Hierarchical structure matters: representing recipes as nested sequences of structured steps, rather than flattened tables, lets models learn procedural dependencies and field interactions that improve performance on real-world synthesis and manufacturing tasks.
RecipeNet is a hierarchical Transformer model designed to learn from recipe data—ordered sequences of steps with structured fields—used in materials science, pharmaceuticals, and manufacturing. Unlike traditional tabular methods that flatten this data, RecipeNet captures both field interactions within steps and dependencies across steps, achieving better performance on recipe-based tasks.