Six-phase agentic workflow: 11 vs. 5 correct rare-disease diagnoses on 50 benchmark cases
A preprint extracted a six-phase workflow from seven rare-disease AI systems and ran it against a single GPT-5 prompt on 50 RareBench cases: 11 correct top-ranked diagnoses versus 5, with p=0.0703 falling short of the authors' 0.05 threshold.
A six-phase agentic workflow placed the correct rare-disease diagnosis first in 11 of 50 benchmark cases. The same GPT-5 with a single prompt managed 5.
A preprint published September 14 on Research Square reviewed seven AI diagnostic systems for rare diseases, selected from 1,193 publications after reading 111 full texts. Rare diseases are hard to recognize: symptoms combine differently across patients, and clinicians struggle to match a free-text case description against the full body of medical knowledge. Older tools required pre-coding complaints into a standard medical ontology; a standalone language model produced diagnoses from raw text with no further verification.
The authors found the same sequence in all seven systems: parse the case description, extract clinically significant features, generate candidate diagnoses, check against medical reference databases, reorder the list, record the rationale for each ranking. They ran this six-phase workflow and a single GPT-5 prompt on the same 50 cases from RareBench (a benchmark dataset of rare-disease descriptions) with identical parameters.
The 11 vs. 5 gap is noticeable, but the p-value came in at 0.0703 against the authors' threshold of 0.05, and 50 cases is too few for a firm conclusion. Across 19 evaluations on 33,738 cases from all seven systems, mean top-1 accuracy was 51.2%, ranging from 22.0% to 77.5%. That spread shows architecture and model choice affect results at least as much as the workflow itself. Still, the review turned the vague word "agentic" into an observable sequence of six steps with an explicit order of operations. Each step can be removed and its contribution measured individually. This is the first paired test on the same model where the workflow is described at that level of precision.
Related organization on Eternal Search:
Open the related Eternal Search page
Sources
[1] researchsquare.com
[2] t.me
Selected because the preprint delivers the first direct paired experiment testing an explicit multi-step agentic workflow against a baseline on the same model, turning 'agentic' into an observable and testable sequence. The statistical limits (p=0.0703, n=50) and the cross-system variance (22–77.5%) are both directly statable facts that resist hype without requiring invented caveats. The Experiment organization on Eternal Search is linked as a relevant destination for readers interested in the research-funding side of this space.