May 27QuarkThe Big Model of HealthScored above the passing line on the 12-doctor National Associate Physician Title Exam, becoming the first large model in the country to successfully cross this threshold.
It is reported that the Quark Health Grand Model, based on Tongyi Qianqian, achieved this breakthrough through massive high-quality data construction and multi-stage post-training strategies.
The new big modeling capabilities can already be invoked directly through Quark Search. Users will find that for serious medical problems Quark will dynamically retrieve books, guides, drug instructions, medical dictionary papers, etc. by analyzing them first and then searching for them. This high search quotient strategy significantly improves the accuracy of complex cases.
The Associate Director title exam assessment covered 12 commonly used disciplines, including Family Medicine, General Internal Medicine, General Surgery, Obstetrics and Gynecology, and so on. In all of the above disciplines, Quark Health Big Model exceeded the passing line and reached the passing line for chief physicians in 4 disciplines, including general medicine and medical oncology.
On the entry-level and intermediate title exams, the smaller Quark Health Large model has a maximum lead of about 7 and 10 points compared to the full-blooded Basic model. Quark's lead extends to a maximum of 30 points when entering the significantly more difficult associate-level exam, which emphasizes comprehensive clinical use. This study validates the great potential of vertical models for performance improvement.
When facing the two types of questions with the "highest misjudgment rate", namely multiple-choice questions and case analysis questions, Quark achieves 711 TP3T in the multiple-choice domain with the help of the "Medical Long Thinking" mechanism; in case analysis questions, Quark improves the correctness rate to 531 TP3T with the combination of the retrieval augmentation and the stepwise inference strategy.