Leveraging LLMs for predictive insights in food policy and behavioral interventions

Abstract Food consumption and production significantly contribute to global greenhouse gas emissions, making them key targets for climate change mitigation. Over the past two decades, food policy initiatives have focused on reshaping production and consumption patterns by reducing food waste and curbing ruminant meat consumption. While evidence on effective interventions is improving, assessing appropriate and context-specific policies remains difficult due to external validity concerns. This paper demonstrates that a fine-tuned large language model (LLM) can accurately predict outcome directions in approximately 80% of empirical studies evaluating dietary interventions. Predictive accuracy improves with richer input detail, peaking at around 75 prompts before declining due to overfitting or saturation. To contextualize these results, we benchmark the LLM against both classical random-effects meta-regression and a prompt-based variant executed entirely within the model. Although traditional approaches yield reasonable magnitude estimates, they lag behind LLMs in directional accuracy and adaptability to diverse intervention formats. Together, our findings suggest that LLMs-especially when fine-tuned on curated evidence-offer a scalable pathway for data-driven, context-sensitive food policy modeling.