Model-Class Revision Policy Reaches 89.5% Exact Recovery in Scientific Discovery

A new discovery policy jointly proposes structural model edits and diagnostic experiments to overcome fixed hypothesis space limitations. In 400 controlled environments, the policy achieved 89.5% exact recovery, outperforming baselines by 10 percentage points using fewer experiments.