We are proud to announce that our paper:

“Large Language Models in Food and Nutrition Science: Opportunities, Challenges, and the Case of FoodyLLM”

has been officially accepted.

This work presents FoodyLLM, a domain-specialized large language model designed specifically for food and nutrition science. While general-purpose LLMs show strong language capabilities, they are not systematically trained on structured recipe–nutrition mappings or food ontologies — limiting their reliability in high-stakes nutrition tasks.

🌱 What FoodyLLM Enables

-- FoodyLLM was fine-tuned on 225,000 task-aligned QA pairs covering:

-- Recipe-based nutrient estimation

-- Traffic-light food labeling classification

-- Ontology-based food entity linking (supporting FAIR interoperability)

📊 Key Results

Across all tasks, FoodyLLM significantly outperforms strong general-purpose LLM baselines:

-- Nutrient estimation accuracy improved to 0.91–0.97

-- Traffic-light classification macro F1 increased to 0.86–0.97

-- Ontology-based entity linking macro F1 reached 0.93–0.98 on synthetic data and up to 0.84 on real corpora

These findings clearly demonstrate the importance of domain-specialized LLMs for trustworthy and scalable food & nutrition AI.

🔬 A Strong International Collaboration

This interdisciplinary effort connects researchers from the Jožef Stefan Institute, University of Ljubljana, Ss. Cyril and Methodius University (Skopje), Harvard Medical School, Northeastern University, and Boston University — combining expertise in AI, semantic technologies, nutrition science, and network medicine.

FoodyLLM represents an important milestone within our broader mission of advancing:

-- Trustworthy benchmarking

-- Task-aware LLM evaluation

-- Domain-adaptive AI systems

📄 Paper, 💻 code, and 🤖 model release coming soon.

Stay tuned for more updates.