A new AI system doesn't just answer medical questions. It decides what to ask next.

Most AI tools used in medicine today answer a specific question a clinician puts to them. A system described in Nature Communications this month, called DxDirector-7B, is built to do something different: manage the diagnostic process itself, deciding which questions to ask and which tests to request, largely on its own [s1].

What makes it different

Real-world diagnosis rarely starts with a clean summary of symptoms — it starts with an ambiguous complaint that requires iterative reasoning and follow-up testing to narrow down [s1]. The researchers behind DxDirector-7B designed it to navigate that entire process autonomously, determining what they call optimal diagnostic strategies and asking for a physician to step in only when a specific clinical operation actually requires one [s1]. That's a meaningful departure from existing AI medical assistants, which typically respond to queries a clinician has already framed rather than driving the workup themselves [s1].

How it performed against other systems

The researchers tested DxDirector-7B against several benchmarks. On RareArena, a set of 22,901 rare-disease cases, it scored 36.23% accuracy, ahead of OpenAI's o3-mini (32.96%), Deepseek-R1-671B (32.07%), and the larger MedFound-176B model (23.98%) [s1]. On a set of 344 complex cases drawn from the New England Journal of Medicine's clinicopathologic case series, DxDirector-7B scored 38.4% — ahead of human physicians in the comparison at 32.5%, and ahead of GPT-4o (30.8%) and Deepseek-R1-671B (30.72%) [s1]. On ClinicalBench, covering 1,500 real-world cases, it scored 63.46%, well ahead of Deepseek-R1-671B (47.28%) and MedFound-176B (35.76%) [s1]. On a 1,273-case set drawn from USMLE-style questions across 12 clinical tasks, it scored 50.88% against MedFound-176B's 39.03% [s1].

Notably, DxDirector-7B posted these results with a comparatively small 7-billion-parameter model, outperforming several "state-of-the-art medical and general-purpose LLMs with significantly larger parameters," in the researchers' own framing [s1].

Fewer questions, not just better answers

The paper's other headline claim concerns efficiency rather than accuracy. Across the datasets tested, DxDirector-7B required roughly three clinical operations — tests, questions, or other diagnostic steps — to reach its conclusions, compared with four to eight or more for baseline LLMs performing the same task [s1]. The researchers report the system achieved 97–98% efficiency in requesting only operations it determined were actually necessary [s1]. In a real-world test at a top-tier Chinese hospital covering 160 cases across nine departments, the system's diagnostic conclusions matched what physicians ultimately determined at substitution rates as high as 75% in cardiology and 60–66.7% in gastroenterology and pulmonology [s1].

Where it falls short

The paper is notably direct about the system's limits. Its authors state that DxDirector-7B "performs worse than Deepseek-V3 and Deepseek-R1 on basic science integration," a gap they attribute to its smaller parameter count relative to those larger models [s1]. It also underperforms in departments like dermatology, plastic surgery, and psychiatry — specialties that depend heavily on direct patient contact and physical or visual examination rather than the kind of structured reasoning over test results the system is built around [s1]. The authors are explicit that human oversight remains mandatory for high-risk or emergent conditions, and describe the system as maintaining "a robust safety and accountability framework" specifically because of that limitation [s1].

What readers should take from this

This is a research paper describing a system's performance on defined benchmarks and one hospital's retrospective case set — not a deployed clinical tool, and not something cleared by any regulator for patient use. The comparisons to "human physicians" are drawn from a specific published case series (the NEJM clinicopathologic conferences), not a live head-to-head trial against practicing doctors managing real patients in real time. The substitution-rate figures from the hospital test describe how often the AI's conclusion matched what physicians separately determined — a measure of agreement, not independent proof the AI was correct in cases where physicians might also have been wrong. The authors' own limitations, particularly around basic science reasoning and specialties requiring physical examination, are a useful corrective to reading these accuracy numbers as a blanket claim of AI besting doctors at diagnosis generally.

Sources

• DxDirector: an agentic large language model driving the full-process clinical diagnosis — Nature Communications, 2026-04-23