Research map / Large language models / Language, safety and interpretability

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

ICLR 2025 · Large language models › Language, safety and interpretability · approach: Data and evaluation methodology

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces NutriBench, an 11,857-item human-verified natural-language meal description benchmark for carbohydrate estimation, evaluates 12 LLMs under four prompting strategies, compares with nutritionists, and runs a T1D glucose simulation.

Problem

LLM-based nutrition estimation from natural language meal descriptions lacks a public benchmark, and it is unclear how accurately LLMs estimate meal carbohydrates relative to professionals and how errors affect blood glucose outcomes.

Approach

A publicly released benchmark of natural-language meal descriptions generated by GPT-4o-mini from WWEIA and FAO/WHO GIFT dietary intake records (with human verification), used to evaluate LLMs with Base, CoT, RAG and RAG+CoT prompting, plus a nutritionist comparison and a Tidepool-based glucose simulation.

Key results

• GPT-4o with CoT achieved the highest accuracy of 66.82% (Acc@7.5) with a 99.16% answer rate across all models and methods.
• Averaged over 12 models, CoT raised Acc@7.5 from 43.24% (Base) to 47.46%.
• Fine-tuned Gemma2-27B-FT improved MAE from 13.33 to 9.57 and Acc@7.5 from 45.61% to 61.71% over its Base prompting, but still trailed GPT-4o CoT.

Measured effects reported in the paper

Surprising findings

• RAG did not consistently help: for GPT-4o-mini, RAG+CoT performed worse than CoT, and for natural serving queries CoT alone always beat RAG+CoT.
• Most models had higher error on natural serving descriptions than metric ones, while the GPT family did better with metric servings.

Limitations

• Meal descriptions were generated by GPT-4o-mini and verified by a single author, so the benchmark may carry generation artifacts.
• Nutritionist study used only three nutritionists on 72 queries, and the look-up condition had a single respondent.
• RAG database mostly contained metric 100g entries, limiting usefulness for natural serving queries.

Experimental setup

Datasets: What We Eat in America (WWEIA) FAO/WHO GIFT FoodData Central (FDC)

Benchmarks: NutriBench

Metrics: Acc@7.5 mean absolute error answer rate %TIR %TBR %TAR BGRI

Baselines: Base prompting Chain-of-Thought RAG Nutritionists

Builds on: chain-of-thought prompting retrieval-augmented generation qLoRA

Closest papers by meaning

• Pic2Prep: A Multimodal Conversational Agent for Cooking Assistance (AAAI 2025 · similarity 0.84)
• GlucoFM-Bench: Benchmarking Time-Series Foundation Models for Blood Glucose Forecasting (NeurIPS 2026 · similarity 0.82)
• What can large language models do for sustainable food? (ICML 2025 · similarity 0.81)
• OVFoodSeg: Elevating Open-Vocabulary Food Image Segmentation via Image-Informed Textual Representation (CVPR 2024 · similarity 0.81)
• Knowledge Graph and Large Language Model for Metabolomics (AAAI 2025 · similarity 0.81)
• Knowledge-Powered Recommendation for an Improved Diet Water Footprint (AAAI 2024 · similarity 0.81)