Cambridge Team Simulates a Heatwave With 100 LLM-Driven Agents

A virtual society under a prolonged heatwave

The project, described on the Cambridge Accelerate Programme for Scientific Discovery site, grew out of a research challenge on the impact of heatwaves. It received funding in the 2025 Accelerate-C2D3 funding call, which the programme describes as supporting novel applications of AI for research and innovation. The funding let the team implement an agent-based model simulation.

According to the researchers, the link between heat and adverse outcomes is established, but the social and behavioural drivers behind it are poorly understood. They argue that traditional statistical methods often fail to capture the dynamic, non-linear interactions that shape how people adapt during environmental crises. Their question was whether a society of LLM-driven agents could reproduce the social mechanisms of heatwave vulnerability.

The model draws on two behavioural theories. Agent needs follow Maslow's Hierarchy, which describes how motivation moves from basic survival to social fulfilment. Complex Contagion Theory explains how behaviours spread across a group when reinforced by multiple social contacts.

How the simulation was built

The team used Agent Society's large scale simulation platform, which is designed for LLM-driven generative social science. The simulation involved 100 heterogeneous agents with psychological and social profiles derived from real-world data. Their cognitive functions were powered by the Qwen LLM.

To measure susceptibility, the researchers built a Heat Vulnerability Index based on literature-defined risk factors, including age, socioeconomic status, race and residential characteristics. Agents were sorted into four levels: low, moderate, high and extreme.

The scenario ran for 13 days in three phases: a three-day baseline, a five-day heatwave with temperatures reaching 40°C, and a five-day recovery period. The team recorded time-stamped data on emotional states, behaviours, satisfaction of physiological and social needs, and interpersonal communications.

Early findings on inequality and social reinforcement

According to the researchers' preliminary results, the psychological impacts of the heatwave were distributed inequitably. Highly vulnerable agents showed steeper declines in perceived safety and social connection than low-vulnerability peers. Resilient agents kept up routine self-care such as staying hydrated and seeking cooler environments, while more vulnerable agents took fewer protective actions.

On the spread of risk information, adoption of helpful measures was driven by social reinforcement in cohesive networks rather than broad exposure to warnings or messages. Actions were more likely to be adopted when several close connections, such as friends, family members or neighbours, reinforced them. The team says this suggests real-world heatwave interventions could work through trusted community networks alongside mass alerts.

The authors also argue that the primary threat of a heatwave, in a society with stable access to food and shelter, may be psychosocial, especially for vulnerable people. They describe the work as a simulated, mechanistic basis for prioritising community-level interventions that strengthen social ties within climate resilience strategy.

Validation, data sharing and next steps

The researchers acknowledge that the model must still be validated with real-world data, which they call a big challenge. One idea is to focus on a specific population, such as pregnant women. That would require a rule-based biological model of pregnancy, environmental data and birth outcomes, and would be a longer-term project involving a clinician.

The team is preparing its results for publication and has made its model and data sources available on GitHub for others to use or build on. Members presented a poster at the Population Association of America 2026 Annual Meeting, supported by Accelerate Programme funding. The author, who has a public health rather than computer science background, also used part of the grant to develop machine learning skills.

For public health planning, the work offers a simulated indication that messaging aimed at close-knit networks may help more than broadcast warnings alone. Because the findings come from a simulation and remain preliminary, the authors stress the need for validation against real populations.