Introduces 'Tuned Lens'—layer-wise affine transformations—improving intermediate decoding accuracy and interpretability in transformers.

Nora Belrose, Igor Ostrovsky, Lev McKinney et al.

Negative-entropy FTRL achieves three-world adaptation: adversarial O(√dT logT log|D|T) and corrupted-stochastic O(d logT/Δmin plus corruption terms).

Fang Kong, Canzhe Zhao, Shuai Li

Magnushammer employs contrastive learning with Transformer to improve premise retrieval, achieving 59.5% success on PISA with 4x fewer parameters.

Maciej Mikuła, Szymon Tworkowski, Szymon Antoniak et al.

TRACT method enhances single-step diffusion model quality using Transitive Closure Time-Distillation, achieving FID of 7.4.

David Berthelot, Arnaud Autef, Jierui Lin et al.

Proposes S5 model with parallel reset for long sequences in RL, outperforming RNNs and Transformers in speed and accuracy.

Chris Lu, Yannick Schroecker, Albert Gu et al.

Proposes PaLM-E, a multimodal large language model integrating visual and sensor data for robotics and visual QA.

Danny Driess, Fei Xia, Mehdi S. M. Sajjadi et al.

Consistency models enable one-step high-quality sampling, surpassing diffusion models with FID 3.55 (CIFAR-10) and 6.20 (ImageNet 64x64).

Yang Song, Prafulla Dhariwal, Mark Chen et al.

This study evaluates climate XAI methods on robustness, faithfulness, and other properties, comparing algorithms like Integrated Gradients across models.

Philine Bommer, Marlene Kretschmer, Anna Hedström et al.

Proposed MV-DTSA framework maps time series to binary images, enhancing forecasting accuracy.

Luoxiao Yang, Xinqi Fan, Zijun Zhang

FPT freezes GPT-2 blocks for seven time-series tasks, reaching 0.516 average long-horizon MSE and 74.0% classification accuracy.

Tian Zhou, PeiSong Niu, Xue Wang et al.

Introduces the 'leap' measure to quantify hierarchical complexity, analyzing SGD training time on low-dimensional functions with neural networks.

Emmanuel Abbe, Enric Boix-Adsera, Theodor Misiakiewicz

Hyena, a subquadratic operator combining implicit long convolution and gating, achieves Transformer-level performance on long sequences with 20% less compute.

Michael Poli, Stefano Massaroli, Eric Nguyen et al.

Unifies multicalibration via game dynamics, reducing complexity and improving guarantees in multi-objective learning.

Nika Haghtalab, Michael I. Jordan, Eric Zhao

Proposed a general FTRL/OMD algorithm achieving best performance in adversarial and stochastic settings for linear bandits.

Christoph Dann, Chen-Yu Wei, Julian Zimmert

FLE algorithm uses MLE with probabilistic models to estimate return distributions in offline RL, providing error guarantees.

Runzhe Wu, Masatoshi Uehara, Wen Sun

Immunizing images with imperceptible adversarial perturbations prevents diffusion models from realistic editing.

Hadi Salman, Alaa Khaddaj, Guillaume Leclerc et al.

Riemannian Flow Matching (RFM) enables simulation-free, scalable generative modeling on complex manifolds using spectral distances, achieving state-of-the-art results.

Ricky T. Q. Chen, Yaron Lipman

Proposes decentralized entropic optimal transport (DEOT) with theoretical guarantees, suitable for distributed distribution comparison.

Xiangfeng Wang, Hongteng Xu, Moyi Yang

Applying Byte Pair Encoding (BPE) to symbolic music reduces sequence length by over 50%, enlarges vocabulary, and improves model performance and speed.

Nathan Fradet, Nicolas Gutowski, Fabien Chhel et al.

Solving inverse physics problems using score matching, achieving high accuracy and temporal stability.

Benjamin J. Holzschuh, Simona Vegetti, Nils Thuerey