Papers1262

Diffusion models make great images and videos but are slow because they usually need many tiny steps.

Long chains of thought make AI smarter but also slower, pricier, and limited by memory windows.

Ex-Omni is a new open-source AI system that can understand text or speech and then talk back while moving a 3D face in sync with the voice.

AIRS-Bench is a new test suite that checks whether AI research agents can do real machine learning research from start to finish, not just answer questions.

NanoQuant is a new way to shrink large language models down to 1-bit and even less than 1-bit per weight without retraining on huge datasets.

PlanViz is a new test that checks whether AI image models can plan and draw useful computer-related pictures like routes on maps, work flowcharts, and website screens.

This paper teaches small, local AI models to write deep, insightful research reports by letting writing and planning work together instead of staying separate.

QuantaAlpha is a smart, evolving system that helps find trading signals (called alpha factors) even when markets are noisy and keep changing.

RelayGen is a training-free way to switch between a big model and a small model while one answer is being generated.

Text-to-image models using GRPO used to give the same final reward to every step, which is like giving the whole team the same grade no matter who did what.

This paper teaches a computer to find buttons, text, and icons on screens so it can click and type in the right places, a skill called GUI grounding.

LatentChem lets AI do chemistry thinking quietly inside continuous vectors instead of writing long step-by-step sentences.