This paper presents a study that applies state probing techniques to evaluate the ability of large-scale language models (LLMs) to understand the cooking process. LLMs are trained on vast amounts of procedural text, but because they lack direct observation of real-world phenomena, they struggle to accurately understand intermediate steps in cooking recipes. To address this, we construct a novel Japanese recipe dataset with clear and accurate annotations of ingredient state changes. Based on this dataset, we present three new tasks to evaluate the ability of LLMs to track ingredient state changes during the cooking process and identify ingredients present at intermediate steps. Experimental results using widely used LLMs, such as Llama3.1-70B and Qwen2.5-72B, demonstrate that learning ingredient state knowledge improves understanding of the cooking process, achieving performance comparable to that of commercial LLMs. The dataset is publicly available on Hugging Face.
