Research map / Large language models / AI and society

What can large language models do for sustainable food?

ICML 2025 · Large language models › AI and society · approach: Critiques of LLM claims and judging

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Evaluates six LLMs on four sustainable-food tasks (experimental design, menu design, sensory profile prediction, recipe preference prediction) and proposes LLM-guided integer quadratic programming to cut menu emissions by 79% while keeping patron satisfaction.

Problem

LLMs' potential for reducing food-system environmental impact is underexplored; LLMs struggle to design menus that cut emissions while maintaining human satisfaction, and their preference prediction is weak for fine-grained taste.

Approach

Use LLMs to generate candidate recipes and estimate per-item human preference scores, then solve the subset selection with integer quadratic programming under emissions and animal-welfare constraints, rather than asking the LLM to directly produce the optimized menu.

Key results

• o1-preview's experimental designs were rated more specific than expert food scientists' (p=0.003), with estimated time saved of 45% vs 22% for human collaborators.
• o1-preview+IQP reduced menu emissions by 79% (p=1.37e-18) while keeping satisfaction with choices within 1% of the original menu (n=800 Prolific participants).
• LLM sensory profile prediction was generally not meaningfully better than a nutrition-based baseline; GPT-4o reached 64% accuracy on recipe preference prediction, versus 50% random.

Measured effects reported in the paper

Surprising findings

• Direct LLM menu revision produced fully plant-based menus that significantly reduced patron satisfaction.
• On recipe preference prediction, LLMs showed an omnivore bias and worst accuracy on vegetarian-vs-vegetarian comparisons.

Limitations

• Minimal prompt engineering; all evaluations zero-shot.
• NECTAR sensory data restricted to academic access and only ingredient/nutrition info available.
• Menu study lacked real delivery incentives and did not model food waste or portion size.

Experimental setup

Datasets: NECTAR sustainable protein sensory dataset Food.com recipes and reviews epicurious.com recipe pairs (post-cutoff) UK Deliveroo/Just Eat 36-item menu

Benchmarks: NECTAR sensory profile prediction Food.com recipe preference prediction menu design human subjects study

Metrics: pairwise ranking accuracy accuracy vs random baseline emissions (kg CO2eq) patron satisfaction (1-10) expert-rated accuracy/specificity/complementarity/time saved animal welfare (animals killed per kg)

Baselines: Claude 3.5 Sonnet Gemini 1.5 Pro GPT-3.5 Turbo GPT-4o Llama 3.1 70B Instruct o1-preview Vegetarian subset Remove beef Vegetarian first random baseline

Builds on: predict-then-optimize OptiMUS SatLM Logic-LM

Closest papers by meaning

• Empowering and Assessing the Utility of Large Language Models in Crop Science (NeurIPS 2024 · similarity 0.86)
• Holistically Evaluating the Environmental Impact of Creating Language Models (ICLR 2025 · similarity 0.84)
• LLMCarbon: Modeling the End-to-End Carbon Footprint of Large Language Models (ICLR 2024 · similarity 0.84)
• Assessing Large Language Models on Climate Information (ICML 2024 · similarity 0.84)
• SuperGPQA: Scaling LLM Evaluation across 285 Graduate Disciplines (NeurIPS 2025 · similarity 0.84)
• Knowledge-Powered Recommendation for an Improved Diet Water Footprint (AAAI 2024 · similarity 0.83)