DiabAgent: A Research Suite for Agentic Physiological Control in Type 1 Diabetes Management

Forum topic · 小凯 · 2026-11-03

Summary

DiabAgent is a research suite for agentic physiological control introduced by researchers on arXiv (2611.03601). It targets Type 1 diabetes, where impaired insulin secretion requires continuous, personalized insulin dosing based on meals, exercise, illness, and glucose measurements. The suite comprises three components: the Agentic Physiological Control (APC) framework, which manages glucose through medical-device tools by inferring individual states, accumulating personal experience, and selecting treatment actions; DiabetSim, a simulator covering diverse patients, events, and device conditions for scalable virtual testing; and SweetBench, a benchmark instantiating 32 scenario designs into 212 glucose-control tasks. Using Claude Opus 5.5 with hourly decision intervals, APC achieved a COGI score of 73.5, surpassing the open-source closed-loop baseline OpenAPS at 71.2. Ablation studies show that individual adaptation and contextual information improve control, and combining high-frequency rule-based control with low-frequency agent decisions offers a promising trade-off between performance and compute cost.

DiabAgent: Toward Agentic Physiological Control

Field: NLP / Medical AI Authors: Chengxi Li, Shaoxuan Li, Deyao Hong, Qingle Liu, Yi Huang, Puheng Bian, Minzhi Xu, Lanyue Zhang, Mingze Sun, Xin Li, Calvin Xiao, Shengyuan Wang, Huan-ang Gao, Jiazheng Xu, Kaisen Yang, Hongning Wang Published: 2026-11-02 arXiv: 2611.03601

Abstract

As 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. Type 1 diabetes provides a concrete setting: impaired insulin secretion creates dependence on exogenous insulin, with dosing continually adjusted using information about meals, exercise, illness, and glucose measurements.

The authors introduce DiabAgent, a research suite for agentic physiological control comprising three components:

APC (Agentic Physiological Control) framework — controls glucose through medical-device tools, using ongoing observations to infer individual states, accumulate personal experience, and choose treatment actions.
DiabetSim — a simulator of diverse patients, events, and device conditions, enabling scalable virtual testing and training.
SweetBench — a benchmark built on DiabetSim that instantiates 32 scenario design instances into 212 glucose control tasks.

Key Results

Using Claude Opus 5.5 with a fixed one-hour decision interval, APC achieves a COGI score of 73.5, compared to 71.2 for the open-source closed-loop baseline OpenAPS.
Ablation studies show that individual adaptation and contextual information improve glucose control.
Combining high-frequency rule-based control with low-frequency agent decisions shows potential to balance control performance against computational cost.

Significance

The work provides a framework, simulator, and evaluation foundation for studying agentic participation in human physiological control — moving medical LLM evaluation beyond question answering toward sustained, closed-loop intervention.

---

*Source: arXiv:2611.03601, cross-posted from zhichai.net forum.*