Researchers present a self-evolving agent-harness framework that automatically improves LLM performance without modifying model weights. The system separates change proposal from credit assignment, using deterministic code for measurement and significance testing to ensure trustworthy improvements.
- Patches populate a gated, categorical quality-diversity archive keyed on the pathology addressed rather than the tasks fixed.
- Generalization is measured on a sealed test scored only after evolution.
- Across seven domains with a frozen open-weight model, credited gains range from +9 to +15.5pp.
- The approach retains 86-147% of training gain, demonstrating generalization rather than overfitting.
The winning patch tracks the model's dominant pathology, showing that the diagnose-and-credit loop transfers across different model families.