Role-Aware Multi-Agent Grading of Clinical Summaries
LENS
A role-aware multi-agent pipeline that grades clinical summaries using open-weight large language models.
Overview
A clinical summary is only useful if the reader can trust what was left out of it. LENS grades summaries with a pipeline of open-weight large language models, each agent reading the same text from a different professional role. Quality is not one number. What a summary has to carry depends on who is reading it and what they have to do next, so LENS scores the text through several role-aware graders rather than collapsing it into a single verdict.
Open-weight models keep the grading reproducible and let it run inside a hospital network where a hosted model is not an option. LENS complements the human judgment the lab has collected on summarization work, including a tool built for gathering physician ratings of summarized patient charts, and it applies to the lab’s own summarizers, among them a dual-stage system that condenses emergency department charts offline on embedded devices.