Machine Learning
Audio-LLMs as Data Filtering Agents for Speech-to-Speech Translation Training
Chen and Nakamura propose a method for training an audio-language model (Audio-LLM) to filter noisy, misaligned, or semantically incorrect speech pairs from large-scale mined corpora used for end-to-end speech-to-speech translation (S2ST). Their two-stage Rank-to-Distill strategy first uses a lightw…
Simplex-Constrained Sparse Bagging: Achieving Compression and Calibration via Sparse Posteriors
Simplex-Constrained Sparse Bagging (SCSB) is a novel, model-agnostic framework for compressing and calibrating bootstrap-based bagging ensembles. It addresses the issue of uniform estimator weighting in standard bagging by optimizing a joint pruning and calibration process over the probability simpl…
Stratified Classification Trees via Simultaneous Latent Budget Models
Simultaneous Latent Budget Trees (SLBT) extend classification trees to explicitly handle stratification factors (e.g., time, space, demographics) by replacing standard impurity-based splits with a model-based split rule derived from Simultaneous Latent Budget models. Child nodes are interpreted as l…
Mixup Distillation Transfers Calibration and Linearity Beyond Dark Knowledge
This paper investigates the interaction between knowledge distillation and mixup when mixup is applied only during student training, revealing that the teacher's signal becomes dominated by distributional confusion rather than class structure. Despite this, the student independently learns superior …
xLSTM Outperforms Mamba-2 and Gated DeltaNet in Complex Sequence Modeling Tasks Due to Enhanced Memory Management
This paper compares subquadratic architectures for sequence modeling, specifically xLSTM, Mamba-2, and Gated DeltaNet, on demanding tasks like code and time-series modeling. xLSTM emerged as the top performer, attributed to its superior state tracking and memory correction mechanisms facilitated by …
Data-Driven Dynamic Assortment with Unknown Two-Sided Preferences Achieves Optimal Rate Regret
This paper addresses the dynamic assortment problem on a two-sided service platform with initially unknown customer and seller choice parameters. A novel data-driven algorithm is proposed that learns these parameters and optimizes the platform's objective concurrently. The algorithm's performance is…
Quantized Distributed Optimization Achieves Centralized Stochastic Rates Under Relaxed Geometry
This paper proposes q-PDGD, a quantized stochastic primal-dual method for distributed optimization under communication constraints (random unbiased quantization). The algorithm achieves linear convergence to a noise-dependent neighborhood under both RSI and PL inequalities with constant step-sizes, …
Theoretical Framework for Test-Time Training: Sampling, Oracle Complexity, and Distribution Classes
This work formalizes test-time training (TTT) as sampling from a distribution $\mu^\star$ using an oracle for approximate density estimates. A quadratic lower bound on query complexity is established for general distribution classes, proving the optimality of existing random walk sampling methods. H…
Tree-Structured Orthonormal Decomposition for Compositional Data
PolyILR introduces a canonical orthonormal decomposition of the Aitchison simplex aligned with arbitrary tree topologies, addressing a gap where existing methods ignore hierarchical structure, discard Aitchison geometry, or limit to binary trees. Each coordinate maps to a specific tree node, enablin…
Gradient-Enhanced Renewable Lasso for Distributed Online Estimation Under High Dimensions
This paper introduces a gradient-enhanced surrogate loss for online estimation in high-dimensional generalized linear models with streaming data. The method improves upon existing renewable estimation techniques by approximating cumulative loss with historical summaries, thereby removing batch-numbe…
TaskFusion: Continual Anomaly Detection for Tabular Data with Heterogeneous Feature Spaces
TaskFusion tackles the underexplored problem of continual anomaly detection on tabular data where feature schemas differ across tasks. It uses an Adaptive Growth-Fusion (AGF) model to map heterogeneous inputs into a shared space, augmented by Taskfusion augmentation and outlier-exposure training wit…
Dual Bayesian Affine Estimators for Nonlinear Parameter Learning in State-Space Models
This paper introduces a novel nonlinear parameter estimator for Wiener-type state-space models using a fixed-point architecture that integrates two affine MMSE estimators. Two frameworks, the dual basis-parameter and dual state-parameter estimators, are developed by employing distinct Dynamic Basis …
GenAIR: Generative Archetype-Grounded Item Representations for Enhanced Sequential Recommendation
GenAIR is a novel framework designed to improve sequential recommendation by generating enhanced item representations. It addresses limitations of existing methods by using LLMs to infer item 'Archetypes' representing target audiences and then calibrating these representations with actual user behav…
Data-Driven Dynamic Assortment with Unknown Preferences Achieves Rate-Optimal Learning
This paper introduces a novel data-driven algorithm for dynamic assortment problems on two-sided platforms where customer and seller preferences are initially unknown. The algorithm concurrently learns these latent parameters and optimizes the platform's revenue. Its performance, measured by regret,…
Itô Maps: Any-Step Stochastic Flow Maps for Enhanced SDE Integration and Control
This paper introduces the Itô map, a novel any-step stochastic flow map for integrating stochastic differential equations (SDEs) in a single pass. Unlike previous one-step generative models, the Itô map directly handles stochastic dynamics, enabling efficient posterior sampling and inference-time co…
Batched Robust iHMM for Outlier-Prone Streaming Data
This paper introduces Batched Robust iHMM (BR-iHMM), a novel approach for online infinite hidden Markov models that addresses data outliers and model misspecification. BR-iHMM utilizes generalized Bayesian inference and posterior influence functions to ensure robustness while offering tunable parame…
Synthetic Augmentation: Bias-Variance Trade-offs in Financial ML
Synthetic augmentation in financial machine learning, while addressing data scarcity, introduces a structural bias-variance trade-off. This trade-off is formalized as a modification of the effective training distribution. The core insight demonstrates that while synthetic data can reduce estimation …
Masked Autoencoders Show Untapped Potential for Downhole Metric Prediction
Predicting downhole metrics from surface drilling data is challenging due to limited labeled downhole measurements. While current approaches utilize neural networks like ANNs and LSTMs, Masked Autoencoder Foundation Models (MAEFMs) are underexplored despite their efficacy in time-series modeling. MA…
One-Shot Learning for Complex Nonlinear Oscillators
MEv-SINDy (Multi-frequency Evolutionary Sparse Identification of Nonlinear Dynamics) identifies global frequency-response curves from a single excitation time history by learning governing equations. It utilizes the Generalized Harmonic Balance (GHB) method to decompose complex forced responses into…
Adaptive Canonicalization for Continuous Equivariant Learning
The authors introduce adaptive canonicalization, a framework where the input's standard form is dynamically determined to maximize network predictive confidence. This approach eliminates the discontinuities associated with fixed canonicalization, ensuring continuity and universal approximation while…
Improved Regret Bounds for Contextual Bandits with Latent State Dynamics
This paper introduces a novel approach to contextual bandits where contexts and rewards are governed by a hidden Markov chain. Unlike previous methods, this work directly models dependencies on hidden states and achieves stronger, high-probability regret bounds. The key innovation lies in a fully ad…
Modeling Nonreciprocal Pairwise Comparisons for Improved Decision Analysis
This paper introduces an additive model for analyzing pairwise comparisons where nonreciprocity is not seen as a defect but as a combination of genuine scale variation and random perturbations. It allows for the estimation of noise levels, assessment of scale variation, and assignment of probabiliti…
Generative Path-Law Jump-Diffusion for Discontinuous Stochastic Trajectories
This paper introduces the Anticipatory Neural Jump-Diffusion (ANJD) flow, a novel generative framework for synthesizing forward-looking, càdlàg stochastic trajectories. The methodology addresses challenges in modeling structural breaks and non-autonomous dynamics by framing path synthesis as a seque…
Optimal Active Learning for Linear System Identification
This paper introduces an active learning algorithm for linear system identification that utilizes optimal centered noise excitation. The method, based on ordinary least squares and semidefinite programming, achieves minimal sample complexity. The authors establish tight lower and upper bounds for sa…