Spotlight Others

Stable closed-loop control in latent space is achieved using a novel Coupled Oscillator Network, offering efficient model-based control for complex nonlinear systems directly from image data.

Softmax attention in transformers adapts its attention window to function Lipschitzness and noise, enabling efficient in-context learning.

MARINA-P and M3 algorithms drastically cut downlink and overall communication costs in nonconvex distributed optimization, scaling efficiently with the number of worker nodes.

Boost DNN robustness to corruptions without sacrificing clean image accuracy using Data Augmentation via Multiplicative Perturbations (DAMP)!

Humanoid robots now walk in San Francisco zero-shot, thanks to a novel ’next token prediction’ approach trained on diverse sensorimotor data, enabling real-world generalization and data efficiency.

Neural network learnability under the manifold hypothesis is hard except for efficiently sampleable manifolds.

This research presents novel adjoint methods for efficiently differentiating Lanczos and Arnoldi iterations, unlocking accurate gradients for large-matrix functions in machine learning.

GOPSA: a novel geodesic optimization method significantly improves cross-site age prediction from EEG data by jointly handling shifts in data and predictive variables.

GR2, a novel framework, extends generative retrieval to handle multi-graded relevance, addressing limitations of existing binary-relevance approaches by enhancing docid distinctness and implementing m…

PocketFlow: a novel generative model designs high-affinity protein pockets using prior-informed flow matching, outperforming existing methods.

GenArtist uses a multimodal large language model as an AI agent to unify image generation and editing, achieving state-of-the-art performance by decomposing complex tasks and leveraging a comprehensiv…

FuseFL achieves superior one-shot federated learning performance by leveraging a causal view of data heterogeneity and progressively fusing model blocks, significantly outperforming existing methods w…

FuseAnyPart: Swap facial parts seamlessly using multiple reference images via diffusion, achieving high-fidelity results.

Linear gated neural networks with fast, bounded units self-organize into modular weight structures and unique gating representations, enabling flexible task switching and compositional generalization.

Flex-MoE: A novel framework flexibly handles arbitrary modality combinations in multimodal learning, even with missing data, achieving robust performance.

User Sequence Imagination (USIM) revolutionizes out-of-vocabulary item recommendation by leveraging user sequence imagination and RL fine-tuning, achieving superior performance in real-world e-commerc…

Edge Pruning efficiently discovers sparse, yet accurate, computational subgraphs (circuits) in large language models via gradient-based edge pruning, advancing mechanistic interpretability research.

This paper introduces EP-η and EP-μ, novel EP variants remarkably robust to Monte Carlo noise, achieving improved speed and accuracy.

Frolic: A label-free framework boosts zero-shot vision model accuracy by learning prompt distributions and correcting label bias, achieving state-of-the-art performance across multiple datasets.

Vision-Augmented Prompting (VAP) boosts LLM reasoning by automatically generating images from textual problem descriptions, incorporating visual-spatial clues to significantly improve accuracy across …