AI News

BioDyad is an agentic biomedical machine-learning system designed to coordinate evidence discovery with executable program search. The paper identifies a feedback problem in existing systems: newly discovered evidence should shape candidate construction, execution results should guide later discovery and reuse, and validation must be matched to a limited search budget. BioDyad addresses this with two hierarchies embedded in Monte Carlo graph search. Its scientific hierarchy combines prior biomedical guidance with iterative discovery and stores links between biomedical plans and execution outcomes so that information can be reused across candidates. Its engineering hierarchy advances candidate programs from smoke execution to train/validation evaluation and then to full-data retraining. The authors evaluate the system on all 76 tasks in the BioXArena benchmark, allowing two hours per task and using three matched large-language-model backends. Against four agent methods and a one-shot baseline, BioDyad records the highest penalized all-task score and task success rate under every backend. The results support coordinated biomedical discovery and ML engineering as a way to turn external knowledge into executable programs across heterogeneous biomedical tasks.