Date of Award

6-11-2026

Document Type

Thesis

Publisher

Santa Clara : Santa Clara University, 2026

Department

Computer Science and Engineering

First Advisor

David Anastasiu

Abstract

AI generated food recipes are the second most desired AI innovation, however relatively few commercial systems exist to fill this niche. This is because while the solution is highly desired, cooking shouldn’t be approached with the same method by most LLMs and standard AIs. The leading models struggle to solve the issue of generating a semantically clear recipe while maintaining memory of the recipe status between steps. This is before we can even attempt to solve the flavor problem.

Cooking represents a unique challenge for AI, since it requires minimal loss in memory alongside already high quality text generation making this sub-domain of ML computing very computationally expensive. Rather than tackle the problem head on with improved data sets, we need to find a way to decrease the computational load for generating recipes. We attempt to solve this by proposing a new model: Bouillean.

Bouillean attempts to take text input of ingredients and generate recipes by approaching the problem like a chef. All recipes can be broken down into primary components: protein, veg, starch, sauce, etc. When a chef cooks, they combine different components to make one recipe, and all those components are cooked the same way regardless of the recipe; if a recipe calls for pan seared fish, that fish will be cooked the same way every time, they do not need to generate a new fish cooking method each time. With Bouillean, we took a large data set of existing recipes and trained Bouillean to identify and categorize the different components. This allows the model to focus on recipe composition and generation using reusable recipe structures.

To construct Bouillean, we used the Recipe1M dataset and fine-tuned Llama 3.1-8B to generate Component Recipes, Action Verbs, and Skeletal Recipes from existing recipes. These outputs were then used to train a two-stage recipe generation pipeline in which the first stage generated skeletal recipes and the second stage transformed them into readable cooking instructions. Evaluation was performed on 200 randomly selected recipes from the Recipe1M test partition and compared against AutoChef and LLaVA-Chef. Bouillean achieved an ingredient F-1 score of 0.0303. For instructions generation, it achieved BLEU-1 through BLEU-4 scores of 0.1398, 0.0639, 0.0379, and 0.0271, respectively, along with ROUGE-1, ROUGE-2, and ROUGE-L scores of 0.2099, 0.0359, and 0.1289. While Bouillean did not outperform all baseline models on automated evaluation metrics, the results demonstrate that a component-based approach can generate semantically coherent recipes and serves as a promising direction for future recipe generation systems.

Download

Share

COinS