LLM agent personalizes meal plans to regulate blood glucose

Personalized glucose regulation is a long-standing challenge in precision nutrition because postprandial glucose responses vary widely between individuals. Traditional glycemic index methods fail to capture this heterogeneity and can't dynamically adjust meals based on personal physiological feedback. To address this, a team led by Mingyu Huang and Weiqing Min from the Chinese Academy of Sciences proposed a novel system called the physio-feedback agentic loop, accepted at ACL 2026 Findings. Their approach unifies individualized absorption modeling with dietary intervention, using an LLM-based agent for context-aware reasoning and iterative refinement of real-world meals.

The system has two core components. First, a Physiology-Aware Glucose Predictor uses a learnable Temporal Physiological Absorption Decay Module to model how each person's body absorbs nutrients over time, rather than relying on population averages. Second, a Prediction-Driven Two-Stage Meal Optimization Agent iteratively adjusts meals based on predicted glucose outcomes as explicit feedback. In experiments on multiple public datasets, the method not only improved prediction accuracy but also effectively reduced glucose excursions, meaning smaller post-meal blood sugar spikes. The authors say this is the first time physiological learning has been integrated with an LLM agent for glucose regulation. The work could enable more adaptive nutrition apps and insulin dosing support, though clinical validation would be needed for real-world deployment.