The authors introduce OmniFood-Bench, a comprehensive benchmark constructed from the MM-Food-100K dataset to evaluate Vision-Language Models (VLMs) on nutrient reasoning and personalized health advice. Unlike previous benchmarks focusing on coarse-grained classification, this work assesses three progressive capabilities: basic perception of ingredients and cooking methods, quantitative reasoning for portion size and nutritional profiling, and safety-critical advisory for disease-specific recommendations.
- The benchmark evaluates six state-of-the-art VLMs, including gpt-5.1, gemini-3-flash, and qwen3-vl-8B.
- Experiments reveal a "Semantic-Physical Gap" where models achieve near-human accuracy in naming dishes but fail catastrophically in mass estimation.
- Models frequently hallucinate benign advice for high-risk diabetic profiles, highlighting significant trustworthiness issues in public health applications.
This work establishes a rigorous standard for the trustworthiness of autonomous agents deployed for public health by exposing these critical reasoning failures.