We propose Vittle, a new visual instruction tuning framework that improves robustnessof MLLMs to data distribution shifts by pursuing the minimal sufficient representation.

We introduce ChA-MAEViT - a robust, high-performing model that encourages cross-channel interactions for multi-channel imaging.

We propose GLSim, a training-free framework that combines global and local embedding similarity signals for accurate object hallucination detection in LVLMs, outperforming prior methods.

We propose a general approach for modifying image-based models to improve robustness and temporal consistency on video.

We unify attention heads and task vectors in the mechanistic analysis of in-context learning.

We provide a refined analysis on the excess risk of kernel ridge regression in structured non-i.i.d. setting and apply it to denoising score learning.

This paper answers the question: "under what conditions will model-free reinforcement learning give rise to thinking as a strategy for reward maximization?"

We introduce Weaver, a framework that combines multiple weak verifiers to effectively select responses in repeated sampling, achieving frontier model accuracy without supervised fine-tuning, while reducing verification costs by 99.97%.

We incentivize truthful data sharing using a novel technique inspired by two sample testing.

We show that linear layers, the primary player in large language models, can be composed from fundamental algebraic primitives with exponentially less parameters.

We provide a precise description on how the test risk scales with a suitable norm-based capacity when compared to a classical metric of model size.

We design the first efficient algorithm that solves the general agnostic ReLU regression with optimal query complexity

This paper proposes a new single-loop first-order algorithm for solving linearly constrained bilevel optimizationm problem.

We derive novel matrix factorizations, leading to an efficient streaming algorithm for privately estimating partial sums even when the size of the input is a priori unknown.

This paper gives a pure dp algorithm for all the pairs min cut problem with the same error as private min-st-cut.

We provide the first convergence guarantees and overparameterization bounds for alternating gradient methods in a special class of nonconvex-nonconcave zero-sum games where both the players are shallow neural networks.

We study replicable distribution testing.

MxDs show that dense layers are more faithfully represented by mixtures of specialized sublayers than by sparsely activating neurons, while remaining just as interpretable.

We show that transformers achieve length generalization when training on shorter main task and longer auxiliary tasks together.

The first large-scale multimodal foundation model for astronomy integrating a large number of diverse observational modalities.

We propose the first efficient, training-free online routing algorithm for high-volume LLM serving under token budget constraints, achieving significant improvements in both routing performance and cost efficiency.