Title: CAREAgent: A Clinical Agent Leveraging Structured Reasoning and Integrated Tools for Order Generation

Abstract:

The translation of medical decisions into actionable, executable instructions is a vital function performed by clinical order generation, serving as the essential link between decision-making processes and practical application. However, current agent-based systems primarily concentrate on high-level decisions, frequently neglecting the detailed, executable specifics necessary for precise clinical orders. To bridge this significant gap, we present CAREAgent, a specialized agent designed for the generation of clinical orders.

To facilitate effective training, we developed a novel two-stage approach for constructing agentic reasoning data. Initially, we engineered an agent framework capable of generating verifiable reasoning trajectories that mirror the utilization of real-world clinical tools. Subsequently, these trajectories were refined through a filtering process based on three criteria: format compliance, order validity, and clinical plausibility.

Utilizing this curated dataset, the model underwent supervised fine-tuning to establish a foundation of medical knowledge and fundamental reasoning structures. This was followed by optimization via reinforcement learning, employing multi-dimensional reward functions to bolster complex clinical reasoning skills. Empirical evaluations across various benchmarks confirm the efficacy of CAREAgent. Notably, on the ClinicalBench benchmark—which was excluded from the training process—CAREAgent achieved F1 score improvements of 5.05%, 2.09%, and 0.86% compared to single-agent, multi-agent, and general agentic reasoning methods, respectively.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC