Food4All: An Agentic Framework and Benchmark for Food Resource Navigation with Adaptive User Understanding

Food4All introduces a framework and benchmark for food resource navigation, integrating 686 structured food resources from Indiana. It evaluates six Large Language Models (LLMs) across 300 multi-turn tasks, focusing on requirement grounding and referral accuracy, with the top model achieving 96.33% accuracy but still exhibiting failures in handling constraints and user interaction challenges. This framework is significant for practitioners as it offers a controlled environment to assess and improve agent performance in real-world, constraint-sensitive food assistance scenarios.