Chain-of-Cooking: Culinary Process Modeling

• Chain-of-Cooking is a paradigm that comprehensively models culinary procedures using formal syntactic grammars and graph-based ontologies.
• It leverages multimodal techniques such as diffusion-based image synthesis and action-centric temporal graphs for precise process visualization.
• The framework integrates blockchain-based provenance to ensure transparent supply tracing and real-time verification of food safety.

Chain-of-Cooking refers to a paradigm for comprehensively modeling, generating, and tracing culinary procedures, encompassing compositional recipe grammar, procedural visual reasoning, provenance-aware supply chains, and step-wise semantic planning. It characterizes cooking as a temporally ordered, data-dependent chain of operations, each transforming ingredients via structured actions, and supports applications in procedural generation, process visualization, automated reasoning, and supply provenance.

1. Formal Syntactic and Semantic Foundations

At its symbolic core, Chain-of-Cooking is grounded in generative grammars and graph-based ontologies for culinary processes. Bagler (Bagler, 2022) formalizes cooking grammar as $G=(V,\,\Sigma,\,R,\,S)$, where:

• $V$: Non-terminals for hierarchical constructs (start symbol $S$, Ingredient Phrase $IP$, Processing Phrase $PP$, Processing Action $P$, Ingredient $I$, Quantity & Unit $Q$, Form $F$, Descriptor $T$, Utensil $V$0).
• $V$1: Terminals (ingredient names, units, forms, actions, etc.).
• $V$2: Production rules, e.g.,

$V$3

• Recursion enables nesting (e.g., a $V$4 where $V$5).

This yields an unambiguous derivation tree for each recipe, capturing dependencies, intermediate transformations, and temporal ordering. The grammar supports symbolic generation, parsing, optimization (e.g., nutritional constraints), and robotic compilation.

An alternative action-centric ontology (Kumbhakern et al., 4 Sep 2025) represents recipes as directed acyclic temporal graphs:

• Nodes: $V$6 (process, transfer, plating actions).
• Edges: Material flow ($V$7, $V$8) and temporal constraints ($V$9, $S$0).
• Environments: $S$1.
• Graph operators: Sequential ($S$2), parallel ($S$3), and merge ($S$4). Concurrency and compositional structure are native, reflecting complex real-world protocols (e.g., pan reuse in Full English Breakfast).

2. Procedural Image and Process Visualization Models

Recent advances in generative modeling have enabled explicit visual instantiation of Chain-of-Cooking as step-aligned procedural imagery.

CookingDiffusion Framework

CookingDiffusion (Wang et al., 15 Jan 2025) formalizes cooking-procedural image generation as follows:

• Input: Ordered recipe steps $S$5, each $S$6 is a textual instruction (with optional image $S$7).
• Output: Sequence $S$8, each $S$9 faithful to $IP$0 and sequentially consistent.
• Prompt scenarios:
  • Text-only: procedural prompts $IP$1
  • Image-only: procedural prompts $IP$2
  • Multi-modal: mixture of past $IP$3 and $IP$4 at ratio $IP$5.

Three "Memory Net" modules (Text Memory Net, Image Memory Net, Multi-Modal Memory Net) inject memory-conditioned embeddings into the Stable Diffusion UNet time-embedding stream:

• Memory Nets perform self-attention over historical text/image/mixed representations to yield procedural embeddings $IP$6.
• Injected via zero-initialized linear projections to $IP$7, thus

$IP$8

Evaluation uses FID and Average Procedure Consistency (APC):

$IP$9

with per-step $PP$0 aggregating CLIP-based image-text and text-text similarity over steps.

CoCook Process Visualization Model

The CoCook framework (Xu et al., 29 Jul 2025) augments latent diffusion with three modules:

• Dynamic Patch Selection (DPS): Retrieves visually relevant image patches from history, encoding via CLIP and cross-modal affinities for consistent texture and shape evolution.
• Semantic Evolution (SE): Integrates previous step textual embeddings as latent prompts with learnable cross-attention, propagating semantic context forward.
• Bidirectional Chain-of-Thought (CoT) Guidance: Synchronizes denoising trajectory across steps both forward and backward, enforcing global sequence coherence.

Metrics include FID, CLIP-T, CLIP-I, DreamSim, and human consistency scales. On CookViz and RecipeQA, CoCook achieves FID = 12.43/12.28 and CLIP-I = 0.739/0.820, outperforming SD and StoryGen baselines.

A critical finding is that patch-based reference (vs. frame-level) and semantic evolution are crucial for step-to-step consistency, as shown by ablation studies.

3. Provenance, Traceability, and Supply Modeling

Chain-of-Cooking, when linked to provenance, enables full traceability of food products through compositional manufacturing or cooking pipelines.

The token-recipes paradigm (Westerkamp et al., 2018) applies blockchain-based smart contracts for culinary supply chains:

• Each ingredient or intermediate product batch is a unique NFT—metadata includes quantity, origin, timestamps, and certificates.
• Recipes are codified as smart contracts:

$PP$1

where $PP$2 are input batch types, $PP$3 quantities, $PP$4 output batch type, $PP$5 quantity.

• Recipe application atomically burns/reduces inputs and mints output, storing parent (input) token IDs for full auditability.

Provenance queries are tree walks on lineage maps, as in the pizza example:

• Final pizza NFT's ancestors: crust NFT, sauce NFT, cheese NFT.
• Trace recursively to original flour, water, tomato, and milk sources.

Performance is gas-efficient: on-chain cost scales linearly with $PP$6 (number of unique inputs per recipe), and batch granularity can be tuned to minimize traceability overhead.

This infrastructure enables instant origin verification by end-users, robust safety recall (sub-minute audit), and transparent sustainability claims.

4. Sequential Reasoning and Recommendation Pipelines

Chain-of-Cooking also denotes step-wise semantic reasoning pipelines for complex, under-specified, or fuzzy-categorical user inputs.

The ChefMind system (Fu et al., 22 Sep 2025) demonstrates a hybrid architecture:

• Chain-of-Exploration (CoE): Query refinement via multi-level fuzzy detection and slot-filling. Fuzziness is defined as:

$PP$7

Progressive refinement yields $PP$8 over five retrieval levels.

• Slot normalization maps user expressions to canonical KG node IDs.
• Downstream modules:
  • Knowledge Graph (KG) retrieval with multi-hop path scoring ($PP$9).
  • Retrieval Augmented Generation (RAG) by vector similarity.
  • LLM integration for natural-language synthesis.

Empirical metrics (accuracy, relevance, completeness, clarity) are formalized; ChefMind attains 8.7 vs. 6.4–6.7 for ablations, and only 1.6% unprocessed queries. A plausible implication is that explicit constraint normalization and hybrid constraint-based/dense retrieval outperform purely generative or retrieval-only pathways for ambiguous culinary tasks.

5. Data Foundations and Evaluation Methodologies

Chain-of-Cooking models require robust, temporally annotated, and semantically paired multimodal datasets.

• The YouCookII dataset (Wang et al., 15 Jan 2025): Preprocessing extracts step-wise (text, image) pairs, removed corrupt videos, aligns CLIP-matched frames, and resizes to 256×256, forming $P$01,000 ordered training/evaluation sequences.
• CookViz (Xu et al., 29 Jul 2025): 40,362 intermediate step images, multi-step recipes, language normalization, watermark removal. Benchmarked alongside RecipeQA for both quantitative and human consistency metrics.
• Action-graph conversion (Kumbhakern et al., 4 Sep 2025) leverages LLM-driven simplification, standardization, and domain-adapted NER/relation extraction to map unstructured free-text into formal DSL/graph representations.

Evaluation spans quantitative (FID, CLIP similarity, DreamSim, graph coherence metrics), human annotation (consistency, coherency), and provenance walk latency (sub-10 min for recall).

6. Limitations and Future Directions

Explicitly noted limitations and open directions in current Chain-of-Cooking research are:

• Rare or out-of-distribution ingredients/methods may cause hallucinations or fallback to defaults (Wang et al., 15 Jan 2025).
• Removing too many steps undermines procedural specificity, particularly in multi-modal prompt regimes (Wang et al., 15 Jan 2025).
• Present focus is on still images; full video generation with temporal (frame-to-frame) dynamics remains an open challenge (Wang et al., 15 Jan 2025).
• Real-world blockchain systems must balance granularity (token batch size) against transaction cost, and fine-grained per-gram tracking incurs high on-chain state (Westerkamp et al., 2018).
• Parsing of complex free-text protocols into action-centric graphs requires robust simplification and controlled lexicon standardization; automated step recovery is a nontrivial information extraction task (Kumbhakern et al., 4 Sep 2025).

A plausible direction is cross-synthesis of symbolic, visual, and provenance-aware Chain-of-Cooking systems for unified AI-assisted culinary design, visualization, execution, and audit. Integration of compositional grammar with multimodal diffusion and on-chain provenance would enable fully auditable, semantically precise, and visually guided cooking workflows, scalable from home robots to global food supply.

7. Summary Table: Chain-of-Cooking Approaches

Approach	Formalism	Application Domain
Generative Grammar (Bagler, 2022)	$P$1	Symbolic recipe generation, optimization
StableDiffusion+MemoryNets (Wang et al., 15 Jan 2025)	Diffusion, Memory Nets	Procedural image synthesis, visual instruction
CoCook/CoT Guidance (Xu et al., 29 Jul 2025)	Diffusion+CoT	Step-aligned cook process visualization
Token-Recipes/Blockchain (Westerkamp et al., 2018)	Smart contract graph	Supply chain, provenance, food safety, audit
Action-Graph DSL (Kumbhakern et al., 4 Sep 2025)	Temporal action graph	Precise process modeling, automation, robot execution
ChefMind/CoE (Fu et al., 22 Sep 2025)	Slot/KG/RAG+LLM	Fuzzy intent resolution, recommendation pipelines

All these frameworks realize, in diverse modalities, the principle that cooking is a compositional, stepwise, temporally coherent chain, supporting programmatic generation, visual reasoning, provenance, and semantic interpretability.