Signal RecipeNet: A Hierarchical Transformer for Recipe Data

Summary

A preprint proposes RecipeNet, a hierarchical Transformer architecture for what it calls recipe data, ordered sequences of steps with heterogeneous structured fields that appear in domains such as materials synthesis, pharmaceutical formulation and industrial manufacturing. The model encodes field-level interactions within each step together with sequential dependencies across steps using stacked Transformer encoders. Existing tabular models typically flatten this structure into a fixed schema, failing to capture such hierarchical relationships. The authors report that RecipeNet consistently outperforms existing tabular models across multiple recipe datasets and tasks. This highlights the value of hierarchical and sequential modeling for representing procedural data.