TxAgent, a medical AI system built on a fine-tuned Llama-3.1-8B model, took an Excellence Award in Open Science at the NeurIPS 2025 CURE-Bench Challenge for therapeutic reasoning.
The system uses retrieval-augmented generation to query live sources - the FDA Drug API, OpenTargets, and a biological knowledge base called Monarch - before making recommendations on drug choices, treatment plans, or adverse effects. Rather than relying on knowledge baked into its weights, TxAgent generates function calls at inference time to a tool suite called ToolUniverse, pulling current data as needed. The CURE-Bench competition evaluated entries on three dimensions: answer correctness, tool utilization, and reasoning quality - meaning the chain of thought counted as much as the final output.
The paper's key finding is that tool-retrieval quality was the primary driver of overall performance, which matters more than it sounds. In medical settings, the dangerous failure mode isn't a wrong answer buried in a verbose reply - it's a plausible-looking drug interaction that gets acted on. Tying an AI's reasoning to authoritative, regularly updated sources like the FDA's own API is a more credible safety strategy than hoping training data was comprehensive and current. The open science framing is also pointed: medical AI that can't be audited is a liability, not an asset.
Benchmark performance and clinical reliability are different problems, though. CURE-Bench tests whether a model reasons well over curated scenarios - whether that holds up against the noise and edge cases of an actual clinical workflow is a question this paper leaves open.