New Machine Learning (Statistics) (stat.ML) papers, Tuesday 20 October 2026
51 papers announced on arXiv. Get them ranked for you, free →
Announced Breaks Separate Conformal Reliability from Frequency Calibration
AI summaryThe paper proposes a reset rule at announced boundaries that separates conformal reliability from frequency calibration, achieving minimax per‑time rate for independent scores. Experiments show it outperforms five adaptive methods across synthetic and real data.
miniTE: Diagnosing and Bridging the Execution Gap in Native NVFP4 Reinforcement Learning
AI summaryThis paper introduces miniTE, a framework that diagnoses and bridges cross-engine numerical discrepancies in native NVFP4 reinforcement learning for LLMs. It achieves stable training with near-BF16 performance while delivering up to 1.47x the throughput of standard NVFP4.
Graph-Regularized Synthetic Control for Stable Counterfactual Estimation under Donor Redundancy
AI summaryGraph-Regularized Synthetic Control adds a graph Laplacian penalty to classical synthetic control, encouraging coherent weights across similar donors. It yields more stable, interpretable counterfactuals while matching pre‑ and post‑treatment accuracy, especially in donor‑redundant settings.
Margin Certificates for Voting Classifiers under Mixing Dependence
AI summaryThis paper derives uniform margin certificates and generalization bounds for voting classifiers under stationary mixing dependence. The results provide non-vacuous guarantees and valid confidence intervals on dependent time-series data where standard i.i.d. methods fail.
On Reparameterizing the Score Function
AI summaryThis paper introduces History-Reparameterized policy gradients to alleviate the credit-assignment bottleneck in autoregressive sequence models by backpropagating through the conditioning history. The method uses an optimal control-variate estimator to reduce gradient variance and improves performance on sparse-reward tasks.
Certifiably Correct Fréchet Means in Orbit Spaaces
AI summaryComputes Fréchet means for finite orthogonal group orbit spaces using spectral and SDP relaxations. A new semidefinite certificate guarantees global optimality even when pairwise alignment fails, proven for Gaussian data modulo sign and validated on districting and time‑series clustering tasks.
Censored Flow Matching for Synthetic Time-to-Event Generation
AI summaryCensored Flow Matching (CFM) is a flow-based generative method for time‑to‑event data that handles censoring via pullback sampling, yielding an EM algorithm that converges to the true event‑time distribution and enables realistic synthetic data and accurate density estimation.
CoCoNet: Convex Conformal Network for Multi-Dimensional Uncertainty Regions
AI summaryCoCoNet learns convex conformity scores for multivariate regression by optimizing prediction‑region efficiency on held‑out data, then applies split‑conformal calibration to keep coverage guarantees. Across 30 tasks, the ICNN variant achieves the highest efficiency, ranking first on 22 of 30 tasks.
Constructive Universal Approximation by Infinite-Dimensional Neural Networks via Gaussian-Weighted Ridgelet Transform
AI summaryThe paper extends the ridgelet transform to infinite-dimensional Hilbert spaces using Gaussian-weighted directions, yielding explicit coefficient measures and reconstruction formulas. It proves uniform approximation of function‑valued neural networks for non‑polynomial activations, with a dimension‑free Monte‑Carlo error rate \(O(N^{-1/2})\).
Morpheus-2.0: Multiscale Probabilistic Forecasting with Dual-Axis Attention
AI summaryMorpheus‑2.0 trains a single encoder‑only model that jointly attends across variables and time on a variable‑time grid, sampling horizon lengths during training. The multiscale version handles multiple resolutions, achieving 0.646 relative MASE and 0.455 relative CRPS, the best reported.
Uniform convergence at optimal rate and scale
AI summaryUniform convergence for real-valued function classes now attains the optimal sample complexity of \(O(d/\varepsilon^2)\) at the optimal scale \(γ\), matching the binary VC case. This is achieved via margin‑ε‑nets, extended Haussler covers, and an aligned chaining argument.
The Certification Exponent: Instance-Optimal Sequential Certification of Policy Improvement from History Dependent Logs
AI summaryThe paper shows that the optimal number of episodes needed to certify a policy’s improvement from history‑dependent logs is \(\Theta(\log(1/\delta)/K^\star)\), where \(K^\star\) is the minimum KL divergence to a “bad” law. A plug‑in Kelly certifier attains this bound, and a backward‑induction “coverage” test identifies whether certification is impossible, finite‑rate, or constant‑time.
Private Evaluation of High-Dimensional Functions
AI summaryThis paper introduces generic differentially private black-box evaluators for high-dimensional functions via a reduction to private quasi-concave optimization. The method guarantees outputs within the convex hull of local function evaluations while improving data overhead for user-level optimization.
Denoising Implicit Variational Inference
AI summaryThis paper introduces Denoising Implicit Variational Inference, which applies denoising score matching to directly estimate marginal scores in semi-implicit variational families. This approach outperforms existing implicit variational inference baselines while reducing gradient estimation costs.
Accelerating Finite-Sum Bilevel Optimization with Fully First-Order Oracles
AI summarySICBO is a fully first‑order algorithm for nonconvex‑strongly‑convex bilevel problems with finite sums, achieving a gradient‑norm error ε with \(\tilde{O}(N+\sqrt{N}\kappa^{7/2}\varepsilon^{-2})\) component‑gradient calls, optimal up to log factors for fixed κ.
Conditional Weight of Evidence for Gradient-Boosted Credit Models: Estimation, Additivity and Its Relation to Adverse Action Reasons
AI summaryThis paper introduces conditional weight of evidence for gradient-boosted credit models to assess a characteristic's impact given an applicant's full record. The proposed tree-based estimators accurately recover applicant-specific credit information and complex feature interactions missed by traditional marginal measures.
Gaussian Process Bandits with Paired Auxiliary Observations
AI summaryThis paper introduces a variance-reduction method for Gaussian process bandits that leverages low-cost auxiliary data alongside paired online observations. The approach provides theoretical regret guarantees and demonstrates consistent improvements in cumulative and simple regret over standard GP-UCB across benchmark tasks.
Martingale Posterior Cox Processes
AI summaryMPCP estimates a Cox process intensity by treating unobserved events as latent variables, eliminating complex priors and enabling a GPU‑friendly copula‑based predictive update. It converges theoretically and runs far faster than existing Bayesian methods while matching or exceeding their performance.
Tilted Mondrian: Triage and Mitigation of Fine-grained Disparities in Conformal Prediction
AI summaryTilted Mondrian adjusts standard Mondrian conformal prediction to reduce fine‑stratum coverage and set‑size disparities while preserving group coverage guarantees, using propensity standardization to quantify severity and a method‑agnostic bound to measure achievable fairness gains.
A Statistical Viewpoint on Adam, Shampoo and Muon
AI summaryThe paper proposes a statistical framework that whitens gradients using their covariance, leading to a diagonal method (MND) that matches Adam and outperforms Shampoo variants on GPT-2‑scale models, and a Muon variant (MN‑Muon) that improves stability and validation loss.
Scalable and Adaptive Distribution Regression via Random Features
AI summaryRaDR uses random feature approximations to achieve scalable distribution regression with a finite-sample guarantee, matching the excess risk of exact kernel methods under a feature-budget constraint. A-RaDR adaptively assigns features via a score-guided scheme, improving accuracy while keeping the speed advantage.
Robust Direct Preference Optimization under Preference Label-Flip Attacks
AI summaryThe paper develops defense mechanisms for Direct Preference Optimization against label‑flip attacks, introducing Trim‑K and Soft‑Trim‑K methods that remove or downweight high‑impact pairs. It proves deterministic gradient‑perturbation bounds and shows these defenses achieve uniform policy deviation guarantees under bounded‑region assumptions.
Dimension-free comparison inequalities between $f$-divergences for Gaussian location mixtures
We establish dimension-free comparison inequalities between the Kullback-Leibler (KL) and $\chi^2$-divergences and the squared Hellinger distance for Gaussian location mixtures with compactly supported mixing distributions. We show that both divergences are bounded by the…
Decision-Target Reversal under Informative Label Missingness
AI summaryThe paper shows that informative label missingness can improve classification under one loss but worsen it under another—“decision‑target reversal.” For binary Gaussian discrimination it derives critical cost ratios where this reversal occurs, extending to multiclass settings and demonstrating with simulations and a medical case.
A Distribution-Free Separation Between Learned and Fixed Features
AI summaryA two‑layer ReLU network learns a constructed hypothesis class under any input distribution within any specified runtime, while any linear predictor with fixed features would require super‑polynomial dimension, showing a distribution‑free separation between learned and fixed features.
DR-Learners for Generalizing CATEs from a Randomized Controlled Trial
Randomized clinical trials are essential for evaluating interventions, yet average treatment effects can obscure treatment effect heterogeneity, making conditional average treatment effects (CATEs) a natural target. However, restrictive eligibility criteria and differential…
Reinforcement learning on the EPRI AI-ccelerating Unit Commitment benchmark: diagnosing over-commitment and guiding MILP with learned policies
Reinforcement learning (RL) is increasingly proposed as a fast solver for unit commitment (UC), but standalone agents remain far more expensive than mixed-integer linear programming (MILP). In the first study of RL on the 51-generator, 72-hour benchmark of the Electric Power…
Divergence based Ranked Preference Alignment for LLMs
Alignment of large language models typically relies on pairwise preference feedback, though ranked feedback can capture more comprehensive information across multiple responses within a single annotation. However, ranked-preference alignment is underexplored in practice compared…
Corrected Block Sparse Spectral Optimization: Exact Models, Finite Identification, and Local Dynamics
The block generalized-power method pairs threshold-based support selection with inexpensive Stiefel updates, but the standard block formulation conflates two nonequivalent objectives and leaves inactive components, threshold ties, and post-selection refitting ill-specified. We…
Detection Augmented Learning In Non-Stationary Bandit Environments
We study the problem of non-stationary bandits without knowledge of the environmental changes. We propose Detection Augmented Learning (DAL) a black-box paradigm that comprises three modular components: a bandit algorithm, a change detector, and a forced exploration schedule.…
Fair Regression under Approximate Demographic Parity
Exact demographic parity requires the same prediction distribution across sensitive groups. In regression, this can be too restrictive and reduce predictive accuracy. We instead bound every pairwise Wasserstein distance between group-specific prediction distributions by a chosen…
Statistical Learning Theory for Direct Preference Optimization
Direct Preference Optimization (DPO) learns from pairwise preferences through a logistic objective whose score-difference parameterization follows the Bradley--Terry model. We study how failure of this modeling assumption limits DPO, distinguishing structural misspecification…
Keep the Lineages Alive: Diversity-Preserving SMC for Diffusion LLM Decoding
Inference-time sequential Monte Carlo (SMC) decoding can improve the reasoning accuracy of masked diffusion language models (diffusion LLMs) without post-training by resampling a population of partial generations according to a reward. Its effectiveness therefore depends on the…
Federated Community Detection in Bipartite Networks
We study federated spectral estimation when a large matrix is horizontally partitioned across servers that cannot share raw data. We propose FLICKER, a target-specific gradient-based power iteration that recovers a singular direction associated with a prescribed spectral target…
Mediation and moderation in distance-based regression with triangular fuzzy data: simulation evidence and an application to academic performance
We study mediation and moderation in the distance-based regression model for triangular fuzzy responses, in which the center and transformed spreads of the response are linear in the components of fuzzy and crisp covariates. With a common design, componentwise least squares…
Towards instance-dependent regret optimality in Episodic MDPs with Posterior Sampling
We study regret minimization in finite-horizon episodic Markov Decision Processes (MDPs). While minimax-optimal algorithms are known, tractable approaches to instance-dependent optimality are still lacking. Motivated by this gap, we introduce $\pi_0$-PSRL, a variant of PSRL that…
Truncate or Not? A Scaling Study of Looped Transformers with Truncated Backpropagation
Looped transformers reuse a shared block of layers to increase computational depth without increasing the number of trainable parameters. They are trained with backpropagation through time (BPTT). To save training compute, BPTT is often truncated to the last few loops, although…
A general condition for fast rates in multi-group learning
In multi-group learning, a single predictor must control its excess risk simultaneously over overlapping groups. Tosh and Hsu (2022) obtained $1/\sqrt{n}$ slow rates via a reduction to online learning with sleeping experts, and $1/n$ rates only under realizability (and for a…
Causal Robustness from Synthetic Environments
AI summaryThis paper introduces a method to learn causally robust predictors from a single observational distribution by generating synthetic environments via covariate reweighting. The main result is that this approach achieves robust prediction and invariance against distribution shifts using only observational training data.
Selecting the Best Few with Human Evaluators: Adaptively Combining Ratings and Pairwise Comparisons
This paper considers a decision-maker tasked with identifying the $k$ best items from a large pool of candidates, such as job applicants, LLMs, award nominees, or new product ideas. The decision-maker learns about the items' quality by soliciting human evaluations. Each…
A Simple and General Approach to top-$k$ Selection
This paper considers the Bayesian top-$k$ identification problem, wherein a decision maker is tasked with selecting $k$ items from a pool of items of unknown quality. To assess the items, the decision maker adaptively allocates a fixed sampling budget, where each sample of an…
Local data augmentation for imbalanced binary classification
Facing imbalanced binary classification, we aim at maximizing the recall while limiting the proportion of false detections in the minority class. For doing that, the present work studies the effect of several rebalancing strategies, from the uniform subsampling of the…
R2-GP: Variance-Calibrated Priors for Additive Gaussian Processes
We develop R2-GP, an R2D2-type prior for Gaussian process (GP) generalised additive models (GAMs) that calibrates explained variance ($R^2$) for Matérn covariance kernels. The function variance for standard Matérn kernels depends jointly on the effect magnitude and length-scale.…
WhatIf-Bench: Benchmarking Intervention-Aware Time Series Forecasting
Recent time series foundation models (TSFMs) can consume covariates in-context, extending zero-shot forecasting to problems driven by external factors. This enables scenario analysis: forecasting how a target responds to events that set covariates to chosen future values.…
Which Model for Which Role? Cost-Efficient Pareto Recommendation for LLM Agentic Workflows
LLM agentic workflows increasingly employ different models for specialized roles. Users may prefer different trade-offs between task quality and deployment cost, sometimes also considering latency. Finding good assignments incurs a separate search cost from evaluating multiple…
Dirichlet-Tree Allocation Priors for Bayesian Causal Forests
AI summaryDirichlet-Tree Bayesian Causal Forests regularize treatment tree leaf values by a stick-breaking Dirichlet prior, bounding effect heterogeneity and preventing variance inflation as trees grow. Simulations show improved accuracy over standard BCF when few heterogeneous regions exist.
Cross-Fitting for Neural Posterior Estimation
Neural Posterior Estimation (NPE) fits an amortized conditional density estimator to a simulated training set of complete-data realizations. NPE's training objective, log-likelihood, favors posterior approximations that assign arbitrarily high density to the training points, and…
Cooperative Multi-Agent Reinforcement Learning for Market Making on Prediction-Market Order Books
A market maker provides liquidity by quoting prices to buy and sell an asset, earning the spread while managing inventory risk. We study market makers trained with multi-agent reinforcement learning on prediction markets, specifically Polymarket, whose thin order books pose…
High-dimensional covariate-assisted ranking
Ranking data in modern applications often involves covariate vectors whose dimension grows with the number of objects. Identifying relevant covariates and integrating them with item-specific utilities can yield better predictive performance. In this work, we study a…
Teacher Supervision, Fixed Representations and Calibration: Exact Risk Calculations for Candidate-Scoring Decision Models
Candidate-scoring decision models return a probability distribution over a presented set of answers. Public documentation of one such system, Jev, and of an open counterpart, JevAny, describes three training ingredients that determine what these probabilities estimate: soft…
Learning to Refer: Inference-Time Collaboration in Decentralized Federated Learning
AI summaryEdgeDRM enables inference-time collaboration in decentralized federated learning by learning local density potentials that guide query referrals along a graph, aggregating predictions from visited clients. It outperforms graph‑matched baselines, especially in sparsely connected settings.