New Benchmark Reveals VLM Gaps in Nutritional Reasoning

The increasing integration of Large Vision-Language Models (VLMs) into critical infrastructure, particularly for personalized healthcare and dietary management, highlights a significant challenge: the "Systemic Information Asymmetry" between a food's visual appearance and its actual nutritional content. Existing benchmarks primarily focus on simple food classification, failing to assess the complex reasoning required for real-world dietary tasks. To address this, researchers introduced OmniFood-Bench, a comprehensive benchmark built from the MM-Food-100K dataset. This benchmark evaluates VLMs across three progressive capabilities: basic perception (identifying ingredients and cooking methods), quantitative reasoning (estimating portion size and nutritional profiling), and safety-critical advisory (providing disease-specific recommendations). Experiments with six state-of-the-art VLMs, including gpt-5.1 and gemini-3-flash, exposed a "Semantic-Physical Gap." Models showed near-human accuracy in naming dishes but performed poorly in mass estimation and frequently generated benign yet unsafe advice for high-risk conditions like diabetes. This work establishes a crucial standard for ensuring trustworthiness in autonomous agents used for public health.