Given that the food system accounts for one-third of human-caused greenhouse gas emissions, this paper explores how large-scale language models (LLMs) can contribute to reducing the environmental impact of food production. Drawing on the sustainable food literature and expert collaboration, we define a typology of design and prediction tasks, and evaluate six LLMs on four tasks. For example, in a sustainable protein design task, collaboration with LLMs resulted in an average time savings of 45% compared to collaboration with other expert human food scientists (22%). However, in a sustainable menu design task, LLMs produced suboptimal solutions when instructed to consider both human satisfaction and climate impacts. We propose a general framework that integrates combinatorial optimization to enhance the inference capabilities of LLMs, resulting in a 79% reduction in food choice emissions in a virtual restaurant while maintaining participant satisfaction. In conclusion, we demonstrate that LLMs supported by optimization techniques have the potential to accelerate sustainable food development and adoption.
