NPC-RubricBench
A stage-aware benchmark for evaluating large language models' multi-stage clinical reasoning across the full course of nasopharyngeal carcinoma, based on longitudinal patient records and patient-specific atomic criteria.
Timeline (last 90 days)
NPC-RubricBench is a stage-aware benchmark for evaluating multi-stage clinical reasoning across the full course of nasopharyngeal carcinoma, based on longitudinal records of 405 patients from two cancer centers, covering 941 patient-stage units across six consecutive care stages and 17,688 patient-specific atomic criteria spanning five evaluation dimensions.
Among the 30 general-purpose and medical-specialized LLMs evaluated on NPC-RubricBench, high diagnostic performance does not automatically translate into better downstream care performance.
In the anatomical staging stage, the dominant failure mode is explicit decision errors, accounting for 48.2%.
In the initial workup stage and the post-treatment follow-up stage, the dominant failure mode is decision omission, accounting for 33.4% and 30.4%, respectively.
The methodology of NPC-RubricBench—patient-specific atomic criteria, decision-moment evidence grounding, and stage-aware evaluation—has the potential to be generalized to other tumor types and full-course chronic disease management scenarios.
All Facts (5)
NPC-RubricBench is a stage-aware benchmark for evaluating multi-stage clinical reasoning across the full course of nasopharyngeal carcinoma, based on longitudinal records of 405 patients from two cancer centers, covering 941 patient-stage units across six consecutive care stages and 17,688 patient-specific atomic criteria spanning five evaluation dimensions.
50%UnverifiedAmong the 30 general-purpose and medical-specialized LLMs evaluated on NPC-RubricBench, high diagnostic performance does not automatically translate into better downstream care performance.
50%UnverifiedIn the anatomical staging stage, the dominant failure mode is explicit decision errors, accounting for 48.2%.
50%UnverifiedIn the initial workup stage and the post-treatment follow-up stage, the dominant failure mode is decision omission, accounting for 33.4% and 30.4%, respectively.
50%UnverifiedThe methodology of NPC-RubricBench—patient-specific atomic criteria, decision-moment evidence grounding, and stage-aware evaluation—has the potential to be generalized to other tumor types and full-course chronic disease management scenarios.
50%