OpenDiscoveryTrace Dataset of 558 AI Scientist Reasoning Trajectories
September 10, 2026
OpenDiscoveryTrace provides a public dataset of 558 scientific agent trajectories across drug discovery, materials science, and genomics. It records 9-field-per-step traces, including thoughts, tool calls, and error revisions, to enable auditing of reasoning processes in models like GPT-5.4 and Claude Opus 4.6.
HOW THIS AFFECTS YOU
●
builderYou can leverage these trajectories to build more transparent and auditable autonomous scientific agents.
●
researcherYou can use these structured traces to move beyond final-output evaluation and diagnose systematic reasoning failures.