy: Expectation Maximization for Neurosymbolic Learning

EM-NeSy casts probabilistic neurosymbolic learning as an Expectation-Maximization framework, using arbitrary exact or approximate symbolic inference engines in the E-step to infer posteriors over neural latent variables, and updating neural parameters solely via gradient descent in the M-step—removing symbolic differentiability requirements while mathematically recovering standard end-to-end gradients under exact inference.

Embedding Foundation Model Predictions in Discrete-Choice Models with Structural Guarantees

To fix behavioral validity failures in tabular foundation models applied to discrete choice (such as cost-monotonicity violations and negative values of time), this paper proposes a two-stage adapter that embeds frozen foundation model predictions into a multinomial logit utility, mathematically guaranteeing the preservation of the structural marginal rate of substitution while recovering most of the accuracy advantage.

End-to-End Identifiable and Consistent Recurrent Switching Dynamical Systems

This paper establishes the first end-to-end identifiability theory for nonlinear recurrent switching dynamical systems under general invertible emissions and introduces \(\Omega\text{SDS}\), a flow-based estimator optimizing exact marginal likelihood via expectation-maximisation for consistent and scalable latent dynamics recovery.

Energy-Tweedie: Score meets Score, Energy meets Energy

This paper generalizes the classical Tweedie formula beyond additive Gaussian noise and posterior means to arbitrary normalizable Gibbs (energy-based) noise distributions, proving that the Stein score of the noisy marginal equals the path derivative of a matched kernel scoring rule (such as the Mahalanobis energy score) evaluated at the full denoising posterior, thereby enabling sample-based score estimation, noise-parameter M-estimation, and diffusion sampling along arbitrary noise-parameter paths.

Enhancing Tabular Learners with Context-Aware Semantic Embeddings

To overcome the semantic blindness of tabular learners and the isolation of naive text embeddings, CASE pre-fills the KV cache of a custom-trained Tabular Language Model with representative context rows to establish a persistent semantic anchor, extracting target-aligned row embeddings that significantly boost downstream tabular predictors on semantically rich datasets.

Epistemic Pairwise Maximin Share

This paper introduces Epistemic Pairwise Maximin Share (EPMMS) and Epistemic Groupwise Maximin Share (EGMMS), establishes a polynomial-time 4/5-EPMMS allocation for additive valuations, proves the existence of exact EGMMS allocations for bivalued valuations, and shows that exact EPMMS allocations always exist for instances with three agents or two agent types.

Estimating Implicit Regularization in Deep Learning

Presents an empirical gradient matching framework that treats observed parameter updates or endpoints as solutions to regularized risk minimization, allowing researchers to directly estimate parameterized models of implicit regularization across both toy models and complex deep architectures without analytical derivations.

Estimating the Expected Output of Wide Random MLPs More Efficiently than Sampling

This paper introduces a sample-free estimation framework based on cumulant propagation and Hermite expansions to estimate the expected output of wide random multilayer perceptrons over Gaussian inputs, theoretically achieving an asymptotic complexity of \(O(n/\varepsilon^2)\) compared to Monte Carlo sampling's \(\Theta(n^2/\varepsilon^2)\) and empirically outperforming sampling by orders of magnitude on rare event probability estimation.

Estimation of the Label-Noise Transition Matrix with Performance Guarantees via Selective Classification

The paper learns high-purity subsets through one-sided selective classification and estimates the transition matrix from the empirical frequencies of all noisy labels within those subsets, decomposing error into subset contamination, learning suboptimality, and accepted-sample fluctuations without requiring accurate pointwise class-posterior recovery.

Even More Guarantees for Variational Inference in the Presence of Symmetries

This paper extends theoretical guarantees for location-scale variational inference under target symmetries to forward KL and \(\alpha\)-divergences, eliminating the restrictive log-concavity assumption and proving exact recovery of mean and correlation matrices for multi-modal and heavy-tailed targets.

Even Sharper Bounds for Transductive Learning and Its Applications

The paper proves Bernstein-type supremum concentration through a two-parameter entropy closure for the swap walk, then uses local complexity to analyze transductive empirical risk minimization, removing both earlier sample-imbalance restrictions and an extra logarithmic confidence factor under bounded-loss and appropriate localization conditions, with applications to realizable VC classification and the empirical kernel spectrum.

Exact Gaussian Moment Matching for Residual Networks: A Second-Order Method

Addressing the long-standing intractability of closed-form distribution propagation under arbitrary multivariate Gaussian inputs, this paper derives exact layerwise first- and second-order moment matching for probit, GeLU, ReLU, Heaviside, and sine activations as well as generalized residual layers using Stein's lemma and Gaussian differential identities, establishing provable asymptotic second-order accuracy under a smooth fourth-order probability metric.

Exact Power Indices for Plurality-Voting Ensembles

Addressing the exponential attribution challenge in plurality-voting ensembles with uniform random tie-breaking, this paper introduces enpower, leveraging combinatorial symmetries and ordinary generating functions to compute exact Banzhaf and Shapley power indices in linear time for binary classification and low-degree polynomial time for multiclass tasks.

Exact Topological Compliance: Generating Persistence-Equivalent Graphs

Initiates a first-principles study of strictly persistence-diagram-compliant graph generation, proving an impossibility theorem for local iterative generation and introducing a minimal-sufficient global constraint programming encoding that achieves 100% exact topological compliance across benchmarks.

Expanding Flow Maps

This paper introduces Expanding Generative Flows (EFlows) and Expanding Flow Maps (EFMs), factoring flow maps into a conditional noise expansion operator and a transport map to unify continuous and discrete few-step generation with adaptive dimensionality.

Expressive Power of Deep Homomorphism Networks over Relational Databases

The paper characterizes the expressive power of Deep Homomorphism Networks (DHNs) over relational databases by relating max, sum, and mean aggregation to logical formalisms (UNFO, UNFOC, UQAFO, HML+C, and RHML), settling static analysis decidability, and empirically validating strict expressive separations.

Extending Myerson's Optimal Auctions to Correlated Bidders via Neural Network Integration

Addressing the computational intractability and non-monotonicity of extending Myerson's optimal auctions to correlated bidders, this paper proposes an empirical pipeline combining tiny ReLU-MLP neural interpolation of marginal profits, MIP-based exact monotonicity verification, and constructive revenue fixes, closing the revenue gap to the theoretical greedy upper bound to within 0.26% while guaranteeing exact strategy-proofness.

Feature Learning Dynamics in Infinite-Depth Neural Networks

This paper investigates reused-weight forward-backward coupling in one-layer ResNets under depth-\(\mu\mathrm{P}\) scaling via conditional Gaussian representations, proving that while training-induced coupling persists in the infinite-width limit, it decays as a higher-order \(\mathcal{O}(L^{-2})\) depth effect, thereby rigorously establishing the Neural Feature Dynamics (NFD) forward-backward SDE system.

Finding Koopman Invariant Subspaces via Personalized PageRank

Addressing subspace truncation and leakage in high-dimensional Extended Dynamic Mode Decomposition (EDMD), this paper proves that Koopman invariant subspaces inherently induce exact zero-block structures in finite-sample EDMD matrices, and presents a graph-theoretic approach that recasts row-normalized EDMD matrices as Markov chains to detect invariant subspaces and bound multi-step leakage via Personalized PageRank (PPR).

Fisher information and the geometry of memorization in neural networks

This paper reveals a loss-curvature geometric decoupling between generalization structure and sample memorization: projecting layer weights onto the top-\(k\) eigenspaces of the Fisher Information Matrix (FIM) selectively erases unstructured sample memorization with negligible loss in generalization, partially recovers true semantics corrupted by label noise, and restores task generalization in overfitted in-context regression transformers.

Fixed Universal Transformers

Analogous to a universal Turing machine, this paper shows that a fixed-weight transformer (either an explicit sparse construction or a randomly initialized backbone) can exactly simulate any target transformer within a given architectural family purely via a target-specific input embedding, while establishing an unavoidable exponential lower bound \(H^{\Omega(L)}\) on the embedding dimension under attention matching.

Flexformer: Flexible Linear Transformer with Learnable Attention Kernel

Flexformer builds on spectral representation theory by parameterizing random Fourier feature frequencies and scaling factors into end-to-end learnable parameters, constructing stationary and nonstationary attention kernels that strictly subsume the softmax kernel while maintaining linear \(O(N)\) time and space complexity.

Flexible Routing via Uncertainty Decomposition

This paper introduces an uncertainty decomposition routing framework based on higher-order calibration that separates total loss into aleatoric and epistemic components, enabling zero-shot tripartite decision-making (predict, route, abstain) across arbitrary loss functions and cost regimes without retraining.

Flow-based Spectral Kernel Learning for Nonstationary Attention

Flow-based Spectral Kernel Attention (FSKA) formulates expressive nonstationary attention via normalizing flow spectral density estimation with positive random features for linear complexity, while leveraging kernel expressiveness to remove query–key projections and fix orthogonal value projections, reducing projection parameters by 40%–47%.

Forest-Guided Semantic Transport for Label-Supervised Manifold Alignment

FoSTA learns supervised forest geometries independently within each domain using extended semi-supervised RF-GAP affinities, leverages class-semantic profiles to infer cross-domain correspondences via scalable hierarchical optimal transport without predefined anchors, and constructs a unified low-dimensional embedding that preserves both semantic structure and continuous biological variation.

Fréchet Regression on the Bures-Wasserstein Manifold

To tackle non-convexity, potential non-existence of minimizers, and boundary collapse caused by unavoidable negative weights during extrapolation in Fréchet regression over the symmetric positive definite (SPD) Bures-Wasserstein manifold, this paper establishes a verifiable Spectral Dominance of Positive Weights condition ensuring well-posedness and a landscape free of local maxima, alongside provably correct projection-free Riemannian gradient descent and pairwise stochastic Riemannian optimization algorithms.

From Average Sensitivity to Small-Loss Regret Bounds under Random-Order Model

Addressing online learning under the random-order model, this paper introduces a time-varying approximation accuracy mechanism into the batch-to-online transformation framework, deriving general small-loss regret bounds governed by the concave conjugate of the offline algorithm's average sensitivity without requiring convexity or smoothness assumptions, and establishing breakthrough polynomial-time guarantees for online submodular minimization.

From Collapse to Improvement: Statistical Perspectives on the Evolutionary Dynamics of Iterative Training on Contaminated Sources

From a rigorous statistical estimation perspective, this paper derives the exact risk recurrence of generative models iteratively trained on mixtures of real and synthetic data, proving that contamination-agnostic empirical risk minimization can avoid model collapse and achieve statistical consistency as long as fresh real data is continuously injected with appropriate sample size scaling.

From Matrix Inversion to Constraints: Provably Tighter Confidence Regions for Importance Weights in Label Shift

Addressing excessive conservatism in importance weight estimation under unsupervised label shift, this paper proposes MaC-LP, reformulating interval Gaussian elimination into a direct matrix constraint system to derive provably tight finite-sample confidence intervals via linear programming that significantly reduce downstream conformal prediction set sizes.

From Persistence to Survival: Hypothesis Testing, Effect Sizes and Vectorisation for Topological Features

STRAND reframes topological persistence diagrams as survival time-to-event data, leveraging the persistence survival function to unify non-parametric two-sample hypothesis testing (log-rank test and hazard ratios) with 1-Wasserstein stable compact vectorisation under a single mathematical representation.

Function graph transformers universally approximate operators between function spaces

This paper lifts functions to probability measures supported on their graphs (graph measures) and introduces graph-preserving Function Graph Transformers, proving that finite compositions of standard softmax self-attention layers and pointwise MLPs universally approximate nonlinear operators between continuous function spaces as well as negative-order Sobolev distribution spaces, while unifying cross-domain arbitrary query decoding via product-measure self-attention.

Gaussian Mixture Models in Hilbert Spaces via Kernel Methods

This paper develops a density-free Gaussian mixture modeling framework for infinite-dimensional and dynamic random objects in separable Hilbert spaces by minimizing the Maximum Mean Discrepancy (MMD) via spectral projections, establishing closed-form objectives, Wasserstein approximation guarantees, and state-of-the-art clustering performance on complex biomedical and geometric datasets.

Generalization in Neural Networks Through the Lens of Magnitude Potential

This paper introduces the Magnitude Potential Ratio (MP ratio), a geometric metric grounded in metric magnitude theory that computes the coverage of a data point within its class versus the entire dataset in logit space via an exact closed-form leave-one-out formulation, demonstrating strong negative correlation with sample memorization and serving as an early indicator of grokking.

Generating from Discrete Distributions Using Diffusions: Insights from Random Constraint Satisfaction Problems

Grounding discrete data generation in random CSP theory and statistical physics phase transitions, this paper proves and empirically demonstrates that continuous Gaussian diffusions systematically outperform masked discrete diffusions on discrete solution generation, shows that learned neural denoisers can match ideal belief propagation limits, and establishes that topological variable ordering pushes sampling capacity to optimal dynamical phase transitions.

Generative Conformal Prediction with Optimized Coverage Allocation

Addressing the excessive conservatism of conventional conformal prediction under multimodal or heterogeneous distributions, this paper proposes ORCA, a three-stage generative framework that formulates rank-dependent coverage allocation as a mixed-integer linear program to produce efficient prediction sets with finite-sample validity and asymptotic convergence to the oracle highest-density region.

Geometric Analysis of Neural Regression Collapse via Intrinsic Dimension

This paper provides the first systematic study of neural multivariate regression collapse through the lens of intrinsic dimension, establishing the over-compressed and under-compressed regimes, and proposing the intrinsic dimension gap as a surrogate for policy evaluation without costly rollouts.

Geometry-Aware Optimal Transport: Fast Intrinsic Dimension and Wasserstein Distance Estimation

Addressing the fundamental discretization error bottleneck in large-scale optimal transport, this work derives a solver-free Monte Carlo estimator for sampling error and intrinsic dimension via semi-dual closed forms, followed by a coupled diagonal Richardson extrapolation scheme that significantly debiases Wasserstein distance estimation.

Geometry of Relaxed Fair Regression: A Unified Framework for Aware and Unaware Settings

This paper formulates demographic parity penalized regression under an optimal transport perspective, deriving the first continuous closed-form transport maps for relaxed fair regression in both aware and unaware settings under squared Wasserstein-2 and Total Variation penalties, and introducing a simple, highly efficient post-processing algorithm.

Global Convergence in Deep Networks via the Second Law of Thermodynamics

By deriving a Rényi-divergence change-of-measure inequality rooted in the second law of thermodynamics, this paper establishes non-asymptotic global linear convergence for regularized continuous-time Langevin dynamics in finite-width deep networks without confining weights to a local neighborhood of initialization.

Gluing Local Contexts into Global Meaning: A Sheaf-Theoretic Decomposition of Transformers

This paper frames Transformer representations over paraphrase graphs via cellular sheaves, proving that the sheaf Laplacian kernel (\(H^0\)) isolates invariant factual content while its spectral complement (\(H^1\)) concentrates causal influence and multi-layer Hodge harmonic mass predicts steering fragility across architectures without training.

Guidance for Prior Change via Density Ratio Estimation

Addressing the bias and expensive retraining issues in amortized diffusion-based Simulation-Based Inference (SBI) under evolving priors, this work proposes learning an unbiased test-time score guidance term via Density Ratio Estimation (DRE) over noisy diffusion states, eliminating both extra simulator invocations and Gaussian reverse kernel approximations.

Hierarchical Conformal Classification

This paper introduces Hierarchical Conformal Classification (HCC), a framework that integrates class taxonomies (DAGs) into the structure and semantics of conformal prediction sets, searching over non-overlapping leaf covers with dynamic LCA pruning to deliver compact, user-aligned prediction sets with rigorous finite-sample coverage guarantees.

High-Arity Sample Compression

This paper introduces high-arity sample compression schemes for product spaces across both partite and non-partite settings, proving that any non-trivial compression scheme implies high-arity PAC learnability via a coordinates-revealing Doob martingale and Azuma's inequality.

High-Dimensional Learning Dynamics of Attention-Indexed Models

For attention-indexed models with extensive-rank attention matrices, the population loss is governed by finitely many trace order parameters, while online SGD follows an infinite hierarchy of matrix moments that admits exponentially accurate finite-degree truncations; tied attention induces automatic symmetry breaking to achieve weak recovery in \(\Theta(d^2 \log d)\) samples, whereas untied attention exhibits a fast-slow timescale separation that governs its recovery transition.

High-Probability Minimax Adaptive Estimation in Besov Spaces via Online-to-Batch

Addressing the lack of high-probability error bounds and unknown noise variance adaptation in nonparametric Besov regression, this paper develops a wavelet-based online learning algorithm with adaptive gradient clipping and multiscale expert aggregation, establishing the first noise-level-aware high-probability minimax-optimal risk bounds under sub-exponential noise via refined online-to-batch conversion.

How Do Transformers Learn to Represent Symmetries?

This paper investigates how vanilla Transformers learn geometric invariances and equivariances under finite data augmentation, uncovering a hierarchy where angle-preserving symmetries are learned far more efficiently than non-angle-preserving ones, and elucidating the interpretable internal mechanisms across attention layers, residual streams, and LayerNorm.

How Long Does Infinite Width Last? Signal Propagation in Long-Range Linear Recurrences

By mapping the Wick expansion of recurrent weight powers to cycle counts in random permutation commutators, this paper derives the first exact finite-width formula for hidden state signal energies in linear recurrences, pinpointing a sharp \(t \sim c\sqrt{n}\) depth-width critical scaling threshold at which infinite-width theory breaks down and standard Glorot initialization becomes unstable.

Human-AI Teaming Through the Lens of Calibration

This paper establishes a unified theoretical framework for human-AI teaming through statistical calibration, proving that common prediction combination methods fail to preserve the human's calibration, while delegation requires a finely calibrated rejector that incurs irreducible excess risk whenever the human relies on unobserved features.

Hybrid Probabilistic Zonotopes for Identifiable and Refinable Predictive Uncertainty

Introduces the Hybrid Probabilistic Zonotope (HProbZ) output head that decouples predictive uncertainty into discrete modal, bounded drift, and irreducible Gaussian noise generators, enabling parameter identifiability via exact convolution likelihood and \(\mathcal{O}(1/k)\) single-pass observation-driven variance contraction.

Hyper Input Convex Neural Networks for Shape Constrained Learning and Optimal Transport

This paper introduces Hyper Input Convex Neural Networks (HyCNNs), integrating dual-lane affine operations and maxout gating with dedicated skip connections to ensure strict input convexity while enabling exponential-in-depth piecewise-affine refinement, substantially outperforming classical ICNNs in high-dimensional optimal transport and single-cell perturbation modeling.

Identified-Set Geometry of Distributional Model Extraction under Top-k Censored Queries

This paper formalizes top-k censored LLM API access as a partial identification problem, deriving the exact total-variation diameter \(U_K\) of the identified set and a computable minimax KL lower bound, while showing that per-position distribution fidelity sharply decouples from downstream task capability transfer.

Implicit Bias of Mirror Flow in Homogeneous Neural Networks: Sparse and Dense Feature Learning

This paper derives a conserved dual balance equation for mirror flow in homogeneous neural networks via convex duality, defines the induced Q-margin, and proves directional convergence to max-margin KKT points, unifying sparse and dense feature learning across different mirror potentials.

Improved State Mixing in Higher-Order and Block Diagonal Linear Recurrent Networks

This paper introduces Higher-order Linear Recurrent Units (H-LRU) and Block-Diagonal LRUs (BD-LRU) equipped with a row-wise \(L_1\) joint gate normalization and an associative parallel scan, breaking the expressivity bottleneck of diagonal state transitions while preserving sub-quadratic training throughput and dynamical stability.

In-Expectation Convergence of Stochastic Gradient Methods under Heavy-Tailed Noise

Addressing the long-standing theoretical dilemma where classical stochastic gradient methods are widely believed to diverge under heavy-tailed noise (\(p\)-th moment bounded for \(p \in (1,2]\)), this work introduces a novel three-step analysis framework that establishes optimal in-expectation convergence rates for SMD, ASMD, SGD, and SGDM in both convex and nonconvex settings without algorithmic modifications or domain compactness.

Inference and Uncertainty Quantification for Streaming r-PCA

This paper resolves two fundamental open problems in streaming \(r\)-PCA by establishing sharp operator-norm convergence for general rank under sub-Gaussian data without non-vanishing remainder terms, deriving high-dimensional Gaussian approximations for subspace estimation errors via tangent linearization, and proposing a consistent online multiplier bootstrap for practical inference.

Inference and Uncertainty Quantification for Streaming r-PCA

This paper resolves two fundamental open problems in streaming \(r\)-PCA by establishing sharp operator-norm convergence for general rank under sub-Gaussian data without non-vanishing remainder terms, deriving high-dimensional Gaussian approximations for subspace estimation errors via tangent linearization, and proposing a consistent online multiplier bootstrap for practical inference.

InfoFlow: A Framework for Multi-Layer Transformer Analysis

This paper establishes a fundamental depth separation between single-layer and multi-layer Transformers on retrieval tasks, proves that softmax attention can only efficiently retrieve the maximum-scoring token, and introduces InfoFlow—a coarse-grained framework tracking discrete accessible information sets with rigorous parameter complexity laws.

Information from coincidences

This paper establishes a single non-asymptotic algebraic "mixed coincidence identity" that simultaneously identifies the log-mixed count \(\log \mathbb{E}_{\nu}[\prod_{i=1}^W \pi_i^{\alpha_i}(x)]\) as a Boltzmann coincidence weight, an exponential-family normalizer, a maximum-entropy value, and a KL-barycenter optimum, unifying concentration inequalities, hypothesis testing, and PAC-Bayes bounds with empirical validation on large language model vocabularies and genomic sequences.

Instance-Adaptive Online Multicalibration

Addressing the limitation where standard online multicalibration algorithms are confined to the worst-case \(\widetilde{\mathcal{O}}(T^{2/3})\) minimax lower bound, this paper proposes an efficient unified algorithm based on adaptive dyadic grid refinement and a sleeping-experts game that dynamically interpolates from \(\widetilde{\mathcal{O}}(\sqrt{T})\) to \(\widetilde{\mathcal{O}}(T^{2/3})\) depending on instance complexity, and proves that this rate is tight.

Instance-Adaptive Online Multicalibration

Presents the first instance-adaptive online multicalibration algorithm that dynamically refines a dyadic prediction grid via sleeping experts and minimax LP feasibility, interpolating between \(\widetilde{O}(\sqrt{T})\) benign and \(\widetilde{O}(T^{2/3})\) worst-case calibration rates without prior knowledge of the sequence structure.

Integral Probability Boundaries

Addressing the issue where standard Integral Probability Metrics (IPMs) collapse distribution discrepancies into a single scalar and yield loose or vacuous bounds at extremes, this paper introduces Integral Probability Boundaries (IPBs)—a non-linear generalization establishing deep connections with Neyman-Pearson testing and hockeystick divergence alongside finite-sample estimation algorithms.

Intrinsic-dimension empirical Bernstein inequalities for bounded self-adjoint operators

For bounded self-adjoint Hilbert-Schmidt operators and random matrices, this paper derives the first fully data-driven empirical Bennett and Bernstein concentration inequalities that bypass ambient dimension by scaling exclusively with intrinsic dimension, achieving asymptotic sharpness with oracle limits in non-isotropic estimation and infinite-dimensional kernel learning.

Intrinsic-Dimension Empirical Bernstein Inequalities for Bounded Self-Adjoint Operators

Establishes the first fully empirical Bennett and Bernstein inequalities for bounded self-adjoint Hilbert-Schmidt operators that depend strictly on the intrinsic dimension rather than ambient dimension, seamlessly extending to infinite-dimensional spaces and asymptotically recovering optimal oracle rates.

Intrinsic Riemannian Cross-covariance for Manifold-valued Random Objects

Addressing the lack of second-order dependency descriptors on non-Euclidean spaces, this paper introduces an intrinsic, footpoint-invariant Riemannian cross-covariance and correlation via geodesic parallel transport, proves asymptotic consistency and normality, and validates its robustness and unbiasedness across spheres, SPD manifolds, and Kendall's shape space.

Is a Linear Probe Evidence of a Linear Representation?

The paper proves that high linear probe accuracy cannot distinguish between a low-dimensional linear world model and a curved kernel machine, and introduces three asymmetric diagnostics—reverse predictivity, spectral concentration, and OOD cross-style transfer—to demonstrate that pretrained vision models lie strictly between the two regimes.

Is Backpropagation Optimal? When Synthetic Gradients Improve Sample Efficiency

This paper revisits backpropagation through the theoretical lens of gradient estimation sample efficiency, proving that when gradient feedback involves uncertainty and nodes satisfy local sufficiency, synthetic gradients can achieve lower estimation mean squared error than backpropagation through local bootstrapping.

Is \(\sqrt{d}\) Separation Necessary for Gradient EM to Learn Gaussian Mixtures in High Dimensions?

This paper rigorously proves that an \(\Omega(\sqrt{d})\) separation between ground-truth mixture component means is unavoidable in high dimensions for randomly initialized hybrid gradient EM to achieve global convergence in sub-exponential time, establishing an almost tight dimension-dependent barrier even in the over-parameterized regime.

Isotropic Activation Functions Enable Deindividuated Neurons and Adaptive Topologies

The paper introduces orthogonal-equivariant isotropic activation functions that eliminate the basis-dependence of neurons, enabling function-invariant layerwise SVD diagonalization, dynamic real-time neurogenesis and neurodegeneration, and asymptotic 50% lossless sparsification of dense networks.

Iterative Chow Filtering for Learning with Distribution Shift

Shows that the weaker notion of \(\mathcal{L}_1\) sandwiching polynomials suffices for efficient PQ learning under distribution shift, introducing Iterative Chow Filtering (ICF) to close the long-standing computational gap for DNF formulas and \(\mathsf{AC}^0\) circuits between standard PAC and selective domain adaptation.

Latent Process Generator Matching

This paper generalizes generator matching from conditioning on static latent variables to time-inhomogeneous Feller latent stochastic processes, rigorously establishing the pushforward Kolmogorov forward equation and gradient equality to unify and expand auxiliary dynamics marginalization across discrete and continuous spaces.

Learn from your own latents and not from tokens: A sample-complexity theory

Addressing the exponential sample complexity \(\mathcal{O}(vm^{L+1})\) required by token-level self-supervision on hierarchical data, this paper proves that predicting one's own latent representations reduces the sample complexity to \(\mathcal{O}(vm^3)\) independent of tree depth \(L\), and reveals that explicit multi-scale stacking like H-JEPA is theoretically redundant due to implicit hierarchical learning in models like data2vec.

Learning Distributions from Multiple Data Providers

This paper initiates the study of discrete distribution learning under restricted conditional sampling from heterogeneous, overlapping data providers, proving that co-occurrence graph connectivity is necessary and sufficient for pointwise consistency while completeness is necessary and sufficient for PAC learnability, with optimal sample complexity spanning a polynomial continuum from \(\widetilde{\Theta}(n/\varepsilon^2)\) to \(\widetilde{\Theta}(n^2/\varepsilon^2)\).

Learning Energy-Based Models from Stochastic Interpolants using Spatiotemporal Differences

Addressing the catastrophic failure modes where temporal methods break down under support mismatch and spatial methods collapse in multimodal regimes, this paper introduces Spatiotemporal Noise-Contrastive Estimation (stNCE), a unified binary classification framework that jointly models space-time density variations for efficient, consistent, single-NFE energy-based model training.

learning fractional-order dynamics from a single trajectory

For non-Markovian dynamical systems exhibiting long-range memory, this paper introduces FO-GS, a decoupled row-wise profiled least-squares grid search algorithm that establishes the first high-probability non-asymptotic statistical recovery guarantees of order \(\mathcal{O}(t^{-1/2})\) for both the fractional order \(\boldsymbol{\alpha}\) and transition matrix \(A\) from a single trajectory.

Learning Lotteries with Minimal Stakeholder Violations from Membership Queries

Addressing conflicting stakeholder standards in AI deployment where unanimous consensus is impossible, this paper formalizes the problem of minimizing the number of rejecting agents (Min-Viol) via binary membership queries, proves its APX-hardness, derives sample complexity bounds independent of population size \(n\), introduces a polynomial-time LP relaxation, and proves an exact matching deterministic lower bound of \(n-K\) on unqueried agents.

Learning Orthonormal Bases for Function Spaces

By reformulating orthonormal basis learning as continuous paths on the infinite-dimensional orthogonal Lie group manifold, this paper parameterizes low-rank skew-adjoint generators with neural networks and solves them via structure-preserving Cayley integration, achieving discretization-free, energy-conserving, and strictly orthonormal basis adaptation.

Learning Reveals Invisible Structure in Low-Rank RNNs

This paper extends the low-rank RNN framework from steady-state activity to gradient descent learning by deriving closed-form ODEs in a reduced overlap space, establishing a fundamental distinction between loss-visible overlaps that dictate function and loss-invisible overlaps that shape learning trajectories and encode past training history.

Learning Sparse Compositional Functions with Norm-Constrained Neural Networks

Breaking free from classical parameter-count limitations, this paper establishes the first approximation rates and statistical excess risk bounds for directed acyclic graph (DAG) sparse compositional functions under multiplicative Frobenius norm constraints, rigorously proving how overparameterized deep neural networks leverage hierarchical locality to avoid the curse of dimensionality.

Learning through Internalization

This paper formally defines the "internalization" process by which neural networks absorb multi-step computational scaffolding into their internal weights, empirically exposes that width benefits internalization more than depth alongside a severe degradation in out-of-distribution generalization, provides the first provable convergence guarantee for internalizing sparse parities with a one-layer linear attention transformer, and unifies the paradigm under positive distribution shift.

Learning to Decide with AI Assistance under Human-Alignment

This paper formalizes AI-assisted repeated binary decision-making as a two-armed online contextual learning problem with full feedback, proving that human-alignment compresses the cumulative regret bound from \(\Omega(\sqrt{|\mathcal{H}||\mathcal{B}|T})\) down to \(\mathcal{O}(\sqrt{|\mathcal{H}|T\log T})\) (and \(\mathcal{O}(\sqrt{T\log T})\) under weak conditions), with empirical robustness demonstrated across two real human-subject studies under partial alignment.

Learning to Price with Persuasion

This paper initiates the learning-theoretic study of monopolistic screening with persuasion under unknown buyer prior distributions, establishing sample and query complexity bounds across value samples, demand queries, and joint learning, while developing the first additive FPTAS for the underlying non-convex optimization.

Learning to Target with Network Interference

Focusing on online adaptive targeting under network spillovers, this paper proves that reducing aggregated outcomes to a standard linear bandit suffers from super-linear regret lower bounds, and develops near-optimal bandit algorithms under full, partial (degree only), and unknown support regimes.

Learning under Localized Minority Imbalance

To address localized minority imbalance (LMI) where minority instances are non-uniformly under-observed across feature regions, this paper proposes Stratified Ensembling (SE)—a tree-based feature partitioning procedure paired with global minority pooling—achieving superior generalization on unbiased target distributions.

Learning with Multiple Correct Answers: Regret Bounds under Different Feedback Models

This paper formalizes online learning with multiple correct answers under mistake-unknown, mistake-known, and set-valued feedback, establishes tight minimax mistake bounds in the realizable setting via tree-based combinatorial dimensions, and reveals a separation of linear, sublinear, and constant regret rates in the agnostic setting.

Length Generalization for Transformers via Compression

This paper establishes a polynomial compressed length complexity bound for the formal transformer specification languages C-RASP+ and C-RASP1 via power word compression, tightening the required sample size bound from doubly exponential to singly exponential and resolving empirical paradoxes surrounding the C-RASP length generalization hypothesis.

Leveraging Data Symmetries to Select an Optimal Subset of Training Data under Label Noise

This paper addresses the theoretical breakdown of cutstats in high dimensions by proving that leveraging group invariant representations under orthogonal and permutation symmetries eliminates the curse of dimensionality and restores the k-NN approximation to the Bayes optimal classifier, enabling downstream-agnostic clean subset selection under severe label noise.

LLM Flow Processes for Text-Conditioned Regression

This paper proposes coupling neural diffusion process trajectory priors with text-conditioned large language model marginal densities via gradient-free closed-form Gaussian convolution product-of-experts guidance, eliminating autoregressive error cascades and achieving calibrated text-conditioned function space selection.

Locality Sensitive Hashing for p-Exponential Kernels with Applications to Density Estimation

The paper proposes a novel mosaic locality sensitive hashing (LSH) scheme based on Poisson hyperplane processes (PHPs) and efficient \(\ell_1\)-biased \(p\)-stable vector sampling, proving that all \(p\)-exponential kernels (\(\exp(-\|x-y\|_p/\sigma)\)) for \(1 < p \le 2\) in bounded regions are exactly LSHable, and applies it to achieve efficient differentially private and relative-error kernel density estimation.

Logistic Bandits with O~(sqrt(dT)) Regret without Context Diversity Assumptions

This paper proposes SupSplitLog, the first \(K\)-armed logistic bandit algorithm that achieves the rate-optimal \(\widetilde{\mathcal{O}}(\sqrt{dT})\) regret bound without any context diversity assumptions (such as strictly positive lower bounds on the minimum eigenvalue of the context covariance matrix), eliminating mandatory random exploration and extending regret bounds to a data-dependent geometric complexity measure.

Logistic Bandits with \(\tilde{O}(\sqrt{dT})\) Regret without Context Diversity Assumptions

SupSplitLog achieves rate-optimal \(\tilde{O}(\sqrt{dT})\) cumulative regret for \(K\)-armed logistic bandits without any context diversity assumption by orthogonally splitting samples into a pilot set for coarse estimation and an estimation set for one-step Newton correction.

Mean Testing under Truncation beyond Gaussian

Addressing the severe polynomial bias barrier \(\mathcal{O}(\nu_{P,p}\varepsilon^{1-1/p})\) encountered in high-dimensional mean testing under arbitrary truncation, this paper completely characterizes the information-theoretic indistinguishability floor alongside a matching optimal \(\mathcal{O}(\sqrt{d})\) second-order U-statistic test, and introduces directional median regularity to overcome this barrier, shrinking the truncation bias down to linear order \(\mathcal{O}(\varepsilon)\).

Measure-to-measure Regression with Transformers

This paper formalizes nonlinear measure-to-measure (M2M) regression and leverages the inherent mean-field and measure-dependent structure of transformers to introduce static pushforward maps and dynamic Transformer Flow Matching (M2M-TFM) operators for continuous distributional extrapolation.

Mini-batch kernel k-means

Presents the first mini-batch kernel \(k\)-means algorithm that reduces per-iteration update complexity to \(\widetilde{O}(k b^2)\) via sparse sliding-window center truncation, with proven dimension-independent \(O(\gamma^2/\epsilon)\) iteration termination and an \(O(\log k)\) expected approximation ratio under early stopping.

Multigroup Fairness and Omniprediction: Separations and Equivalences

The paper proves that standard omniprediction does not even require accuracy in expectation (ruling out calibration and multigroup fairness as necessary conditions), whereas the stronger notion of loss outcome indistinguishability (Loss OI) is strictly equivalent to calibrated multiaccuracy under mild neutrality conditions.

Nearly Optimal Robust Covariance and Scatter Matrix Estimation Beyond Gaussians

Addressing the robust estimation problem for high-dimensional elliptical distributions under the strong contamination model, this paper introduces a multi-stage spatial sign filtering framework with degree-4 sum-of-squares (SoS) relaxations that achieves nearly optimal error matching the Gaussian setting in polynomial time and sample complexity.

Neural Expansion: A Unified Mechanism for How Deep Neural Networks Generalize

This paper introduces Neural Expansion, a geometric framework utilizing the singular value decomposition of penultimate-layer feature Jacobians and directional Generalization Intervals (GIs) to quantify ray coverage around training samples, revealing that overparameterized networks generalize not via instance memorization, but by anisotropically expanding classification corridors along flat, low-curvature directions.

Neural Networks With Dense Weights Are Not Universal Approximators

Combining the continuous graph-theoretic weak regularity lemma with message-passing kernel representations, this paper establishes a deterministic model compression theorem proving that strongly dense feedforward networks lacking privileged dominating weights hit an insurmountable expressivity saturation barrier, and thus fail to universally approximate Lipschitz continuous functions.

No Triangulation Without Representation: Generalization in Topological Deep Learning

This paper extends the combinatorial manifold benchmark MANTRA, demonstrates that standard graph neural networks match or surpass higher-order message passing under proper representations and feature encodings, and introduces a subdivision-based generalization stress test showing that current topological deep learning models exploit combinatorial artifacts rather than genuine topological invariants.

Nonparametric In-Context Learning under Growing Geometric Complexity: Minimax Optimality and Local Geometry-Adaptivity of Transformers

The paper establishes the minimax rate of nonparametric in-context learning on heterogeneous manifold mixtures whose complexity grows with sample size, and proves that a structure-informed transformer approximates a tangent local-polynomial estimator attaining that rate; optimality of a trained predictor still requires sufficient training tasks and a sufficiently small empirical-risk optimization gap.

On Making SE(2)-Invariant Networks Optimal

Grounding on classical invariant theory and quotient manifold theorems, this paper derives provably optimal (minimal and discriminative) \(\mathrm{SE}(2)\)-invariant feature maps, constructing a plug-and-play convolutional module that seamlessly integrates into standard CNNs while achieving superior rotational robustness and up to \(14\times\) speedups over steerable equivariant networks.

On the Convergence of Multicalibration Gradient Boosting

Addressing the lack of multi-round convergence theory for web-scale multicalibration gradient boosting, this paper rigorously models the recursive feature-augmented dynamics as a discrete nonlinear dynamical system, proving an \(\mathcal{O}(1/\sqrt{T})\) global convergence bound for prediction updates and empirical multicalibration error, a linear rate under smooth weak learners, and local quadratic convergence for adaptive rescaling.

On the depth of monotone ReLU neural networks and ICNNs

By bridging neural networks with polyhedral geometry via Newton polytopes, this paper proves that monotone ReLU networks cannot compute or approximate \(\mathsf{MAX}_n\), establishes a sharp linear depth lower bound of \(n\) for ICNNs on \(\mathsf{MAX}_n\), and shows that depth-2 standard ReLU networks can be quadratically more depth-efficient than ICNNs.

On the Tightness and Computational Tractability of Higher-Dimensional Confidence Sequences

This paper systematically investigates multivariate lifts of 1D betting-based confidence sequences for bounded random vectors, proves that the portfolio confidence region achieves the tightest coverage but is computationally intractable, and introduces three tractable outer approximations (bounding box, \(\ell_p\)-ellipsoid, and their intersection) with theoretical between-grid safety guarantees, enabling fast anytime-valid sequential decision-making.

One-Bit Clustering for Two Component Sub-Gaussian Mixture Models

Addressing the extreme low-bit clustering bottleneck in two-component symmetric sub-Gaussian mixture models under severe communication constraints, this paper introduces the first dithered one-bit quantization clustering framework with a modified Lloyd's algorithm, establishing partial recovery rates matching unquantized counterparts, minimax lower bounds with sharp constants, and exact recovery free of center spikiness constraints via Haar random rotation preprocessing.

One-Shot Generative Flows: Existence and Obstructions

This paper develops a structural theory of stochastic processes with independent endpoints that induce zero-acceleration straight-line flows, deriving an Eulerian-Lagrangian PDE characterization, constructing explicit one-shot interpolants for arbitrary non-degenerate Gaussian endpoints, and proving fundamental impossibility theorems for multi-modal targets.

Online Budget Allocation with Censored Semi-Bandit Feedback

This paper introduces the stochastic budget-allocation problem under censored semi-bandit feedback, designing an optimism-based algorithm (UCBowski) that achieves a parameter-free polylogarithmic regret rate \(\mathcal{O}(\mathrm{poly}(K)(\log T)^2)\) under diminishing-returns regimes and a minimax-optimal worst-case regret \(\widetilde{\mathcal{O}}(K\sqrt{T})\) matching an \(\Omega(K\sqrt{T})\) lower bound.

Online Set Learning from Precision and Recall Feedback

This paper introduces an online learning formulation for predicting unknown target subsets under stochastic item-level precision and recall feedback, proving that finite VC dimension characterizes learnability with sublinear regret in both realizable and agnostic settings.

PAC-Bayesian Bounds for Learning Partially Observed Stochastic Linear Time-Invariant State-Space Systems with Inputs and Sub-Gaussian Noise

This paper establishes the first algorithm-independent PAC-Bayesian generalization gap bounds of order \(\mathcal{O}\big(\frac{(\ln N)^r}{\sqrt{N}}\big)\) with optimal logarithmic confidence dependence \(\mathcal{O}(\ln(1/\delta))\) for partially observed stochastic linear time-invariant (LTI) state-space systems with inputs and unbounded sub-Gaussian noise, while also deriving finite-sample parameter estimation and \(H_2\) transfer function error guarantees for MEP, MAP, and Gibbs sampling estimators.

PAC Learning with Bandit Feedback: Sharp Sample Complexity in the Realizable Setting

This paper resolves the sample complexity of multiclass PAC learning with bandit feedback in the realizable setting by introducing the bandit DS dimension \(\operatorname{BDS}(\mathcal{H})\) based on pseudo-boxes, achieving a near-optimal matching bound \(\widetilde{\Theta}(\operatorname{BDS}(\mathcal{H})/\epsilon)\) via the ListCascade algorithm.

PENEX: AdaBoost-Inspired Neural Network Regularization

By reformulating multi-class exponential loss with a SumExp barrier penalty instead of rigid zero-sum constraints, PENEX translates AdaBoost's margin-maximizing resistance to overfitting into deep neural networks, yielding an unconstrained, self-calibrating regularizer that consistently surpasses cross-entropy and established baselines in low-data and noisy regimes.

Performative Prediction with Selective Labels

This paper formalizes performative prediction with selective labels, demonstrates that retraining only on accepted sub-populations can cause arbitrary divergence from performatively stable solutions under deterministic thresholds, and proposes Robust Repeated Risk Minimization (Robust RRM) with a strongly convex upper-bound loss that guarantees a bounded distance to the true stable point while leveraging historical acceptance data to tighten label uncertainty intervals over time.

Playing Markov Games Without Observing Payoffs

This paper extends online learning in symmetric games without observing payoffs to Markov sequential environments, establishing a counter-intuitive symmetry hierarchy and proving that local mimicking and recursive expected payoff redistribution achieve near-optimal sublinear regret against an omniscient adversary under completely unknown transitions.

Prudent-Banker: No Extra Fees for Baseline Safety in Adversarial Bandits With and Without Delays

Under arbitrary unobserved feedback delays, Prudent-Banker is the first algorithm to achieve both minimax-optimal worst-case regret \(\widetilde{\mathcal{O}}(\sqrt{T}+\sqrt{D})\) and \(\widetilde{\mathcal{O}}(1)\) regret relative to a safe baseline policy, accompanied by matching information-theoretic lower bounds.

Pure Exploration Beyond Reward Feedback: The Role of Post-Action Context

This paper introduces fixed-confidence Best Arm Identification (BAI) with post-action context, establishes instance-dependent sample complexity lower bounds for both separator and non-separator causal structures, and proposes G-tracking and extended Track-and-Stop algorithms that asymptotically achieve optimal sample complexity.

Rank-Transformed Dissimilarity Profiles for High-Dimensional Classification

The paper proposes a rank-transformed dissimilarity-profiling framework for high-dimensional classification that converts inter-point geometric distance patterns—often viewed as artifacts of the curse of dimensionality—into low-dimensional rank representations, adaptively capturing location, scale, and higher-order moment shifts while conferring strong robustness against contamination.

Regret–Oracle Complexity Tradeoffs in Agnostic Online Learning

Addressing the doubly-exponential query blowup \(\mathcal{O}(T^{2^{\mathcal{O}(d_{\mathrm{LD}})}})\) in classical agnostic-to-realizable reductions, this paper introduces ADEPT, an adaptive pruning framework utilizing a weak-consistency oracle to dynamically eliminate unrealizable paths, slashing the oracle complexity down to \(\mathcal{O}(T^{d_{\mathrm{VC}}+1})\) while maintaining near-optimal regret.

Robust PAC Learning of Concurrent Stochastic Games

Introduces PAC-CSG, the first PAC learning framework for two-player general-sum concurrent stochastic games under transition uncertainty that combines \(L^1\) confidence sets with an exploration robust MDP to output an \(\varepsilon\)-optimal \(\varepsilon\)-Nash equilibrium or a sound non-existence certificate in \(\widetilde{\mathcal{O}}(R_{\max}^2 H^4 |S|^2 |A| / (p_{\text{reach}} \varepsilon^2))\) trajectory samples.

Sharp Convergence and Sample Complexity of Policy Mirror Descent for Average-Reward MDPs

For ergodic average-reward MDPs, this paper establishes a master Lyapunov recursion for unregularized policy mirror descent that delivers linear convergence across exact, tabular, and single-trajectory linear function approximation settings, achieves a tight \(\widetilde{O}(t_{\mathrm{mix}}^3/\varepsilon^2)\) sample complexity, and matches an information-theoretic critic lower bound.

Spectrum-Adaptive Generalization Bounds for Trained Deep Transformers

To address the exponential depth explosion and dimension-dependent norm constraints in existing generalization bounds for deep Transformers, this paper proposes the first post hoc spectrum-adaptive generalization bound, allowing layerwise and matrixwise selection of Schatten indices after training to match the learned singular-value decay.

The Panel Complexity of Sortition: Is 12 Angry Men Enough?

This paper introduces a novel definition of representative panels based on the Wasserstein distance from statistical learning theory, establishing near-tight panel complexity bounds for social welfare maximization, core fairness, and outlier deviation control in participatory budgeting and facility location.

The Pareto Frontier of Randomized Learning-Augmented Online Bidding

Introduces a novel continuous "bidding function" analytical framework for randomized learning-augmented online bidding, fully closing the Pareto frontier with matching analytical upper and lower bounds for \(R \ge 2/\ln 2 \approx 2.885\), and improving the near-optimal consistency bound to \(e - \Omega(\varepsilon^{1/4})\) as \(R \to e\).

The Price of Locality: Why Forward-Forward Underperforms Backpropagation?

This mechanism-analysis paper combines conditional local convergence theory, spectral analysis of representation kernels and error signals, and interventions on training groups and update spectra to explain why the Forward-Forward Algorithm (FFA) often trails back-propagation (BP); relaxing gradient isolation between CNN12 layers raises accuracy from 53.02% to 78.74%, still below BP's 85.67%.

The Sample Complexity of Multiple Change Point Identification under Bandit Feedback

This paper investigates multiple change point localization under active bandit feedback in continuous domains, introducing the LCP algorithm that decouples coarse multiscale detection from sequential halving refinement to establish rate-optimal non-asymptotic sample complexity and minimax lower bounds governed by local energies.

Tight \(L_\infty\) Sample Complexity for Low-degree and Sparse Boolean Polynomials

Motivated by the need for optimization-safe surrogates in pseudo-Boolean optimization, this paper completely characterizes the tight minimax sample complexity of learning low-degree and sparse Boolean polynomials under subgaussian noise with uniform \(L_\infty\)-error guarantees, proving that noise imposes an unavoidable dimensional penalty relative to the noiseless regime.

Time-Sensitive Anytime-Valid Testing

The paper formulates anytime-valid testing-by-betting as an optimal control problem that maximizes expected rewards over rejection times, establishes a Markovian Bellman representation and equivalence to general e-processes, and derives the closed-form exponential-decay-optimal (EDO) betting rule governed by Rényi divergences.

Training with (Swap) Regret Loss in a Single-Layer Self-Attention Model: A Case Study on the Probability Simplex

This paper proves that training single-layer self-attention models directly on the probability simplex using external and newly introduced swap regret losses yields stationary points whose forward passes exactly recover smoothed fictitious play and multi-head Blum–Mansour dynamics, driving multi-agent game play to coarse correlated and correlated equilibria without algorithm supervision.

Transfer Learning of Linear Regression with Multiple Pretrained Models: Benefiting from More Pretrained Models via Overparameterization Debiasing

This paper investigates transfer learning for high-dimensional linear regression using multiple least-squares pretrained models, derives exact asymptotic test errors using random matrix theory, reveals how overparameterization bias causes inconsistency when scaling the number of pretrained models, and proposes a simple multiplicative debiasing principle that restores statistical consistency.

Transfer Learning Through Conditional Quantile Matching

Addressing the ubiquitous challenge of coexisting covariate and concept shifts in multi-source regression transfer learning, this paper proposes TLCQM, a framework that trains conditional generative models on individual source domains and calibrates synthetic responses via conditional quantile matching at target covariates, theoretically establishing a tighter empirical risk minimization excess risk bound over target-only learning without restrictive shift assumptions.

Tree Search With Predictions

This paper investigates target search on trees guided by machine-learned predictions, establishes that \(O(\log \mathrm{dist}(s,t))\) query complexity is unattainable on general trees, and introduces a recursive exponential search algorithm on \(k\)-spine decompositions that achieves an existentially optimal \(O(k \log \mathrm{dist}(s,t))\) query bound for trees of pathwidth \(k\).

Two-Fidelity Best-Action Identification for Stochastic Minimax Tree

Under a known fast-evaluation bias envelope and slow evaluations unbiased for true node minimax values, 2FFS uses endpoint certificates and recursive budgets to choose adaptively between expansion and sampling, achieving high synthetic-tree accuracy with substantially fewer node visits, while stopping and cost guarantees require additional regularity conditions.

Uncertainty-Aware Probabilistic Constrained Clustering from Entangled Pairwise Supervision

Addressing the entanglement of intrinsic aleatoric ambiguity, expert epistemic judgment, and stochastic recording noise in realistic pairwise supervision, this paper formalizes Uncertainty-Aware Probabilistic Constrained Clustering (UPCC), establishes an angular pairwise objective ProbPair, and builds an Estimator–Corrector–Integrator (ECI-PP) framework that substantially outperforms state-of-the-art deep constrained clustering methods.

Uncovering Challenges of Solving the Continuous Gromov-Wasserstein Problem

The paper reveals that existing continuous Gromov-Wasserstein (GW) solvers catastrophically collapse when data lacks paired correlation, and introduces NeuralGW, the first purely continuous solver built on minimax adversarial dual reformulation without discrete optimal transport approximations.

Why Heavy-Tailed Weight Spectra Predict Model Quality

This paper derives a data-independent Bayes free energy evaluated solely on trained final-layer weights and proves through High-Temperature Marchenko-Pastur theory that scale-based spectral heaviness dictates overfitting transitions while heavy-tailed exponents drive generalization.

Disentanglement as Identifiable Pushforward Factorisation

By formalising disentanglement as the factorisation of smooth pushforward manifold density into statistically independent 1-D seam factors induced by Jacobian SVD, this paper proves that disentanglement is equivalent to Jacobian axis-alignment (C1) and singular value derivative diagonality (C2), guarantees identifiability up to permutation and sign, and reveals how diagonal posterior covariances in VAEs enforce these conditions via an exact Price/Bonnet identity.