MedAI: Evaluating TxAgent's Therapeutic Agentic Reasoning in the NeurIPS CURE-Bench Competition

A team from the MedAI group evaluated the TxAgent therapeutic reasoning system in the NeurIPS CURE-Bench Competition, analyzing how retrieval quality for tool calls shapes model performance in high-stakes clinical tasks [1]. TxAgent is built on a fine-tuned Llama-3.1-8B model that dynamically generates and executes function calls to a unified biomedical tool suite called ToolUniverse, which integrates the FDA Drug API, OpenTargets, and Monarch resources to supply current therapeutic information [1]. The system uses iterative retrieval-augmented generation, an approach that provides greater accuracy and a wider scope of functions for prompt engineers [5]. In medical applications, the accuracy of both the reasoning trace and the sequence of tool invocations is critical because biased or inaccurate training data can make a large language model’s output less reliable [6]. The CURE-Bench NeurIPS 2025 Challenge benchmarks therapeutic-reasoning systems using metrics that assess correctness, tool utilization, and reasoning quality [1]. NeurIPS, held annually in December, is one of the three primary conferences of high impact in machine learning and artificial intelligence research, alongside ICLR and ICML [4]. The conference includes three days of invited talks and refereed paper presentations, followed by two days of workshops and competitions [4]. The MedAI team demonstrated performance gains through improved tool-retrieval strategies and was awarded the Excellence Award in Open Science [1]. The initial submission, dated December 12, 2025, measured 729 KB; a revised version, dated June 15, 2026, measured 731 KB [1]. The work addresses tasks such as drug recommendation, treatment planning, and adverse-effect prediction, all of which demand robust, multi-step reasoning grounded in reliable biomedical knowledge [1].