10M-call study: agent self-improvement gains sit in noise
TL;DR
- Across 10M+ LLM calls, seven evolution methods, and six models, selection performs no better than random at distinguishing helpful artifacts from harmful ones.
- Typical performance margins between evolution methods fall within the variance between independent runs of the same method.
- The authors argue progress depends on better evaluation, not better proposers, citing 'making true improvements detectable and accumulable.'
The selection step at the heart of agent-evolution loops is no better than random at distinguishing helpful candidate artifacts from harmful ones. That finding comes from a 38-page arxiv preprint that ran more than 10 million LLM calls across seven evolution methods and six language models.
The authors state it plainly: 'agent evolution stalls not because methods fail to propose better artifacts, but because noisy evaluations cannot reliably identify them.'
At the method level, the paper reports that typical performance differences are smaller than the variance between independent runs of the same method. The gap between two evolution techniques is often indistinguishable from the gap you get by rerunning one of them.
The abstract names the artifacts in play (prompts, skills, memories). The arxiv listing does not say which seven methods or six models the team benchmarked. The authors' takeaway: shift the field from proposing improvements to 'making true improvements detectable and accumulable.'
Originally reported by paper
Read the original article →Original headline: 10M-Call Study: Typical Agent Self-Improvement Margins Are Smaller Than Run-to-Run Noise