Epicure shows how a food model can move from “what pairs with this ingredient?” to a set of controls for exploring cuisine, chemistry, and recipe context.
Source note: Jakub Radzikowski, Josef Chen. “Epicure: Navigating the Emergent Geometry of Food Ingredient Embeddings.” arXiv:2605.22391, 2026-05-21. https://arxiv.org/abs/2605.22391
Why This Paper Matters
Most food recommendation systems behave like lookup tables with better math. You give them chicken, miso, basil, or rice, and they return nearby ingredients. That is useful, but it leaves out the question a cook actually asks: nearby in what sense?
Chicken can be close to garlic because they often appear together in recipes. It can be close to pork or beef because the ingredients share flavor chemistry or culinary function. It can also be steered toward a cuisine, a texture, a processing style, or a named pantry neighborhood. These are different questions. A good food AI system should expose them as controls rather than hide them inside one generic recommendation.
Epicure is a paper about that control layer. The authors build three related ingredient embeddings and show that the choice of training walks can become a product knob: one model emphasizes recipe co-occurrence, another emphasizes food chemistry, and a third blends the two. The result is less like a static recommender and more like a map with sliders.
The Idea in Plain English
An ingredient embedding puts foods into a mathematical space where similar things sit near each other. In a simple version, garlic and onion may sit near chicken because recipes often use them together. Basil and oregano may sit near each other because they share culinary roles and chemical signals.
Epicure builds three versions of that space from the same ingredients and nearly the same machinery. The difference is what the model is allowed to “walk through” while learning.
Cooc sees ingredient co-occurrence: what appears with what in recipes. Chem sees ingredient-compound relationships: which ingredients share flavor compounds. Core blends the two. Because the three models share the same vocabulary, architecture, and hyperparameters, their differences are easier to interpret. The paper asks: what changes when the map is trained to care more about recipe context or food chemistry?
The second idea is that the map is for more than nearest neighbors. Once concepts become directions, a user can rotate an ingredient toward South Asian cuisine, Mexican pantry ingredients, sweet baking, or a discovered cluster of aromatics. The paper uses SLERP, a spherical interpolation method, to make that steering continuous. At a small angle, the result stays close to the seed ingredient. At a larger angle, the target concept takes over.
What the Researchers Tested
The authors assembled 4.14 million recipes from 11 public sources across several language and cuisine contexts. The corpus is dominated by English RecipeNLG and Chinese XiaChuFang, with smaller Russian, Vietnamese, Spanish, Turkish, Indian-English, Indonesian, and German sources.
They then normalized roughly 200,000 raw ingredient strings into 1,790 canonical ingredients. This step matters because recipe data is messy: the same food can appear under many spellings, preparations, brands, or translations. The paper uses an LLM-assisted canonicalization pipeline, embedding-based clustering, and manual curation to create the final ingredient vocabulary.
From there, the authors build two kinds of graph signal:
- A recipe co-occurrence graph with 203,508 positive-NPMI ingredient-ingredient edges.
- A FlavorDB chemistry graph with 80,019 typed ingredient-compound edges, 2,247 typed compound nodes, and 15 flavor categories.
They train three Metapath2Vec models with 300-dimensional embeddings. The models are siblings, not unrelated experiments. They differ in the walk schema:
- Cooc walks only the ingredient co-occurrence graph.
- Core walks the typed compound graph while injecting repeated ingredient-ingredient walks.
- Chem walks the typed compound graph without pure ingredient-ingredient templates.
The evaluation has three layers. First, the paper checks whether labeled concepts become recoverable as linear directions: sensory categories, basic tastes, USDA nutrient measures, and eight cuisine macro-regions. Second, it checks the geometry itself, including isotropy and food-group/cuisine structure. Third, it looks for emergent factors and modes, then tests whether those discovered modes can be used for navigation.
What They Found
The important finding is not that one model wins everywhere. It is that the three models answer different culinary questions, and those differences can be exposed.
Chemistry Sharpens Directions
All three models recover culinary concepts linearly, but Chem performs best across nearly every probe. The reported ordering is Cooc < Core < Chem across baked-in compound features, held-out basic tastes, USDA nutrients, and cuisine macro-regions. For cuisine separability, mean Cohen’s d is 2.43 for Cooc, 2.70 for Core, and 3.07 for Chem.
This is interesting because many of the probes aren’t simple repeats of chemistry labels. The paper argues that compound-mediated walks act as a structural prior: routing through shared aroma compounds helps organize broader culinary concepts too.
Recipe Context and Chemistry Produce Different Neighbors
For chicken, Cooc retrieves recipe companions such as garlic, onion, black pepper, turkey, and carrot. Core and Chem move toward protein and flavor-profile neighborhoods. For basil, Cooc leans toward pantry context like olive oil, parmesan cheese, black pepper, and white wine, while Core and Chem retrieve Italian-herb chemistry neighbors such as oregano, tarragon, rosemary, and pasta.
That difference is exactly what a chef or product builder would want to choose. Sometimes the user wants “what do I cook with this?” Sometimes they want a substitute or flavor cousin.
The Embeddings Contain Named Culinary Neighborhoods
The unsupervised analysis recovers 20 stable factors per model and 150 to 200 named culinary modes per model. These modes are more than arbitrary clusters. Their mean coherence sits well above random-pair baselines: 0.611 vs. 0.097 for Cooc, 0.833 vs. 0.348 for Core, and 0.703 vs. 0.115 for Chem.
The examples read like useful culinary neighborhoods: sweet baking and dessert ingredients, South Asian whole spice blends, Mexican and Latin American pantry ingredients, East or Southeast Asian aromatics. That gives the model something closer to a navigable atlas.
Steering Works as a Continuous Control
The paper uses SLERP to rotate an ingredient toward a target direction. Rice rotated toward South Asian cuisine retrieves ingredients like curry leaf, masoor dal, urad dal, chana dal, and fenugreek seed. Corn rotated toward Latin American cuisine retrieves salsa verde, tomatillo, queso fresco, fajita seasoning, and corn tortilla.
The angle matters. At 0 degrees, the query is basically the original seed. Around 30 degrees, the target starts to influence the neighborhood. Around 60 degrees, the result is dominated by the target concept. This turns food recommendation into controlled exploration.
Why It Happens
The mechanism is the walk schema. The graph is more than a data structure; it decides what the model repeatedly sees as context.
Cooc learns from ingredients appearing together in recipes, so it tends to preserve practical cooking companions. Chem learns through compound-mediated paths, so it sharpens chemistry-driven and flavor-profile relationships. Core repeats ingredient-ingredient walks while also using compound paths, which creates a more concentrated geometry. The paper reports a participation ratio of 94.2 for Core, compared with 173.6 for Cooc and 183.1 for Chem. Core is not simply broken or collapsed; its concentration is a consequence of the training design.
The bigger lesson is that “embedding bias” does not have to be hidden. If the system designer knows which signal dominates a model, the interface can let users choose it.
What This Means for Builders
For builders of food, recipe, restaurant, grocery, and menu-planning tools, Epicure points to a better interface pattern.
Do not ship one opaque recommendation list and call it personalization. Let the user choose the question:
- Co-occurrence companion: what is commonly cooked with this?
- Flavor-profile peer: what shares chemistry or sensory character?
- Cuisine steering: what happens if this ingredient moves toward South Asian, Japanese, Mediterranean, or Latin American space?
- Mode lookup: what named culinary neighborhood does this ingredient belong to?
- Exploration distance: how far should the recommendation move from the original seed?
This is also useful outside consumer recipe apps. A restaurant R&D team could use these controls to test menu variants. A grocery product team could generate substitution or basket-expansion ideas. A food manufacturer could explore flavor extensions while keeping a seed ingredient recognizable.
The strongest product idea is not “AI suggests five ingredients.” It is an interactive culinary map where model choice, target direction, and travel distance are visible controls.
What This Means for Buyers and Operators
For operators evaluating food AI systems, the paper suggests a simple procurement question: can the vendor explain what kind of similarity the model is using?
A recommendation that comes from recipe co-occurrence is not the same as a recommendation that comes from chemistry. One may be better for meal planning. The other may be better for substitution, flavor matching, or novel product development. If a vendor cannot separate those behaviors, users may get plausible but poorly framed suggestions.
The caveats also matter operationally. Corpus imbalance can make some cuisines better represented than others. FlavorDB coverage means many ingredients receive chemistry signal indirectly rather than directly. LLM-assisted canonicalization means data cleaning choices become part of the model’s behavior. A buyer does not need every detail, but they should know whether the system can inspect, log, and correct those layers.
The practical bar is explainable control, not simply fluent recipe generation.
What to Watch Next
The next step is an interface study. The paper proposes a chef-facing tool that exposes model choice, closest-mode lookup, and SLERP angle in one place. That would test whether cooks actually use these controls, which controls they understand, and whether the outputs improve real menu or recipe work.
The second thing to watch is whether the three siblings become a continuous family. Right now, Cooc, Core, and Chem are three fixed points on a chemistry-vs-recipe-context spectrum. A future model could expose that mix as a smooth parameter.
The third thing to watch is cross-modal grounding. If the same canonical ingredient vocabulary connects embeddings, recipes, images, sensory descriptors, and nutrition data, the steering operator could move across more than ingredient names. A user might steer from ingredient to recipe text, dish image, menu item, or flavor description.
Limitations and Caveats
The paper is careful about several limits.
First, the corpus is imbalanced. English RecipeNLG and Chinese XiaChuFang dominate the data, while several other cuisine regions have much smaller support. The authors argue that the model ranking remains stable, but resolution inside smaller regions is still limited.
Second, chemistry coverage is uneven. Only 523 of the 1,790 canonical ingredients retain active ingredient-compound edges after filtering. The remaining non-hub ingredients still participate, but their chemistry signal is indirect.
Third, the pipeline depends on LLMs for canonicalization, cuisine tagging, and mode labeling. The embeddings themselves are trained over graph walks rather than LLM judgments, but the node vocabulary and labels are shaped by LLM-assisted preprocessing.
Fourth, the paper says code and trained artifacts are not released at this time. That limits independent reproduction and makes this a promising research artifact rather than a production-ready public benchmark.
Source
Jakub Radzikowski, Josef Chen. (2026). Epicure: Navigating the Emergent Geometry of Food Ingredient Embeddings. arXiv preprint arXiv:2605.22391. Available at: https://arxiv.org/abs/2605.22391