TTT-Discover: Learning to Discover at Test Time

TTT-Discover introduces a test-time reinforcement learning framework designed to solve verifiable scientific discovery problems, such as optimizing GPU kernels or mathematical bounds. Unlike standard reinforcement learning that targets average reward, this approach optimizes for maximum reward on a single, specific problem, treating the model policy as a disposable tool to reach a breakthrough artifact. By utilizing an entropic objective and a modified search strategy that balances exploration of unvisited states with exploitation of high-performing lineages, the system achieves significant performance improvements, including kernels up to 50% faster than existing human-designed solutions. This method enables models to adapt to out-of-distribution tasks without requiring extensive retraining, demonstrating that test-time learning effectively pushes the frontier of discovery in domains where even incremental improvements carry substantial weight.