A benchmark for assessing large language models on molecular-to-food and food-to-molecular prediction tasks

Large Language Models (LLMs) have demonstrated remarkable proficiency in general-purpose tasks, yet their capacity for fine-grained reasoning in knowledge-intensive domains (KIDs) remains largely unexplored. This study addresses this gap by investigating LLM performance in the specialized field of food chemistry. We introduce a novel benchmark comprising two core tasks: Molecular-to-Food Prediction (MFP) and Food-to-Molecular Prediction (FMP). To support this benchmark, we curated and standardized the FlavorDB dataset, creating a robust testbed for flavor-molecular association. We evaluated two open-source models (Kimi-K2, DeepSeek-V3.2) and three closed-source models (Gemini-3-Pro, GPT-5.1, and Seed-1.8) under zero-shot and one-shot in-context learning settings. Our systematic analysis yields six key findings that characterize the capabilities and limitations of current LLMs in this domain. For instance, in the FMP task, Gemini-3-Pro achieved the highest zero-shot F1 score of 0.556, while Kimi-K2 led the one-shot setting with an F1 score of 0.522. In-context learning consistently improved performance across models, most notably boosting Kimi-K2's F1 score from 0.451 to 0.496 on complex multi-food tasks. Critically, we identify four recurring categories of domain-specific reasoning errors, which illuminate the fundamental challenges general-purpose models face when applied to fine-grained scientific inference in food chemistry. This work not only establishes a framework for rigorously evaluating LLM potential in knowledge-intensive domains but also provides a critical foundation for advancing practical applications, including flavor optimization and the development of novel food products.