pith. sign in

arxiv: 2605.15918 · v1 · pith:3EOTBXACnew · submitted 2026-05-15 · 🧬 q-bio.QM

The Impact of Heatwaves on Population Health: A Large Language Model-Enhanced Agent-Based Simulation

Yuanhao Liu , Yuanfei Liu , Tian Lu , Hengyang Zhang , Zuowei Wang , Ying Dai This is my paper

Pith reviewed 2026-05-19 17:36 UTC · model grok-4.3

classification 🧬 q-bio.QM
keywords heatwavesagent-based modelinglarge language modelspsychosocial impactsvulnerabilitycomplex contagionclimate resiliencepopulation health
0
0 comments X

The pith

Large language model simulations show heatwave health impacts are mainly psychosocial and hit vulnerable groups hardest.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that an LLM-powered agent-based simulation of a 100-agent virtual community facing a 13-day heatwave reveals primarily psychosocial impacts that are unequally distributed by vulnerability. Higher-vulnerability agents experience steeper drops in perceived safety and social connection and show reduced protective behaviors, while lower-vulnerability agents maintain more adaptive routines. Risk information spreads through complex contagion, relying on repeated reinforcement inside tight social networks rather than single broad exposures. If these patterns hold, heat-risk strategies would need to combine targeted support for vulnerable individuals with community information pathways instead of uniform alerts. A sympathetic reader would care because rising heat events make it urgent to understand why some groups suffer more and how social mechanisms shape resilience.

Core claim

In this LLM-enhanced agent-based simulation, heat-related impacts are primarily psychosocial and unequally distributed. Agents with higher Heat Vulnerability Index scores experience larger declines in perceived safety and social connection than lower-vulnerability agents. Vulnerability also shapes adaptive capacity, with highly vulnerable agents showing behavioral constriction through reduced engagement in protective actions while more resilient agents maintain self-care routines. At the collective level, risk-information diffusion follows a pattern of complex contagion driven by repeated social reinforcement within cohesive networks.

What carries the argument

An LLM-enhanced agent-based model in which 100 heterogeneous agents receive Heat Vulnerability Indices from demographic risk factors and are tracked across baseline, heatwave, and recovery periods for changes in perceptions, behaviors, and information spread.

If this is right

  • Interventions should combine targeted support for vulnerable groups with community-based information pathways.
  • Psychosocial factors and social networks should be central to models of climate resilience.
  • Adaptive capacity varies by vulnerability level, so uniform policies may fail those most affected.
  • Risk communication works better through repeated reinforcement in cohesive networks than through broad one-time exposure.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same modeling approach could be applied to other climate hazards such as floods to compare mechanism patterns across event types.
  • Running the simulation with varied network structures or information sources could test which community features most improve collective resilience.
  • Policymakers might use such models to compare the effects of different intervention timings and targeting rules before real-world deployment.

Load-bearing premise

The large language model generates agent behaviors and perceptions that accurately represent real human psychosocial responses to heat stress and community social dynamics without empirical calibration or validation against observed data.

What would settle it

A real-world survey or observational study during an actual heatwave that finds no significant difference in perceived safety, social connection, or protective actions between high- and low-vulnerability groups would challenge the simulation results.

read the original abstract

Extreme heat events are increasing in frequency and intensity under climate change, but the socio-behavioral mechanisms that shape community resilience remain insufficiently understood. This study uses a Large Language Model-enhanced agent-based model to simulate responses to a prolonged heatwave in a virtual society. One hundred heterogeneous agents were assigned a Heat Vulnerability Index based on demographic risk factors and observed over 13 simulated days covering baseline, heatwave, and recovery periods. The simulation shows that heat-related impacts are primarily psychosocial and unequally distributed. Agents with higher vulnerability experienced larger declines in perceived safety and social connection than agents with lower vulnerability. Vulnerability also shaped adaptive capacity. More resilient agents maintained routine self-care and protective behaviors, whereas highly vulnerable agents showed behavioral constriction, marked by reduced engagement in protective actions. At the collective level, risk-information diffusion followed a pattern of complex contagion, with adoption driven more by repeated social reinforcement within cohesive networks than by broad exposure alone. These findings suggest that LLM-enhanced simulation can help identify behavioral and social mechanisms of climate resilience and inform heat-risk interventions that combine targeted support for vulnerable groups with community-based information pathways.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper describes an LLM-enhanced agent-based simulation of 100 heterogeneous agents assigned a Heat Vulnerability Index and observed over 13 simulated days (baseline, heatwave, recovery) of a prolonged heat event. It claims that heat impacts are primarily psychosocial and unequally distributed, with higher-vulnerability agents showing larger drops in perceived safety and social connection, behavioral constriction (reduced protective actions), and collective risk-information diffusion exhibiting complex contagion driven by repeated social reinforcement rather than broad exposure.

Significance. If the LLM-generated heterogeneous behaviors prove representative of real psychosocial responses, the results could illuminate vulnerability-driven mechanisms of climate resilience and support targeted interventions combining support for high-risk groups with community information pathways. The approach is a methodological strength in using LLMs to explore emergent social dynamics in ABMs without relying solely on traditional datasets.

major comments (2)
  1. [Methods] Methods (agent behavior generation and simulation setup): The central claims about unequal psychosocial impacts, behavioral constriction, and complex contagion depend entirely on the LLM producing realistic agent perceptions and actions. No empirical calibration, comparison to heatwave survey data, epidemiological records, or prior behavioral studies is reported, nor is any sensitivity analysis on prompt variations or LLM stochasticity provided. This is load-bearing for the reported patterns in the Results.
  2. [Results] Results (vulnerability-stratified outcomes and contagion analysis): The manuscript presents declines in safety/social connection and the complex-contagion pattern without quantitative validation metrics, robustness checks against alternative vulnerability assignments, or falsification tests that would distinguish LLM artifacts from genuine mechanisms.
minor comments (2)
  1. [Abstract] Abstract: The phrasing 'LLM-enhanced simulation can help identify behavioral and social mechanisms' should be tempered to reflect the current absence of external validation.
  2. [Methods] Notation: Clarify how the Heat Vulnerability Index is operationalized as a continuous or categorical variable and how it is injected into the daily LLM prompts.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments, which identify key areas where additional transparency and checks can strengthen the manuscript. We have revised the paper to expand the Methods with prompt details and sensitivity analyses, and to augment the Results with quantitative metrics and robustness checks. We respond to each major comment below.

read point-by-point responses

  1. Referee: [Methods] Methods (agent behavior generation and simulation setup): The central claims about unequal psychosocial impacts, behavioral constriction, and complex contagion depend entirely on the LLM producing realistic agent perceptions and actions. No empirical calibration, comparison to heatwave survey data, epidemiological records, or prior behavioral studies is reported, nor is any sensitivity analysis on prompt variations or LLM stochasticity provided. This is load-bearing for the reported patterns in the Results.

    Authors: We agree that the absence of direct empirical calibration against survey or epidemiological data is a limitation, as the simulation relies on LLM-generated behaviors without external anchoring. The study is framed as exploratory to probe plausible mechanisms in a controlled setting rather than as a calibrated predictive model. In the revised manuscript we have added a detailed Methods subsection describing the prompt templates, agent initialization logic, and the specific LLM configuration used. We have also incorporated a sensitivity analysis varying temperature, prompt phrasing, and stochastic seeds, with results showing that the main qualitative patterns (vulnerability gradients and complex contagion) persist across these variations. The Discussion now explicitly notes the lack of real-world calibration and positions the work as hypothesis-generating. revision: yes

  2. Referee: [Results] Results (vulnerability-stratified outcomes and contagion analysis): The manuscript presents declines in safety/social connection and the complex-contagion pattern without quantitative validation metrics, robustness checks against alternative vulnerability assignments, or falsification tests that would distinguish LLM artifacts from genuine mechanisms.

    Authors: We accept that the original Results lacked sufficient quantitative safeguards. The revised manuscript now reports effect sizes and within-agent change statistics for the safety and social-connection declines, together with adoption-threshold and reinforcement-count metrics for the contagion process. We have added robustness checks that reassign the Heat Vulnerability Index using alternative demographic weightings and compare outcomes against a null model in which agent responses are randomized. These additions are included to help separate the reported patterns from potential LLM-specific artifacts. revision: yes

Circularity Check

0 steps flagged

No significant circularity in the simulation-based derivation

full rationale

The paper describes an exploratory agent-based simulation in which 100 heterogeneous agents are assigned a Heat Vulnerability Index and then interact over 13 simulated days, with daily perceptions, protective actions, and social diffusion generated by an LLM. The reported findings (unequal psychosocial impacts, behavioral constriction in high-vulnerability agents, and complex contagion in risk information) are direct outputs of running this model rather than quantities derived by reducing equations to fitted parameters or by self-citation chains. No load-bearing self-citations, ansatz smuggling, or self-definitional steps are present in the manuscript; the derivation chain consists of the simulation setup itself and is therefore self-contained as a modeling study.

Axiom & Free-Parameter Ledger

3 free parameters · 1 axioms · 0 invented entities

The simulation depends on multiple uncalibrated parameters for agent heterogeneity and the core assumption that LLM outputs faithfully model human behavior, none of which are derived from external benchmarks in the abstract.

free parameters (3)
  • Number of agents
    Set at 100 to represent a virtual society; scale chosen without stated justification from real population data.
  • Simulation duration and periods
    13 days covering baseline, heatwave, and recovery; specific lengths and transitions are modeling choices.
  • Heat Vulnerability Index assignment
    Based on demographic risk factors but exact weighting, thresholds, and mapping to agent traits not detailed.
axioms (1)
  • domain assumption Large language models can generate realistic heterogeneous agent behaviors and perceptions in social simulations of climate events
    Invoked as the foundation for assigning responses to heat stress, safety, social connection, and protective actions.

pith-pipeline@v0.9.0 · 5740 in / 1418 out tokens · 57063 ms · 2026-05-19T17:36:29.979353+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

16 extracted references · 16 canonical work pages · 2 internal anchors

  1. [1]

    R., Liu, R., Richardson, S

    “LLM Social Simulations Are a Promising Research Method.” arXiv preprint arXiv:2504.02234. Bai, J., Bai, S., Chu, Y., Cui, Z., Dang, K., Deng, X., ... & Zhu, T. (2023). Qwen technical report. arXiv preprint arXiv:2309.16609. Bail, Christopher A

    work page arXiv 2023

  2. [2]

    Can Generative AI Improve Social Science?

    “Can Generative AI Improve Social Science?” Proceedings of the National Academy of Sciences 121(21):e2314021121. Badr, H. S., Zaitchik, B. F., Kerr, G. H., Nguyen, N. L. H., Chen, Y. T., Hinson, P., ... & Gardner, L. M. (2023). Unified real-time environmental-epidemiological data for multiscale modeling of the COVID-19 pandemic. Scientific data, 10(1),

    work page 2023

  3. [3]

    Brian JL Berry, L Douglas Kiel, and Euel Elliott

    Retrieved September 28, 2025 (https://www.wsj.com/tech/ai/meta-is-delaying-the-rollout-of-its-flagship-ai-model). Brian JL Berry, L Douglas Kiel, and Euel Elliott. Adaptive agents, intelligence, and emergent human organization: Capturing complexity through agent-based modeling. Proceedings of the National Academy of Sciences, 99(suppl_3):7187–7188,

    work page 2025

  4. [4]

    Complex Contagions and the Weakness of Long Ties

    “Complex Contagions and the Weakness of Long Ties.” American Journal of Sociology 113(3):702–734. Cheng, W., Li, D., Liu, Z., & Brown, R. D. (2021). Approaches for identifying heat-vulnerable populations and locations: A systematic review. The Science of the total environment, 799, 149417. https://doi.org/10.1016/j.scitotenv.2021.149417 Guilbeault, Dougla...

    work page doi:10.1016/j.scitotenv.2021.149417 2021

  5. [5]

    Beyond Static Responses: Multi-Agent LLM Systems as a New Paradigm for Social Science Research

    “Beyond Static Responses: Multi-Agent LLM Systems as a New Paradigm for Social Science Research.” arXiv preprint arXiv:2506.01839. Inostroza, L., Palme, M., & de la Barrera, F. (2016). A Heat Vulnerability Index: Spatial Patterns of Exposure, Sensitivity and Adaptive Capacity for Santiago de Chile. PloS one, 11(9), e0162464. https://doi.org/10.1371/journa...

    work page doi:10.1371/journal.pone.0162464 2016

  6. [6]

    Scaling Laws for Neural Language Models

    “Scaling Laws for Neural Language Models.” arXiv preprint arXiv:2001.08361. Klinenberg, Eric

    work page internal anchor Pith review Pith/arXiv arXiv 2001

  7. [7]

    Chicago: University of Chicago Press

    Heat Wave: A Social Autopsy of Disaster in Chicago. Chicago: University of Chicago Press. Kovats, R. S., & Kristie, L. E. (2006). Heatwaves and public health in Europe. European journal of public health, 16(6), 592–599. https://doi.org/10.1093/eurpub/ckl049 Kozlowski, Austin C., and James Evans

    work page doi:10.1093/eurpub/ckl049 2006

  8. [8]

    Simulating Subjects: The Promise and Peril of Artificial Intelligence Stand-Ins for Social Agents and Interactions

    “Simulating Subjects: The Promise and Peril of Artificial Intelligence Stand-Ins for Social Agents and Interactions.” Sociological Methods & Research. doi:10.1177/00491241251337316. Li An, Volker Grimm, Abigail Sullivan, BL Turner Ii, Nicolas Malleson, Alison Heppenstall, Christian Vincenot, Derek Robinson, Xinyue Ye, Jianguo Liu, et al. Challenges, tasks...

    work page doi:10.1177/00491241251337316

  9. [9]

    Li, F., Yigitcanlar, T., Nepal, M., Nguyen, K., Dur, F., & Li, W. (2024). Assessing heat vulnerability and multidimensional inequity: Lessons from indexing the performance of Australian capital cities. Sustainable Cities and Society, 115, 105875. https://doi.org/10.1016/j.scs.2024.105875 Macy, M. W., & Willer, R. (2002). From factors to actors: Computatio...

    work page doi:10.1016/j.scs.2024.105875 2024

  10. [10]

    https://doi.org/10.3390/ijerph15071433 Miles-Novelo, A., & Anderson, C. A. (2023). Avoiding a grim future: The climate crisis and its effects on human aggression and violence. Advances in Environmental and Engineering Research, 4(2), Article

    work page doi:10.3390/ijerph15071433 2023

  11. [11]

    https://doi.org/10.21926/aeer.2302034 National Disease Control and Prevention Administration. (2022). Public Health Protection Guidelines for High Temperature and Heatwaves. Beijing: NCDPC Newport, Cal

    work page doi:10.21926/aeer.2302034 2022

  12. [12]

    Park, Joon Sung, Joseph C

    Retrieved September 28, 2025 (https://www.newyorker.com/culture/open- questions/what-if-ai-doesnt-get-much-better-than-this). Park, Joon Sung, Joseph C. O’Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein

    work page 2025

  13. [13]

    Generative Agents: Interactive Simulacra of Human Behavior

    “Generative Agents: Interactive Simulacra of Human Behavior.” arXiv preprint arXiv:2304.03442. Pearce, M., Garcia, L., Abbas, A., Strain, T., Schuch, F. B., Golubic, R., Kelly, P., Khan, S., Utukuri, M., Laird, Y., Mok, A., Smith, A., Tainio, M., Brage, S., & Woodcock, J. (2022). Association between physical activity and risk of depression: A systematic r...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1001/jamapsychiatry.2022.0609 2022

  14. [14]

    Navigating Artificial General Intelligence Development: Societal, Technological, Ethical, and Brain-Inspired Pathways

    “Navigating Artificial General Intelligence Development: Societal, Technological, Ethical, and Brain-Inspired Pathways.” Scientific Reports 15(1):1–22. Ray, D. K., West, P. C., Clark, M., Gerber, J. S., Prishchepov, A. V., & Chatterjee, S. (2019). Climate change has likely already affected global food production. PloS one, 14(5), e0217148. https://doi.org...

    work page doi:10.1371/journal.pone.0217148 2019

  15. [15]

    These variables were normalized and coded on a standardized scale (1-3) to ensure comparability, with higher values indicating greater vulnerability

    (Cheng et al., 2021; Inostroza et al., 2016), socioeconomic status (combining income and education levels), race/ethnicity (accounting for documented disparities) (Inostroza et al., 2016), residential characteristics (Li et al., 2024), and health proxies (using marital status as an indicator of social isolation) (Li et al., 2024). These variables were nor...

    work page 2021

  16. [16]

    relevant to heatwave

    Two-level mixed-effects modeling of health vulnerability and emotional scores Surprise Joy Fear Anger Sadness Disgust Predict ors Estima tes std. Err or Estima tes std. Err or Estima tes std. Err or Estima tes std. Err or Estima tes std. Err or Estima tes std. Err or (Intercept ) 4.72 *** 0.24 7.77 *** 0.11 2.17 *** 0.10 1.04 *** 0.08 1.76 *** 0.07 1.51 *...

    work page 2007