Source-linked AI summary
Do Recipes Have Personas? Characterizing and Generating Creator Style in Attributed Procedural Graphs
Lei Jiang
TL;DR
The paper addresses how to recover and generate individual creators’ procedural styles when LLMs produce semantically sound but homogenized workflows. It introduces a creator-mapped recipe-graph dataset and combines stylometric graph analysis with persona-conditioned generation. Results show complementary strengths between few-shot semantic reasoning and structured topological control, motivating an ensemble, while the framework remains limited by its culinary scope and unresolved fusion needs.
Problem
The paper asks how creator-specific procedural personas can be identified and how complete execution graphs can be generated for unseen dishes despite lexical overfitting and homogenized LLM workflows.
Method
The paper introduces VRT, represents recipes as procedurally aligned graphs, analyzes lexical and topological stylometry, and evaluates persona-conditioned graph generation with normalized physical-state edges.
Results
Few-shot LLMs show semantic assignment strengths, while the structured two-stage model provides stronger macro-planning and topological control; both achieve competitive operational fidelity.
Takeaways & Limitations
Procedural persona can be represented as a creator’s physical execution-graph footprint, with ensemble generation combining global semantic reasoning and localized topology.
Takeaways & Limitations
The framework is built in the culinary domain, and future work proposes neuro-symbolic fusion and post-hoc calibration to validate constraints and repair violations.
Abstract
from arXiv · showhide
While large language models (LLMs) possess vast zero-shot procedural knowledge, their tendency to produce homogenized logic often obscures the unique, idiosyncratic execution processes of individual human creators. In this paper, we investigate the computational discovery of procedural personas from unstructured data. To achieve this, we introduce ViralRecipesTrans, a new dataset of procedurally aligned execution flow graphs extracted from popular culinary video transcripts and explicitly mapped to specific creators. We formulate procedural stylometry as a graph learning and process discovery task, revealing a fundamental duality: while traditional lexical classifiers overfit via semantic leakage, discrete topological metrics successfully capture the rigid physical constraints of a creator's workflow. Building upon this characterization, we extend our framework into a novel generative task--predicting a creator's exact structural execution graph for unseen dishes. We expose a fundamental dichotomy in style generation between global macro-planning and local structural execution. Our results demonstrate that few-shot LLMs dominate semantic assignment but suffer from persistent macro-planning deficits, whereas our structured two-stage model achieves superior topological control via rigid Markovian priors. Together, an ensemble approach to procedural generation combines the strengths from both sides, dynamically synthesizing global semantic reasoning with localized topological footprints to automate the discovery and generation of personalized workflows.
1 Introduction
The paper reframes recipe understanding as procedural stylometry: creators differ not only in language but in the topology of executable workflows. It introduces VRT and studies both persona identification and creator-conditioned graph generation, emphasizing a trade-off between global planning and local execution.
- Motivation: LLMs can parse bounded culinary knowledge, but their homogenized outputs obscure creators’ idiosyncratic workflow structures.The paper locates the distinctive signal of human authorship in how tasks are physically approached, parallelized, and paced.
- Motivation: Recipes are executable directed acyclic graphs whose physical and chemical constraints permit substantial creator-specific structural variation.This combination of text and topology makes recipes an environment for intra-class stylometry.
- Contributions: ViralRecipesTrans provides procedurally aligned execution DAGs extracted from viral culinary-video transcripts and mapped to individual creators.The dataset supports investigation of stylistic fingerprints in both linguistic expression and workflow structure.
- Research questions: The paper asks whether topology can overcome lexical overfitting and whether persona conditioning can predict complete execution graphs for unseen dishes.The research questions explicitly contrast textual signatures with structural topology and include task ordering and structural ellipsis.
- Research questions: A central generative trade-off separates global semantic macro-planning from local Markovian structural execution, motivating an adaptive ensemble.The ensemble is proposed to synthesize these complementary forces for operational fidelity.
2 Related Work
Prior work moved from lexical authorship features toward structural signatures, while persona modeling and retrieval-augmented generation expanded procedural generation capabilities. However, enforcing specific topological structures during generation remains challenging.
- Stylometry and persona modeling: Traditional computational stylometry uses N-grams and TF-IDF, but these features can overfit vocabulary in constrained domains.Recent work therefore explores structural signatures based on linguistic dependency trees to improve cross-topic robustness.
- Procedural generation: RAG-based LLMs excel at procedural extraction, yet existing controllable-generation frameworks still face persistent difficulty enforcing valid physical topologies.The related work situates this challenge alongside neural logic constraints for structurally controlled generation.
3 The ViralRecipesTrans (VRT) Dataset: Construction and Statistics
VRT is built by discovering and filtering culinary video accounts, collecting qualified transcripts, converting them into execution graphs, and characterizing their lexical and structural diversity. The resulting corpus covers 97 accounts and more than 5,000 videos.
- Dataset scope: VRT maps culinary-video transcripts to execution DAGs and contrasts real-world creator variation with standardized procedural recipe corpora.The authors acknowledge a limited pool of influencers with qualified English transcripts and video formats.
- Account discovery and filtering: An automated YouTube Data API pipeline uses layered cooking queries followed by filters for data density and procedural diversity.The filtering process targets volume, format, repertoire diversity, transcript quality, and valid procedural graphs.
- Account discovery and filtering: Accounts require at least 10,000 subscribers, 50 uploaded videos, 30% recent long-form uploads, and 30 unique dishes.Three sampled videos must also provide sufficient English transcripts and compile into valid procedural graphs.
- Corpus scale: 97 unique accounts and more than 5,000 videos span niche home cooks with 14K subscribers to major channels with 7.7M subscribers.Three accounts overlap between the long- and short-format groups.
- Transcript processing: For each qualifying account, the pipeline retrieves up to 50 recipe videos per format, ranked by total views, and quality-checks transcripts through action-graph construction.Transcripts are retained only when graphs contain at least two cooking actions and two distinct ingredients.
- Descriptive statistics: The corpus contains 570 unique cooking verbs; its five most frequent actions comprise 37% of action tokens, while 47% of per-account actions are hapax legomena.These statistics indicate both a shared core vocabulary and substantial creator-specific rarity.
4 Detecting and Characterizing Recipe Personas
Procedural DAG topology captures medium-level workflow differences and creator-specific structural signals more robustly than lexical stylometry alone. Across archetype and scalability analyses, topology separates formats and preserves a stable persona-ranking signal as creator classes increase.
- Archetype-Level Structural Classification: Topological projection maps recipe DAGs to discrete structural metrics that classify three procedural archetypes, achieving Macro F1 0.783.The projection removes semantic features and uses metrics including depth-to-width ratio, branch rate, and maximum layer depth.
- Archetype-Level Structural Classification: The written archetype is perfectly isolated, while video workflows exhibit distinct cinematographic structural topology.The table caption attributes this separation to a fundamental difference between written and video-derived workflows.
- Archetype-Level Structural Classification: Video-derived workflows are elongated and deeply decomposed into sequential micro-actions, with increased ingredient reuse compared with shallow, parallelized written recipes.Depth and branching behavior govern the boundary between written and video formats.
- Archetype-Level Structural Classification: Short-form workflows retain deep sequencing but compress graph size and prune complex preparation toward linear dump-and-bake assembly.This compression changes operational behavior, not only the number of graph elements.
- The Illusion of Lexical Stylometry: Lexical accuracy does not establish procedural-style capture because lexical features conflate a creator’s repertoire with their procedure.Dish and ingredient inputs in the generative task remove niche-specific vocabulary, leaving structural signal as the transferable cue.
- The Scalability of Procedural Personas: As creator classes scale from K = 2 to K = 50, Top-1 Accuracy and Macro F1 decline, but structural persona signal strengthens relative to the random baseline.The growing gap from 1/K and stable AUC near 0.78 at K = 50 indicate ranking power despite weaker rigid classification boundaries.
- The Scalability of Procedural Personas: These results motivate generative modeling because discriminative classifiers reach a ceiling when procedural variance behaves as a distributional preference rather than a fixed boundary.Generation is presented as necessary to characterize and replicate a creator’s idiosyncratic routing.
5 Modeling Procedural Variance for Persona Generation
Persona-conditioned recipe generation predicts a creator’s physical execution graph rather than reproducing spoken style, evaluating semantic assignment and topology separately. Results expose complementary strengths: few-shot LLMs handle ingredients better, while structured generation controls macro-structure, motivating adaptive ensembling.
- Task formulation: Persona-conditioned generation predicts an execution graph for a dish and ingredients under a target creator, rather than the creator’s dialogue or vocabulary.The output models action states and ingredient assignments as the creator’s physical workflow.
- Evaluation paradigm: Topological alignment metrics replace text-overlap measures to evaluate whether generated graphs capture creator-specific operational structure.Ingredient F1 evaluates aligned steps, while normalized Edge F1 compares normalized state-ingredient tuples and penalizes structural hallucinations.
- Generation paradigms: Generation paradigms range from zero-shot and few-shot semantic prompting to step-by-step autoregressive prediction and LLM-free structured two-stage generation.The two-stage approach separates structural topology from ingredient assignment and applies creator-specific statistical priors.
- Performance comparisons: Few-shot LLMs achieve the highest Ingredient F1, whereas S2S achieves the lowest Step Error across long and short video formats.Few-shot IF1 is 0.344 and 0.385 versus S2S at 0.294 and 0.348; S2S Step Error is 0.350 and 0.332 versus 0.373 and 0.351.
- Performance comparisons: Few-shot and S2S models reach competitive nEF1 through complementary mechanisms, but microscopic step prediction fails to dominate because isolated predictions fracture overarching contextual logic.Few-shot nEF1 is 0.474 and 0.478, while S2S reaches 0.456 and 0.455 on long and short formats.
- Ensemble generation: An ensemble learned selector is motivated by the trade-off between LLM semantic reasoning and S2S topological control, dynamically routing generation to different models.LLM-driven approaches provide strong semantic reasoning but require black-box, token-intensive API calls at every step.
6 Ablation Study
The ablation study shows that model scale does not resolve macro-planning deficits, while few-shot prompting improves semantic assignment only until context dilution begins.
- Larger LLM backbones do not necessarily improve adherence to rigid procedural topologies.The comparison spans proprietary and open-weight architectures.
- Semantic assignment peaks at k = 9 few-shot exemplars before additional complex graphs cause context dilution from k = 11.Too many procedural exemplars overwhelm attention to stylistic signals.
7 Conclusion and Future Work
The paper frames procedural style as a creator-specific topological footprint and combines structured generation with LLM semantic agility through an ensemble. It identifies physical-constraint validation and broader-domain transfer as directions for future work.
- Procedural style is encoded in the physical shape of execution graphs rather than vocabulary.The paper introduces ViralRecipesTrans, operational-fidelity metrics, and a structured two-stage generation algorithm around this framing.
- The proposed ensemble explores the trade-off between structured macro-planning and few-shot LLM micro-semantic agility.
- Future work targets neuro-symbolic fusion and post-hoc calibration to validate physical constraints and repair violations.
- The framework is built in the culinary domain, while applying topological stylometry elsewhere remains a future direction.