본문 바로가기

Culinary Class Wars: Evaluating LLMs Using ASH in Cuisine Transfer Task

문서 유형
연구출판물 · 학술대회 발표논문
등재
Scopus
저널 순위
Scopus Q3
피인용
Scopus 1
FWCI (Scopus)
0.85
ISSN
2375-9232

저자

  1. Lee, Hoonick제1저자Korea University, Seoul, Republic of Korea
  2. Gim, MoganHankuk University of Foreign Studies, Yongin, Republic of Korea
  3. 최동희교신저자Pusan National University, Busan, Republic of Korea
  4. Kang, Jaewoo교신저자Korea University, Seoul, Republic of Korea

초록

Computational gastronomy increasingly leverages data mining and AI to reason about recipes, cuisines, and culture. Recent advances in large language models (LLMs) such as GPT-4, LLaMA, Mistral, and Gemma have demonstrated strong instruction-following and generative abilities, making them promising tools for creative tasks. Yet, despite these advances, LLMs often fall short when required to generate culturally faithful and contextually coherent recipes. We study LLMs in cuisine transfer-adapting a base dish to a target cuisine-as a structured setting for creativity under cultural constraints. We introduce the AS H (authenticity, sensitivity, harmony) benchmark and a generator-evaluator pipeline: six open-source LLMs generate 4,800 recipes from 800 standardized prompts spanning 20 base dishes and 40 cuisines, and nine LLM evaluators, together with a stratified subset of 1,000 human ratings, assess cultural fidelity and overall quality. Across models and cuisines, sensitivity scores are consistently higher than authenticity and harmony, revealing a tendency to overvalue surface-level cultural cues; agreement patterns cluster by model family, and some evaluators (e.g., 11ama2:13b) align more closely with human ratings. We discuss implications for 'LLM-as-a-judge' in creative domains within computational gastronomy.

키워드

ASH BenchmarkComputational GastronomyCreativity EvaluationCuisine TransferCultural AILarge Language ModelsLLM-as-a-JudgeRecipe Generation

식별자

Scopus EID
2-s2.0-105035379137