AI scientists produce results without reasoning scientifically

Researchers evaluated LLM-based scientific agents across eight domains in over 25,000 runs, systematically analyzing their reasoning patterns. They found agents ignore evidence in 68% of traces and revise beliefs via refutation only 26% of the time—failures that persist even with high-quality reasoning examples. The base LLM accounts for 41.4% of outcome variance versus 1.5% for agent architecture, indicating current systems execute workflows but lack the self-correcting mechanisms defining scientific inquiry.