Recursive Self-Improvement

This direction centers on systems where the self-improvement loop itself is the object of study, covering metrics and oversight for self-improvement rates, closed-loop experiments in automated machine learning and AI scientists, evolutionary program search integrated with test-time learning, component-wise self-improvement, and constraints such as reward hacking, self-play collapse, data gating, and criteria for invalid runs. Work is valued for reproducible closed loops, public run logs, explicit reporting of invalid runs and failure rates, rather than only best-case results.