An exploration of using large language models to integrate farmer behaviour into an agricultural systems model of the Peruvian Andes

The implications of climate change on agro-pastoral farming systems in the Peruvian Andes are not fully understood. There is already a significant impact on agricultural productivity from current climate variability and extreme weather in the region. This is exacerbated by chronic poverty in many rural areas and the need for improved government-led strategic planning. Tools to assist with policy planning for climate change adaptations that achieve environmental and social resilience are vital, and these require collaboration with rural communities to incorporate the complexities of behavioural responses to climate change, market dynamics, and policy shifts in agricultural and water management. 

In this study we further develop a recent agricultural systems model (the TELLUS model; Pilditch et al., in review). The model is an agent-based simulation focussed on the behaviour of interacting populations of individual farming agents. TELLUS offers the opportunity to analyse the impact of interventions/policies in light of key scenarios and conditions of interest, with potential to uncover unforeseen emergent behaviours within farming systems (e.g., tipping points, amplifiers, system adaptations) and potential unintended consequences of scenarios and policies (e.g., increasing in equalities; increased system fragility). A difficulty in applying such models to specific case studies is in choosing valid parameter values, especially for model behaviour associated with human behaviour and decision-making.

Our work over recent years includes extensive fieldwork in the Cordillera Negra and Cordillera Blanca, involving interviews and workshops with farming communities, and collaboration with regional NGOs. These interactions have been instrumental in understanding local challenges and priorities. The challenge in terms of modelling this system is turning information gained from qualitative methods (e.g. interviews) into parameter values for the model. Our novel approach is to assess the extent to which modern AI systems, specifically, Large Language Models (LLMs) can help perform this task.  We leverage the reasoning abilities of LLMs to directly estimate relevant model parameters from automated interview transcription/translations. We will discuss the extent to which this integration has aided the creation of a TELLUS model tuned specifically to the Peruvian Andes context. Our approach will hopefully serve as a novel tool, combining empirical research, community involvement, and advanced computational modelling, to explore future climate scenarios and the potential effects of policy interventions.