Food is the cornerstone of human survival and social development. Food intelligence empowers various tasks—such as recipe generation, intelligent nutritional meal planning, dietary recommendations, and food-health correlation analysis—through the deep integration of perceptual computing and cognitive reasoning. This drives the digitalization of recipe cooking, the intelligent management of dietary practices, and the precision of nutritional services.
The Food Computing team at the Institute of Computing Technology, Chinese Academy of Sciences, in collaboration with Institute of Intelligent Computing Technology, Chinese Academy of Sciences, Suzhou and Zhongke Shenjian Technology Co., Ltd., Suzhou, has developed a domain-specific large language model for food intelligence—FoodSky. The study first integrated authoritative food data resources to construct a multi-source fused food corpus, FoodEarth, for training the FoodSky model. It then proposed a topic-selective state-space model and a hierarchical topic-aware retrieval-augmented generation algorithm, which can incorporate topic-relevant information and retrieving external knowledge bases for enhancing FoodSky’s ability to capture fine-grained food semantics and generate food-related text.
The model was evaluated on the Chinese National Chef Examination and Dietetic Examination, achieving the accuracy of 83.3% (chef exam) and 91.2% (dietitian exam), reaching expert-level performance. This enables FoodSky to provide professional guidance in culinary practices, dietary planning, and nutrition-health service.
The developed large language model, FoodSky, is expected to serve as an intelligent assistant in cooking, nutrition, and health, further advancing food intelligence.
This research was supported by the Beijing Natural Science Foundation (No. JQ24021) and the National Natural Science Foundation of China (No. 62125207). The findings were published in the Patterns (Cell Press).
downloadFile