Scholarlink

Large Language Models (LLMs) are increasingly applied to culinary tasks such as recipe generation, yet their limitations in producing high-quality, executable recipes remain underexplored. This paper systematically investigates deficiencies in LLM-generated recipes, with a focus on ingredient proportions. We conduct a thematic analysis of user comments from YouTube cooking channels and perform experiments evaluating multiple LLMs on ingredient scaling tasks. Our findings reveal significant challenges, including incorrect ingredient quantities, poor texture, and flavor issues, although users appreciate innovative combinations in some recipes. We also explore fine-tuning through Direct Preference Optimization (DPO) to improve ingredient quantity scaling, yielding modest gains (8 – 15% across evaluation metrics). While the results are promising, they also highlight the need for more advanced methods—such as multi-agent systems for validating ingredient proportions, hybrid models combining structured culinary knowledge with LLMs, and targeted pretraining on verified datasets. We anticipate these approaches will address fundamental knowledge limitations and contribute to the development of more robust and context-aware AI cooking assistants.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Improving LLMs for Recipe Generation via Reinforcement Learning Fine-Tuning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Prashant Shaw,
  • Asanka Wasala,
  • Rohitt R. Punjj

摘要

Large Language Models (LLMs) are increasingly applied to culinary tasks such as recipe generation, yet their limitations in producing high-quality, executable recipes remain underexplored. This paper systematically investigates deficiencies in LLM-generated recipes, with a focus on ingredient proportions. We conduct a thematic analysis of user comments from YouTube cooking channels and perform experiments evaluating multiple LLMs on ingredient scaling tasks. Our findings reveal significant challenges, including incorrect ingredient quantities, poor texture, and flavor issues, although users appreciate innovative combinations in some recipes. We also explore fine-tuning through Direct Preference Optimization (DPO) to improve ingredient quantity scaling, yielding modest gains (8 – 15% across evaluation metrics). While the results are promising, they also highlight the need for more advanced methods—such as multi-agent systems for validating ingredient proportions, hybrid models combining structured culinary knowledge with LLMs, and targeted pretraining on verified datasets. We anticipate these approaches will address fundamental knowledge limitations and contribute to the development of more robust and context-aware AI cooking assistants.