Proposes a gradient-based algorithm matching the theoretical lower bound for multi-fidelity best-arm identification.

Riccardo Poiani, Rémy Degenne, Emilie Kaufmann et al.

DEFT leverages Doob's h-transform to fine-tune small networks for fast, high-quality conditional diffusion sampling, outperforming existing methods.

Alexander Denker, Francisco Vargas, Shreyas Padhy et al.

Using batch-reused SGD on two-layer neural networks enables learning low-dimensional polynomials near the information-theoretic limit, surpassing classical p-dependent complexity.

Jason D. Lee, Kazusato Oko, Taiji Suzuki et al.

Introduces Discrete Guidance (DG) leveraging continuous-time Markov chains for conditioned sampling in discrete spaces, outperforming existing models.

Hunter Nisonoff, Junhao Xiong, Stephan Allenspach et al.

Introducing Structured State Space Duality (SSD) framework links SSMs with attention, enabling 2-8× faster Mamba-2 with competitive language modeling performance.

Tri Dao, Albert Gu

Extends stochastic optimal control (SOC) to infinite-dimensional Hilbert spaces, deriving Doob’s h-transform for diffusion bridges, enabling resolution-free function space sampling.

Byoungwoo Park, Jungwon Choi, Sungbin Lim et al.

Proposes adaptive FTRL with stability, penalty, bias matching for Θ(T^{2/3}) regret, improving BOBW bounds across environments.

Taira Tsuchiya, Shinji Ito

BRAID leverages conservative reward modeling to fine-tune diffusion models, outperforming offline data with 15-20% reward gains while avoiding invalid designs.

Masatoshi Uehara, Yulai Zhao, Ehsan Hajiramezanali et al.

LatProtRL employs reinforcement learning in latent space to optimize protein fitness, outperforming baseline methods in key benchmarks.

Minji Lee, Luiz Felipe Vecchietti, Hyunkyu Jung et al.

Phased Consistency Models (PCM) outperform LCM in 1-16 step generation, applicable to high-res images and videos, enabling state-of-the-art few-step text-to-video synthesis.

Fu-Yun Wang, Zhaoyang Huang, Alexander William Bergman et al.

Using Abacus Embeddings, Transformers achieve 99% accuracy on 100-digit addition.

Sean McLeish, Arpit Bansal, Alex Stein et al.

Zamba is a 7B hybrid model combining Mamba backbone and shared attention, outperforming comparable transformers in speed and memory efficiency.

Paolo Glorioso, Quentin Anthony, Yury Tokpanov et al.

Proposes a verified safe reinforcement learning framework combining curriculum learning, incremental verification, and multi-controller strategies, extending safety horizon by 10x.

Junlin Wu, Huan Zhang, Yevgeniy Vorobeychik

AnalogCoder uses training-free LLMs to generate Python code for automatic analog circuit design, achieving over 83% success on 24 tasks.

Yao Lai, Sungyoung Lee, Guojin Chen et al.

Proposes a membership-query-based active learning framework for solving 0-1 integer programs with unknown knapsack constraints, improving query efficiency and solution quality.

Rosario Messana, Rui Chen, Andrea Lodi et al.

Proposes diffusion-based learning of multi-layer latent energy-based priors, improving sampling and hierarchical representation in generative models.

Jiali Cui, Tian Han

Cross-Layer Attention (CLA) shares KV heads between adjacent layers, reducing cache size by 2× with minimal accuracy loss.

William Brandon, Mayank Mishra, Aniruddha Nrusimha et al.

DIAMOND employs a diffusion-based world model achieving 1.46 human-normalized score on Atari 100k, surpassing previous methods.

Eloi Alonso, Adam Jelley, Vincent Micheli et al.

Proposes Scientific Generative Agent (SGA), integrating LLMs and differentiable simulation for physical scientific discovery, outperforming baselines with significant improvements.

Pingchuan Ma, Tsun-Hsuan Wang, Minghao Guo et al.

Proposes DML-IV, a double machine learning-based nonlinear IV regression method, reducing bias and enabling high-quality policy learning.

Daqian Shao, Ashkan Soleymani, Francesco Quinzan et al.