Google method curbs agent test memorization and boosts generalization
Google Cloud AI Research and university teams developed Regularized Recursive Self-Improvement to stop self-improving agents from memorizing test tasks. RRSI limits edit budgets, tracks past attempts, and uses a critic to block task-specific hacks so agents generalize beyond training. Tests showed up to fourteen point training gains and smaller but consistent improvements on unseen benchmarks.
Where this came from. One outlet reported this. We read it, then wrote the sixty words above.