Megadose The AI curve, compressed.

The Resolution Hypothesis: Agent Evolution's Key Is Identifying Improvements

· ArXiv · AI/CL/LG ·
Selection performed no better than random at distinguishing helpful agent changes from harmful ones, the researchers report.

The study tested seven evolution methods across six models, using more than 10 million LLM calls. Useful prompts, skills and memories were proposed, but noisy evaluations let selection errors accumulate across rounds. At the method level, typical performance gaps of a few percentage points were smaller than differences between independent runs of the same method. The authors argue that progress depends on making real improvements reliably detectable and retainable.

ArXiv · AI/CL/LG's note

score 6

Categories: Research