Evaluation-driven Scaling for Scientific Discovery

SimpleTES is a framework that scales test-time discovery loops with parallel exploration, feedback-driven refinement, and local selection.

Across 21 problems in six domains, it achieved state-of-the-art results, including faster LASSO, improved quantum circuit routing, and new Erdős overlap constructions.

The work shows that scaling evaluation-based search can materially improve AI-assisted scientific discovery.

It also produces trajectories that can be used to post-train downstream models.