Why AI Models ‘Think’ Longer: New Theory Explains Reasoning BreakthroughNew theory by CDS PhD student Nikolaos Tsilivis, CDS’ Julia Kempe, & others links AI models’ longer answers to better reasoning…18h ago18h ago
Supercharged Information Synthesis: CDS Alum Teaches AI Models What Information Actually MattersCDS PhD alum Vishakh Padmakumar tackled summarization’s hardest challenge: choosing what matters.5d ago5d ago
Lightweight AI Monitors Can Catch 93% of Attacks That Break Safety Rules into Benign-Looking StepsCDS researchers developed a lightweight AI safety monitor that blocks 93% of decomposition attacks with 90% lower cost.Nov 5Nov 5
Pinpointing Where AI Models Hide Their Concepts: From Safety to Dogs to Mathematical ReasoningNew research from CDS reveals the specific attention heads where language models store concepts like safety and reasoning.Oct 29Oct 29
Asking AI in Multiple Languages Unlocks More Diverse PerspectivesPrompting AI in multiple languages helps it reflect a broader range of perspectives.Oct 24Oct 24
Meet the Fellow: Jian QianJian Qian joins CDS as a Faculty Fellow & Courant Instructor, bringing research in machine learning theory and interactive decision-making.Oct 22Oct 22
When AI Learns What Makes an Image Probable, Simple Beats Complex by 10¹⁴⁰⁰⁰Natural images vary in probability by up to 10¹⁴⁰⁰⁰, CDS researchers find, challenging assumptions about how images are structured.Oct 17Oct 17
Most Chatbots Miss Half the World’s Values, Study FindsCDS PhD student Lily Zhang and Meta collaborators created a new dataset to help build LMs that better reflect global human values.Oct 15Oct 15
How LLMs Could Transform the Study of Human GroupsCDS Faculty Fellow Ilia Sucholutsky and collaborators outline how LLMs could reshape how we study group cognition at societal scales.Oct 10Oct 10
CDS’ Grace Lindsay Launches YouTube Channel on “AI for the Planet”Grace Lindsay’s new YouTube series turns climate-AI research into five-minute videos for a wide audience.Oct 8Oct 8