The study introduces the Evidence-Governed Evolution (EGE) framework, which separates candidate generation from confirmation and adoption of self-generated modifications by autonomous AI agents. The framework requires candidates to achieve a predefined primary objective in a fresh paired audit and to pass a reference-relative non-inferiority test that protects claimed capabilities before bounded canary deployment. The study also presents a calibration lower bound for adaptive candidates, a reference-replacement bridging test, and a cumulative degradation boundary; Monte Carlo stress tests demonstrate the calibration method only and do not evaluate deployed agents.
Source:Hugging Face 社区论文精选 · Read the original
Leave a Reply