NutriBench: A Dataset for Evaluating Large Language Models on Nutrition Estimation from Meal Descriptions

TL;DR

NutriBench is the first public dataset for evaluating nutrition estimation from natural language meal descriptions, derived from WWEIA and FAO/WHO GIFT across 11 countries and annotated with macronutrients and calories. The paper systematically benchmarks twelve LLMs using Base, CoT, RAG, and RAG+CoT prompts, showing GPT-4o with CoT achieves the top accuracy (66.82%) and near-perfect answer rate, often outperforming professional nutritionists in speed. The authors introduce Retri-DB for knowledge-grounded prompting, demonstrate that CoT improves performance especially on complex meals, and reveal cultural and dietary variability in model errors. They also show that fine-tuning with FoodData Central data and real-world risk simulations for diabetes further underscore the practical value and limitations of LLM-based nutrition estimation for professionals and laypersons alike.

Abstract

Accurate nutrition estimation helps people make informed dietary choices and is essential in the prevention of serious health complications. We present NutriBench, the first publicly available natural language meal description nutrition benchmark. NutriBench consists of 11,857 meal descriptions generated from real-world global dietary intake data. The data is human-verified and annotated with macro-nutrient labels, including carbohydrates, proteins, fats, and calories. We conduct an extensive evaluation of NutriBench on the task of carbohydrate estimation, testing twelve leading Large Language Models (LLMs), including GPT-4o, Llama3.1, Qwen2, Gemma2, and OpenBioLLM models, using standard, Chain-of-Thought and Retrieval-Augmented Generation strategies. Additionally, we present a study involving professional nutritionists, finding that LLMs can provide comparable but significantly faster estimates. Finally, we perform a real-world risk assessment by simulating the effect of carbohydrate predictions on the blood glucose levels of individuals with diabetes. Our work highlights the opportunities and challenges of using LLMs for nutrition estimation, demonstrating their potential to aid professionals and laypersons and improve health outcomes. Our benchmark is publicly available at: https://mehak126.github.io/nutribench.html

Figures (14)

• Figure 1: GPT-4o-mini answers a query from NutriBench using different prompting strategies.
• Figure 1: Unique meals and food items sampled from each country from FAO/WHO GIFT
• Figure 2: Accuracy and Answer Rate (defined in Section \ref{['sec:experiments']}) for the evaluated LLMs across different prompting strategies on NutriBench.
• Figure 3: The countries from which meals were extracted from the FAO/WHO GIFT. The indices on the map correspond to the countries in Table \ref{['tab:who_fao']}.
• Figure 4: Difference in MAE between natural and metric serving descriptions. Natural servings are preferred by most models (orange), whereas the GPT family prefers metric (blue).