This paper proves that Transformers trained via gradient descent cannot efficiently learn majority Boolean functions, with error growing exponentially with input dimension.

Bo Chen, Zhenmei Shi, Zhao Song et al.

ffstruc2vec employs a flat, flexible framework for structural node embeddings, enhancing scalability and interpretability in large graphs.

Mario Heidrich, Jeffrey Heidemann, Rüdiger Buchkremer et al.

Proposed RIM manifold optimization reduces complexity from O(n^3) to O(n), excelling in tasks like image denoising.

Jinghui Yuan, Fangyuan Xie, Feiping Nie et al.

Reinforcement learning optimizes language models for inference time objectives, improving pass@k and voting accuracy by reducing gradient variance.

Yunhao Tang, Kunhao Zheng, Gabriel Synnaeve et al.

OmniLearn dynamically scales worker batches, cutting heterogeneous-cluster training time by 14–85% and improving asynchronous accuracy by up to 6.9%.

Sahil Tyagi, Prateek Sharma

ScalingNoise improves video generation quality and consistency by guiding noise selection.

Haolin Yang, Feilong Tang, Ming Hu et al.

Proposes collaborative self-play (CSP) framework with Reinforced Self-Training (ReST) to improve tool use and uncertainty calibration in language models.

Jacob Eisenstein, Reza Aghajani, Adam Fisch et al.

Týr-Pruner employs end-to-end global sparsity distribution optimization via supernet construction and evolutionary search, retaining 97% performance with 50% parameter removal.

Guanchen Li, Yixing Xu, Zeping Li et al.

Proposes sublinear algorithms for Wasserstein and TV distances from streaming data, enabling scalable fairness and privacy auditing.

Debabrota Basu, Debarshi Chanda

ADiT unifies molecule and material generation via a shared latent diffusion framework, achieving SOTA results with high speed.

Chaitanya K. Joshi, Xiang Fu, Yi-Lun Liao et al.

This study evaluates transformer models for molecular property prediction, showing domain adaptation with physicochemical tasks significantly outperforms larger data scaling.

Afnan Sultan, Max Rausch-Dupont, Shahrukh Khan et al.

CoPL enhances LLM personalization using graph-based collaborative filtering and LoRA experts, improving preference estimation.

Youngbin Choi, Seunghyuk Cho, Minjong Lee et al.

EPPO leverages evidential uncertainty to adapt quickly in non-stationary environments, outperforming state-of-the-art PPO variants in continuous control tasks.

Abdullah Akgül, Gulcin Baykal, Manuel Haußmann et al.

Q♯ algorithm uses distributional RL with KL regularization to optimize large language models post-training, achieving superior performance.

Jin Peng Zhou, Kaiwen Wang, Jonathan Chang et al.

SpargeAttn introduces a training-free universal sparse attention method, accelerating inference by 2.5-5x across language, image, and video models without performance loss.

Jintao Zhang, Chendong Xiang, Haofeng Huang et al.

Gradient-based method uncovers multi-dimensional refusal cones in LLMs, revealing independent mechanisms governing refusal behavior.

Tom Wollschläger, Jannes Elstner, Simon Geisler et al.

ONPO uses optimistic mirror descent for general-preference alignment, achieving O(1/T) gap and AlpacaEval scores up to 48.6.

Yuheng Zhang, Dian Yu, Tao Ge et al.

This paper introduces PLDM, a latent dynamics planning method that leverages reward-free offline trajectories, outperforming model-free RL in generalization and trajectory stitching.

Vlad Sobal, Wancong Zhang, Kyunghyun Cho et al.

Llamba uses cross-architecture distillation from Llama-3.x, achieving 2-3x inference throughput with only 0.1% training data, suitable for edge devices.

Aviv Bick, Tobias Katsch, Nimit Sohoni et al.

Ensemble perspective on SNN; membrane potential smoothing + guidance boosts accuracy to 83.2% on CIFAR10-DVS with 4 timesteps.

Yongqi Ding, Lin Zuo, Mengmeng Jing et al.