SWE-EVO is a new test (benchmark) that checks if AI coding agents can upgrade real software projects over many steps, not just fix one small bug.
MatSpray turns 2D guesses about what materials look like (color, shininess, metal) into a clean 3D model you can relight realistically.
Flow Matching is like teaching arrows to push points from a simple cloud (source) to real pictures (target); most people start from a Gaussian cloud because it points equally in all directions.
SAM Audio is a new AI that can pull out exactly the sound you want from a noisy mix using text, clicks on a video, and time ranges—together or separately.
The paper proposes the Laws of Reasoning (LORE), simple rules that say how much a model should think and how accurate it can be as problems get harder.
RadarGen is a tool that learns to generate realistic car radar point clouds just from multiple camera views.
ReCo is a new way to edit videos just by telling the computer what to change with words, no extra masks needed.
InsertAnywhere is a two-stage system that lets you add a new object into any video so it looks like it was always there.
Visual grounding is when an AI finds the exact thing in a picture that a sentence is talking about, and this paper shows today’s big vision-language AIs are not as good at it as we thought.
3D-RE-GEN turns a single photo of a room into a full 3D scene with separate, textured objects and a usable background.
The paper introduces UCoder, a way to teach a code-generating AI to get better without using any outside datasets, not even unlabeled code.
The paper introduces Canon layers, tiny add-ons that let nearby words share information directly, like passing notes along a row of desks.