Sibyl-AutoResearch: Self-Evolving Trial-and-Error Harnesses Beat Paper-Generating AI Research Systems

arXiv·low signal

Sibyl-AutoResearch argues that autonomous research systems fail not because they can't execute experiments, but because they lose trial experience — weak evidence becomes prose, pilot signals become broad claims, and process failures don't change later behavior. The framework introduces a self-evolving harness that retains and learns from failed experiments. Directly relevant to anyone building autonomous research or self-improving agent pipelines.