AI & Computer Science · AI Papers
Learning to Discover at Test Time (Mert Yuksekgonul et al.)
LISTEN TO THIS EPISODE

Mert Yuksekgonul and colleagues introduce TTT-Discover, which updates a language model while it works on a particular problem.
The method combines reinforcement learning with reuse of promising attempts, aiming to find one exceptional solution.
The paper reports experiments in mathematical optimization, GPU kernel engineering, algorithm competitions, and single-cell denoising.
Its comparisons examine what changes when learning modifies the model itself, alongside search over candidate solutions.
The discussion considers verifiable rewards, computational costs, evaluation limits, and the distinction between solving one problem and learning to solve new ones.
Read about the original source: Learning to Discover at Test Time
Part of Marginalia: books, papers, and ideas for learning and debate.