We are pleased to announce that our work, “FoodBench-QA: Overview of the Shared Task on Grounded Food and Nutrition Question Answering”, has been accepted for presentation at the CL4Health Workshop at LREC 2026.

FoodBench-QA introduces a comprehensive benchmark for evidence-based question answering in the food and nutrition domain, requiring systems to ground their predictions in food composition databases and structured domain ontologies. The shared task is designed to reflect real-world challenges where accuracy, interpretability, and regulatory alignment are essential.

The benchmark is structured around three core tasks:

-- Nutrient estimation from recipe ingredients, evaluated in accordance with EU Regulation 1169/2011

-- FSA traffic-light classification for fat, salt, sugars, and saturates

-- Food named entity recognition and linking to the Hansard Taxonomy, FoodOn, and SNOMED CT ontologies

The shared task attracted participation from five international teams, demonstrating growing interest in grounded AI approaches for food science and nutrition.

The best-performing systems achieved strong results across tasks, including:

-- Up to 93.57% accuracy for protein estimation

-- Macro F1 scores up to 0.90 for traffic-light classification

-- Entity linking performance ranging from 60.71% to 80.89% on real-world text, and up to 95.75% on synthetic datasets, depending on the ontology

These results highlight both the progress and remaining challenges in developing trustworthy, explainable, and regulation-aware AI systems for food and nutrition applications. The findings further emphasize the importance of robust benchmarking and adaptive model selection, aligning with the broader vision of AutoLLMSelect and AutoLearn-SI to advance transparent and reliable AI systems.

We thank all participating teams and collaborators for their valuable contributions.