ATHENA-R1: An AI Agent That Reasons Over Biomedical Tools

Treatment reasoning in medicine is inherently iterative: it requires identifying missing information, gathering evidence from multiple sources, evaluating contraindications and comorbidities, and revising conclusions as new data emerges. Most current AI systems struggle with this because they lack a structured way to seek and integrate evidence before forming a recommendation.

Researchers introduce ATHENA-R1, an AI agent trained to perform treatment reasoning across all FDA-approved drugs since 1939 by interacting with a universe of 212 real biomedical tools (covering indications, interactions, adverse events, disease phenotypes, and more). The agent learns to identify knowledge gaps, call the appropriate tools, incorporate retrieved evidence, and iterate until it reaches a grounded decision.

To train the agent without expensive human-annotated reasoning traces, the team developed a two-level self-learning framework: multi-agent systems first generate tools, tasks, and reasoning trajectories for supervised fine-tuning, followed by reinforcement learning with scientific feedback that rewards evidence gathering, valid tool use, logical coherence, and correctness.

ATHENA-R1 significantly outperforms frontier models like GPT-5 on both drug reasoning benchmarks and complex patient treatment cases. It also generates interpretable reasoning traces and has been positively evaluated by physicians and rare disease experts, with adverse event hypotheses validated in electronic health records from over 5 million patients.

This Ledger Entry expands how readers think about agentic AI in high-stakes domains by showing that treatment reasoning — which requires knowing what evidence to seek before forming a conclusion — can be effectively trained through self-generated trajectories and reinforcement learning over a universe of real biomedical tools, achieving strong performance, interpretability, and clinical validation without relying on human-annotated reasoning traces.