LLMs vs Traditional Models for Recipe Nutrient Estimation
A study by CGU-ILALab evaluates methods for estimating nutrients from unstructured recipe text, comparing TF-IDF with Ridge Regression, DeBERTa-v3, and LLMs like Gemini 2.5 Flash. Under EU Regulation 1169/2011 tolerance criteria, TF-IDF offers moderate accuracy with near-instant inference, while DeBERTa-v3 performs poorly due to data scarcity. Few-shot LLM inference and a hybrid TF-IDF+LLM pipeline show promise.
Key facts
- Study compares TF-IDF, DeBERTa-v3, and LLMs for nutrient estimation.
- Uses EU Regulation 1169/2011 tolerance criteria.
- TF-IDF achieves moderate accuracy with fast inference.
- DeBERTa-v3 performs poorly under data scarcity.
- Few-shot LLM inference (Gemini 2.5 Flash) shows potential.
- Hybrid TF-IDF+LLM pipeline also evaluated.
- Published on arXiv with ID 2604.25774.
- Focus on dietary monitoring from unstructured recipe text.
Entities
Institutions
- CGU-ILALab
- arXiv