MAP-AgMO: Multi-Agent framework for Personalized Agricultural Multi-Objective decision-making

Josué Kpodo, Parisa Kordjamshidi, A. Pouyan Nejadhashemi

Keywords: Multi-agent LLMs, multi-objective optimization, agricultural decision support, Pareto optimization, MCDM, benchmark evaluation, DSPy

TL;DR: MAP-AgMO is a multi-agent LLM framework for personalized farm decision-making under conflicting objectives, paired with the first benchmark for evaluating LLM reasoning in agricultural multi-objective settings.

Abstract: Feeding a growing and increasingly affluent global population demands substantial increases in demand for agricultural products. Achieving this requires decision support tools that can navigate the inherently conflicting objectives farmers face daily, including maximizing crop yield, minimizing water use, reducing costs, and promoting environmental sustainability. Yet a critical question remains unexplored: How effectively can multimodal AI models, trained primarily on internet-scale general corpora, understand and reason about domain-specialized, multi-objective agricultural problems? In such settings, relevant language, contextual knowledge, and structured decision criteria are often sparse or underrepresented in the training data. We propose MAP-AgMO, a multi-agent LLM framework in which specialized agents would collaboratively handle problem formulation, multi-objective optimization (generating Pareto-front non-dominated solutions via dedicated solvers), and Multi-Criteria Decision Making [1] to deliver personalized, farmer-specific recommendations. Built on top of DSPy [2] and inspired by the workflows of agricultural Extension human agents [3], the envisioned framework would be tool-augmented, drawing on weather data retrieval, database summarization, and end-to-end crop growth simulation via a crop model [4]. Most importantly, farmer preferences would be elicited both implicitly and explicitly through natural language interaction, enabling the framework to generalize across a variable number of conflicting objectives. Finally, to rigorously evaluate and diagnose the language understanding and reasoning of current SOTA LLMs in this setting, we propose the first benchmark dataset for agricultural multi-objective reasoning. Besides measuring performance, the benchmark is being designed to expose where these LLMs could systematically fail (e.g., quantitative trade-off reasoning, simulation-grounded inference, and preference consistency). We expect these findings to inform future domain-adapted model development. Ultimately, we present this work-in-progress to solicit feedback and foster collaboration.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 92