Zhou Pengfei, Min Weiqing, Fu Chaoran, Jin Ying, Huang Mingyu, Li Xiangyang, Mei Shuhuan, Jiang Shuqiang
Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China.
University of Chinese Academy of Sciences, Beijing, China.
Patterns (N Y). 2025 Apr 22;6(5):101234. doi: 10.1016/j.patter.2025.101234. eCollection 2025 May 9.
Food is the cornerstone of both survival and social life. With the increasing complexity of global dietary cultures, there is a growing demand for food intelligence to enable tasks like recipe recommendations and diet-disease correlation discovery. To address this, we introduce the food-oriented large language model (LLM) FoodSky, which offers fine-grained perception and reasoning on food data. We constructed a food corpus, FoodEarth, from various authoritative sources to enhance FoodSky's knowledge. We also developed the topic-based selective state space model and hierarchical topic retrieval augmented generation algorithms to improve FoodSky's ability to capture fine-grained food semantics and generate context-aware food-relevant text. Extensive experiments show that FoodSky significantly outperforms general-purpose LLMs on the Chinese National Chef Examination and Dietetic Examination, achieving an accuracy of 83.3% and 91.2%, respectively. Beyond enhancing culinary creativity and promoting healthier eating patterns, FoodSky aims to establish a new benchmark for domain-specific LLMs in addressing real-world food-related challenges.
食物是生存和社会生活的基石。随着全球饮食文化日益复杂，对食物智能的需求不断增长，以实现食谱推荐和饮食与疾病关联发现等任务。为解决这一问题，我们引入了面向食物的大语言模型（LLM）FoodSky，它能对食物数据进行细粒度的感知和推理。我们从各种权威来源构建了一个食物语料库FoodEarth，以增强FoodSky的知识储备。我们还开发了基于主题的选择性状态空间模型和分层主题检索增强生成算法，以提高FoodSky捕捉细粒度食物语义并生成上下文感知的食物相关文本的能力。大量实验表明，FoodSky在中国国家厨师考试和营养师考试中显著优于通用大语言模型，准确率分别达到83.3%和91.2%。除了增强烹饪创造力和促进更健康的饮食模式外，FoodSky旨在为特定领域的大语言模型在应对现实世界中与食物相关的挑战方面树立新的标杆。
