Probability

Kyurae Kim, Samuel Gruffaz, Ji Won Park et al.

For researchers using Langevin Monte Carlo for sampling, this work extends theoretical guarantees to the overdamped regime, showing the exponential integrator remains stable and effective.

Shi Chen, Zhengjiang Lin, Kaizhao Liu et al.

Provides rigorous statistical guarantees for transformer performance as context length grows, addressing a key theoretical gap for practitioners scaling models.

Hugo Koubbi, Borjan Geshkovski, Philippe Rigollet

This work addresses representation collapse in transformers, a key issue for AI practitioners, but it is incremental as it builds on existing theoretical frameworks without introducing new methods.

Yunbum Kook, Santosh S. Vempala

This is an incremental survey that reviews an existing method with broad applications in areas like isoperimetric inequalities, optimization, and Markov chains, but does not introduce new results.

Mitia Duerinckx, Borjan Geshkovski, Stefano Rossi

It offers a mathematical foundation for a widely observed empirical phenomenon in large language models, addressing a key limitation in transformer-based architectures.

Giuseppe Bruno, Shi Chen, Zhengjiang Lin et al.

For theorists studying transformer scaling, this provides precise phase transition boundaries and limiting laws, but the analysis is restricted to i.i.d. keys and fixed queries, limiting direct applicability.

Eric Babson, Moon Duchin, Annina Iseli et al.

This work provides foundational theoretical tools for analyzing random MST, a widely used but mathematically understudied object, benefiting researchers in probability and combinatorial optimization.

Lifu Wei, Yinuo Ren, Naichen Shi et al.

It provides a computationally efficient and unbiased method for inference-time guidance in diffusion models, addressing the bottleneck of repeated score/gradient evaluations.

Paata Ivanisvili, Xinyuan Xie

For mathematicians, these are incremental improvements on known inequalities and bounds, with no broad impact beyond the specific problems.

Victor Kawasaki-Borruat, Clara Grotehans, Pierre Vandergheynst et al.

This provides a rigorous basis for manifold-aware sampling and generative modeling, addressing a fundamental challenge in high-dimensional data analysis.

Luca Avena, Gianmarco Bet, Bernardo Busoni

For researchers and practitioners relying on LLMs for reasoning tasks, this study reveals that current models lack robust probabilistic reasoning despite strong performance on standard benchmarks.

Molly Wang

This work addresses the critical problem of managing irreversible actions and associated risks for developers and users of recursive LLM agents, providing a framework for safer agent design.

Samuel N. Cohen, Filippo de Feo, Jackson Hebner et al.

This work addresses infinite-dimensional PDEs and optimal control problems, which are foundational in applied sciences like physics and stochastic systems, representing a novel paradigm rather than an incremental improvement.

Thomas L. Draper, Feras A. Saad

This work provides the first known space lower bounds for entropy-efficient random sampling, addressing a fundamental question in information theory and algorithm design.

Brice Huang, Mark Sellke

This work provides theoretical evidence for computational hardness in spin glass optimization, impacting physics and algorithm design, though it is incremental on prior conjectures.

Eric Li

This work advances the understanding of the Duke–Erdős–Rödl problem for extremal graph theory, providing tight bounds at the one-third threshold and clarifying the role of adjacent-edge conditions.

Yuzhou Gu, Mark Sellke

Resolves a long-standing conjecture in information theory, showing that Gaussian measures are not optimal for certain entropy properties.

Guy Bresler, Alina Harbuzova

This work establishes a hardness partial order for planted tensor models, which is incremental but provides formal reductions that could impact theoretical computer science and machine learning by linking conjectured-hard problems.

Fabiola Ricci, Claudia Merger, Sebastian Goldt

For researchers studying learning dynamics and generalization in neural networks, this work provides a theoretical and experimental framework linking Fourier properties of data to sample complexity and training speed.

Sourav Chatterjee, Persi Diaconis, Susan Holmes

For statisticians and machine learning researchers, this work provides a unified theoretical framework with finite-sample guarantees for set size estimation across diverse domains.