Main results of the paper: 22 agent / backbone configurations on the 64 BioXArena-Pro tasks across 8 domains. Click a task name for its description, or a score for the details of that run.
Every task is scored by the official evaluator and normalized to 0–1; a task without a valid score (timeout, no or invalid submission, agent failure) counts as 0, and any run longer than 2 h (7,200 s) is a failure even if a submission was written. Successful: tasks (out of 64) whose submission passed the task grader. Task Avg: mean normalized score over all 64 tasks, with failed tasks scored 0 (used for the ranking). Success-only Avg: mean score over the successful tasks only.
† Two task scores taken from MLEvolve with Opus 5, because the API returned empty completions for Opus 5.5 on those tasks. ‡ Run through a gateway for the native Gemini protocol on a separate machine of the same hardware class; scores are comparable, agent times are not.
Mean score over the 64 tasks (×100); yellow cells are the CLI means over its two backbones.
Mean score over the 64 tasks (×100); yellow cells are the scaffold means (rows) and backbone means (columns).