Papers200

Training big language models usually needs super-expensive, tightly connected GPU clusters, which most people do not have.

This paper builds a new test, called MURGAT, to check whether AI models can back up each small fact they say with the right part of a video, audio, or figure.

This paper introduces Causal-JEPA (C-JEPA), a world model that learns by hiding entire objects in its memory and forcing itself to predict them from other objects.

Voxtral Realtime is a speech-to-text model that types what you say almost instantly, while keeping accuracy close to the best offline systems.

Knowledge graphs are like giant fact maps, and keeping every fact correct is hard and important.

LiveMedBench is a new, always-updating test for medical AIs that keeps test questions safely separated from training data to avoid cheating by memorization.

Modern image editors can now follow visual prompts like arrows and scribbles, which opens a new way for attackers to hide harmful instructions inside images.

The paper asks a simple question: which kind of step-by-step reasoning helps small language models learn best, and why?

SceneSmith is a smart team of AI helpers that turns a short text like 'a cozy study with books and a desk' into a full 3D home scene you can drop right into a robot simulator.

WorldCompass teaches video world models to follow actions better and keep pictures pretty by using reinforcement learning after pretraining.

Robots often get confused by wordy instructions, so this paper tells them exactly where to touch instead of what to do in sentences.

This paper teaches a language model to improve its own math answers by first writing several drafts and then learning to beat its best draft.