Learning to Discover at Test Time | Nebius Science Papers Club
Can an LLM learn from its own attempts while solving a problem and use that experience to discover a new state-of-the-art solution?
Instead of searching with a frozen LLM, TTT-Discover uses reinforcement learning at test time, allowing the model to learn from the problem it’s currently solving — reaching new state-of-the-art results in mathematics, GPU kernels, algorithm competitions, and biology, all with an open model.
Join Federico Bianchi, Staff Scientist at Together AI and co-author of the paper, for a 20-minute research talk followed by live Q&A and an open discussion.
Read the paper before the webinar →
About Nebius Science Papers Club
Nebius Science Papers Club is a webinar series led by Nebius researchers. Each session brings together paper authors and practitioners to discuss new ideas, research, and discoveries in AI.
The session is open to researchers, engineers, students, and anyone curious about the paper.
Register to get the recording
Try Nebius AI Cloud console today
