DiabAgent agentic framework for physiological control
Original title: DiabAgent: Toward Agentic Physiological Control
DiabAgent research suite includes APC framework, DiabetSim simulator and SweetBench benchmark showing improved glucose control over OpenAPS baseline.
DiabAgent provides the Agentic Physiological Control framework, DiabetSim simulator and SweetBench benchmark for agentic control of type 1 diabetes glucose levels. APC uses medical-device tools and ongoing observations to infer states, accumulate experience and choose treatment actions. Using Claude Opus 5.5 and one-hour decision interval APC achieves a COGI of 73.5 compared to 71.2 for OpenAPS. Ablations show individual adaptation and contextual information improve control.
- COGI of 73.5 for APC with Claude Opus 5.5
- OpenAPS baseline COGI 71.2
- SweetBench has 212 glucose-control tasks from 32 scenarios
- DiabetSim simulates patients and events
- combining high-frequency rule-based and low-frequency agent decisions balances performance and cost
From the original article
1 min read at arxiv.orgAs pretraining scales, medical language models have advanced rapidly in clinical question answering. Their ability to achieve effective physiological control through sustained intervention still lacks systematic quantitative evaluation.
By Chengxi Li · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.