Topics

How do AI models really learn? A new framework maps their internal learning paths—revealing why some concepts emerge early and others spark sudden leaps in intelligence.

LLMs hallucinate confidently. Uncertainty estimation flags errors in NLP & code—strong on open-source models, misses nuances in GPT-4o. A plug-and-play step toward safer, self-aware AI.

The system analyzed two decades of comments and identifies accounts that appear designed to increase social division rather than support any single political side.

A new method reveals internal thinking steps from leading systems and raises questions about possible copying between models from different countries.

Researchers create an AI that reads studies, analyzes data, and predicts molecule behavior to speed up science.

New method manipulates internal concepts in AI systems for better outputs but also exposes potential security weaknesses.

Engineers at UC San Diego have created a better way to teach AI systems to handle tough problems mixing words and pictures, like math questions with charts.