Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

LLMs as World Models: Data-Driven and Human-Centered Pre-Event Simulation for Disaster Impact Assessment

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that pretrained LLMs prompted with pre-event geospatial, building, socioeconomic, and street-view data can simulate human-perceived earthquake intensity in advance, matching USGS DYFI reports with correlation of 0.88 and…

desk verdict A useful proof-of-concept that overclaims: the headline metrics don't match the tables and no distance-only baseline is tested, so the 'world model' conclusion outruns the evidence. read the letter →

arxiv 2506.06355 v1 pith:4FEZO2W7 submitted 2025-06-02 cs.CY cs.CEcs.CLcs.CV

Lingyao Li , Dawei Li , Zhenhui Ou , Xiaoran Xu , Jingxiao Liu , Zihui Ma , Runlong Yu , Min Deng This is my paper
classification cs.CYcs.CEcs.CLcs.CV
keywords largelanguagemodelsworldearthquakesimulationModifiedMercalliIntensityDidYouFeelItpre-eventdisasterassessmentmultimodaldatastreetviewimagery
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper sets out to show that a pretrained large language model can act as a 'world model' for sudden-onset disasters: given only pre-event data about an earthquake, it can estimate how people at specific locations would perceive and report the shaking. The authors build a virtual-sensor pipeline that feeds each sampled location a bundle of earthquake parameters, site-condition (VS30) values, building characteristics from open map data, census-block socioeconomic variables, and a street-view image, then asks the model to rate the location on the Modified Mercalli Intensity scale. Evaluated on the 2014 Napa and 2019 Ridgecrest earthquakes, the pipeline's zip-level and county-level averages align with crowd-sourced 'Did You Feel It?' reports, with best reported correlation of 0.88 and RMSE of 0.77. If this holds, emergency managers could get cheap, human-centered impact forecasts before a quake instead of waiting for post-event sensors and surveys.

What carries the argument

The load-bearing machinery is the 'LLM as virtual sensor' prompt: a role-based template that gives the model earthquake parameters, distance from epicenter, VS30 site condition, open-map building summaries, census-block socioeconomic indicators, and a street-view image, and asks for chain-of-thought reasoning followed by a Modified Mercalli Intensity rating (I–XII). Point-level ratings are averaged per zip code or county, and those averages are compared with crowd-sourced community intensities using Pearson correlation and RMSE. The framework also wraps the base prompt with retrieval-augmented generation and in-context demonstrations, which the paper shows lower RMSE even with a few examples.

What would settle it

A prospective check: issue pre-event MMI forecasts for all zip codes before a future California earthquake, then compare each zip's predicted intensity with actual crowd-sourced shaking reports; if accuracy is substantially worse in zip codes with few or no reports than in the original top-100 set, the pre-event generalisation claim is not supported.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that LLMs can effectively simulate human-centric seismic risk: when prompted as seismic experts with rich multimodal pre-event features, the models produce area-level MMI estimates that track real crowd-sourced shaking reports, with the best cases reaching correlation of 0.88 and RMSE of 0.77. The claim is explicitly a pre-event simulation claim, not a post-event damage assessment: the model never sees observed shaking or damage reports, only the earthquake source parameters, local site conditions, building stock, demographics, and street imagery. The paper also reports that retrieval-augmented generation and in-context learning improve accuracy, that visual street-view input matters more than structured numerical features, and that closed-source and larger models generally perform better. These results are presented as evidence that LLMs have learned enough about seismic attenuation and the built environment to serve as synthetic observers for pre-event planning.

Load-bearing premise

The evaluation assumes that testing only the 100 zip codes with the most 'Did You Feel It?' responses is a fair test of general pre-event simulation, even though those areas are likely the ones closest to the epicenter or most densely populated.

Editorial extensions

If this is right

  • Pre-event MMI maps could be generated for any earthquake source scenario using only open data, giving emergency managers a human-impact forecast before shaking starts.
  • Because RAG and ICL improve predictions with only a handful of demonstrations, the approach can be adapted to new regions without fine-tuning.
  • Street-view imagery carries more predictive signal than structured numeric features, which argues for investing in visual built-environment data in hazard modeling.
  • The models internalize distance-based attenuation but not site-condition (VS30) effects, so operational systems should either add physics-based priors or train on data that makes site effects salient.
  • Closed-source and larger models generally perform better, making model choice a practical trade-off between cost, privacy, and simulation accuracy.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • An implication left implicit in the paper, and not tested here, is that the headline accuracy may depend on the evaluation's restriction to the 100 response-rich zip codes, where distance-to-epicenter likely dominates; a test on sparse-report or held-out areas would clarify whether the model truly generalizes pre-event.
  • The same 'virtual observer' setup could transfer to other sudden-onset hazards with crowd-sourced intensity labels, such as floods, heat waves, or storms, where street-view and census data are available.
  • One testable extension suggested by the feature analysis is that translating numeric features into natural-language descriptions, such as 'very soft soil' instead of 'VS30 346 m/s', might improve alignment by matching the LLM's training distribution.
  • If distance alone drives most predictions, the model may be rediscovering a ground-motion decay curve rather than performing human-centered reasoning; a controlled comparison that removes distance from the prompt would separate those explanations.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes an LLM-based framework for simulating perceived earthquake impacts (Modified Mercalli Intensity, MMI) at zip-code and county scales, using multimodal inputs including earthquake source parameters, geospatial site conditions, building characteristics, socioeconomic data, and street-view imagery. The framework is evaluated on the 2014 Napa and 2019 Ridgecrest earthquakes, with USGS 'Did You Feel It?' (DYFI) reports as ground truth. The authors report correlations and RMSEs across several open- and closed-source LLMs, and analyze the effects of retrieval-augmented generation (RAG), in-context learning (ICL), feature ablations, and model scaling. The central claim is that LLMs can effectively simulate human-centric seismic risk and support pre-event disaster planning.

Significance. If the central claim were strongly supported, the paper would contribute a novel human-centered, multimodal approach to pre-event seismic impact simulation, with practical value for scenario-based planning. The paper's strengths include a systematic multi-model comparison, a clear pipeline description, ablation studies on input modalities, a reasoning-output analysis, and the use of publicly available datasets. The framework is reproducible in principle, and the comparison against DYFI ground truth is an appropriate target for a human-perception-oriented measure. However, the quantitative evidence as presented does not substantiate the headline claim: the abstract's numbers do not match the main table, the evaluation set is selected in a way that biases toward near-source, high-response areas, no non-LLM baseline is provided, and the RAG experiments leak reported MMI values into the prompt. These issues are load-bearing for the paper's central claim, though they appear correctable with additional experiments and revised reporting.

major comments (5)
  1. [Abstract and Table 1] The abstract's headline metrics are not supported by the reported results. The abstract states 'high correlation of 0.88 and low RMSE of 0.77 as compared to real reports at the zip code level,' but Table 1 shows that the best zip-level correlation is Corr_Z = 0.77 (GPT-4o, Napa) and the best zip-level RMSE is RMSE_Z = 0.92 (GPT-4.1-mini, Ridgecrest). The values 0.88 and 0.77 appear to be county-level numbers (Corr_C and RMSE_C) from different models/events. The central quantitative claim must be corrected, or the analysis and abstract must be revised to report consistent, clearly labeled metrics.
  2. [Section 4, Data Sampling and Evaluation] The evaluation restricts the test set to 'the top 100 zip codes with the highest number of responses' of DYFI. These zip codes are disproportionately near the epicenter or densely populated, where the distance-to-epicenter gradient dominates the MMI signal. The evaluation therefore does not measure predictive skill for unobserved, sparsely populated, or far-field areas, which is exactly the setting relevant to the paper's pre-event simulation claim. The authors should evaluate on all zip codes with DYFI data, or a random or spatially stratified holdout subset, and report performance separately for near-source and far-source regions.
  3. [Section 3.3 and Figure 5] No distance-only or physics-based baseline is reported. The prompts contain magnitude, epicentral distance, and depth, and Figure 5 shows that predicted MMI is strongly controlled by epicentral distance. A fitted ground-motion prediction equation or a simple attenuation curve could plausibly achieve comparable correlation and RMSE on this selected subset. Without such a baseline, the reported LLM performance does not establish that the model contributes anything beyond distance decay. Add at least one non-LLM baseline and compare it on the same evaluation set.
  4. [Section 3.5, RAG] The RAG experiment injects ground-truth labels into the prompt. The text states that RAG provides 'the reported MMI within the prompt as the retrieved context.' Since reported MMI is exactly the target variable being predicted, the RAG and ICL improvements may reflect retrieval of known outcomes rather than enhanced simulation ability. The authors should clarify whether the retrieved demonstrations come from the same event and whether the example MMI values are part of the query; if so, the RAG/ICL results cannot be used to support the main claim and must be re-analyzed with strict separation of retrieval context and labels.
  5. [Sections 1 and 4, Pre-event framing] The paper frames the contribution as 'pre-event simulation,' but the evaluation uses actual earthquake source parameters (magnitude, epicenter, depth, distance) from USGS ShakeMap in the prompt. These parameters are not known before an earthquake occurs. The experiments therefore test a post-event nowcast, not a true pre-event simulation. To support the claim, the authors should either evaluate with hypothetical scenario parameters (e.g., plausible ruptures for a planning region) or explicitly reframe the contribution as a rapid post-event assessment tool, adjusting the title and discussion accordingly.
minor comments (5)
  1. [Table 1 and throughout] There are recurring typographical errors, such as 'Ridgcrest' in Table 1 and in the 'Combined Effect' figure caption, which should be corrected.
  2. [Figure 5 caption] The caption for Figure 5 is ambiguous about the axis labels; the text says 'the x-axis is the predicted MMI' in some places and 'close to epi.' in others, making it hard to interpret the distance and VS30 relationships. Please clarify the axes and the direction of the distance axis.
  3. [Appendix D, Data Leakage Test] The leakage test removes only city and state names, but the prompt still contains the earthquake's magnitude, depth, epicenter coordinates, and distances, which are the most informative features for memorization or shortcutting. The conclusion that leakage is ruled out is too strong; a more thorough test should also remove or perturb distance and magnitude or use a held-out earthquake.
  4. [Section 3.5 and Table 3] The RAG/ICL demonstration analysis reports improvements without any uncertainty quantification or statistical significance testing across the two events; the observed differences in RMSE may be within noise. Please report confidence intervals or repeated-sampling variability.
  5. [Section 2.2 and Title] The term 'world model' is used loosely; the framework is essentially a regression-style mapping from input features to an intensity label through an LLM, which does not fit standard definitions of world models that predict future states or dynamics. Please either justify the terminology or temper the framing.

Circularity Check

1 steps flagged · score 5.0 of 10

Main zero-shot LLM-vs-DYFI comparison is independent, but the RAG improvement claim is partly circular because reported MMI labels are inserted into the prompt as retrieved context.

  1. fitted input called prediction [Section 3.5 (Prompting Techniques, RAG) and Figure 3 (Demonstration analysis on RAG and ICL)]
    "RAG: It combines information retrieval with text generation that allows LLMs to access external data sources at inference. We provide LLM with a set of multimodal features and the reported MMI within the prompt as the retrieved context to ground their MMI estimates."

    The 'reported MMI' supplied as RAG context is the same USGS DYFI ground-truth value used to compute RMSE and correlation in the evaluation. When the number of RAG demonstrations k is varied, the LLM's output is directly conditioned on the target label, so the measured decrease in RMSE with increasing k reflects how closely the model copies or interpolates the supplied labels rather than its independent ability to simulate shaking. The paper's conclusion that 'demonstration techniques can enhance the capacity of LLMs' is thus forced by the experimental design for the RAG sub-claim. The main zero-shot Table 1 results do not use RAG and therefore remain an independent comparison.

full rationale

The headline alignment claim is not circular in its central, zero-shot form: the LLM receives feature bundles (distance, VS30, building, socioeconomic, street view) and outputs MMI; the USGS DYFI labels are external ground truth, and no model parameters are fitted to those labels. The paper even includes a leakage test removing city/state names. The RAG experiment, however, supplies the reported DYFI MMI inside the prompt as retrieval context, and then evaluates the model against the same DYFI values; the reported improvement from demonstrations is a by-construction effect of giving the answer as context. This is a genuine but bounded circularity: it affects the RAG/ICL performance claim, not the zero-shot alignment result. The abstract's '0.88 and 0.77 at the zip code level' numbers appear in Table 1 only as county-level metrics (Corr_C=0.88 and RMSE_C=0.77), which is a reporting inconsistency rather than a circularity. The choice of the top-100-by-DYFI-response zip codes and the absence of a distance-decay baseline are validity concerns about generalizability and alternative explanations, but they do not make the prediction equivalent to its input by construction. Overall, the paper has one partially circular sub-claim, while the principal empirical comparison retains independent content, warranting a moderate score rather than a high one.

Assumptions & free parameters 3 free parameters · 5 assumptions · 1 invented entities

The framework rests on externally sourced features and a pretrained LLM, so it contributes no fitted parameters in the usual sense. The hand-chosen design decisions that shape the results are the 100-zip cutoff, the 50-point sampling rate, and the 100-meter building buffer. The key assumptions are that DYFI is a valid ground truth for human perception, that the selected features suffice, that a pretrained LLM can act as a synthetic observer, and that the leakage test is strong enough to rule out memorization. No new physical or ontological entity is introduced; the 'virtual sensor' is a framing device.

free parameters (3)
  • Top-100 zip code cutoff = 100
    Zip codes are ranked by DYFI response count and only the top 100 are evaluated; this hand-chosen cutoff selects high-response areas and affects all reported metrics.
  • Samples per zip code = 50
    Each evaluated zip code is represented by 50 randomly sampled points; changing this number would change the variance of the averaged MMI prediction.
  • Building buffer radius = 100 m
    OpenStreetMap building features are summarized within a 100-meter radius of each sample point; this radius is chosen without sensitivity analysis.
assumptions (5)
  • domain assumption DYFI zip-level MMI is a valid ground-truth measure of human-perceived shaking at the spatial scales used in this study.
    The paper aggregates USGS 'Did You Feel It?' reports into zip code and county MMI labels and uses them as targets; the validity of this aggregation and its comparability to LLM point estimates is assumed.
  • domain assumption The selected features (distance, VS30, building summary, census data, street view) are sufficient to estimate perceived MMI.
    These features define the prompt, and the paper cites prior literature for their relevance, but it does not perform feature selection or show that the set is complete.
  • ad hoc to paper A pretrained LLM can act as a synthetic observer whose text output maps onto the MMI scale.
    This is the core capability being tested; the paper assumes the LLM's reasoning and rating are meaningful proxies for human perception.
  • ad hoc to paper The source parameters supplied in the prompt (epicenter, magnitude, depth, distance) are consistent with a pre-event simulation.
    The experiments use the true source parameters of past earthquakes, which would not be known before an event, and the prompt sets the event date to 2025-06-01 for both historical cases.
  • ad hoc to paper The leakage test is sufficient to rule out memorization of the two earthquakes.
    The test in Appendix D removes only city and state names while keeping epicenter coordinates, magnitude, depth, and distance, so it does not fully control for the model recognizing the events.
invented entities (1)
  • LLM as virtual sensor / synthetic observer
    purpose: Conceptual frame for treating the LLM's output as a simulated human perception of shaking.
    This is a metaphor for the LLM's role, not a concrete entity with falsifiable handles; its validity rests entirely on the empirical correlation results.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LLMs as World Models: Data-Driven and Human-Centered Pre-Event Simulation for Disaster Impact Assessment." pith.science (2026). https://pith.science/paper/4FEZO2W7

@misc{pith2026250606355,
  author       = {Pith},
  title        = {Pith review of: LLMs as World Models: Data-Driven and Human-Centered Pre-Event Simulation for Disaster Impact Assessment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4FEZO2W7}},
  note         = {Machine review of arXiv:2506.06355}
}
read the original abstract

Efficient simulation is essential for enhancing proactive preparedness for sudden-onset disasters such as earthquakes. Recent advancements in large language models (LLMs) as world models show promise in simulating complex scenarios. This study examines multiple LLMs to proactively estimate perceived earthquake impacts. Leveraging multimodal datasets including geospatial, socioeconomic, building, and street-level imagery data, our framework generates Modified Mercalli Intensity (MMI) predictions at zip code and county scales. Evaluations on the 2014 Napa and 2019 Ridgecrest earthquakes using USGS ''Did You Feel It? (DYFI)'' reports demonstrate significant alignment, as evidenced by a high correlation of 0.88 and a low RMSE of 0.77 as compared to real reports at the zip code level. Techniques such as RAG and ICL can improve simulation performance, while visual inputs notably enhance accuracy compared to structured numerical data alone. These findings show the promise of LLMs in simulating disaster impacts that can help strengthen pre-event planning.

Figures

Figures reproduced from arXiv: 2506.06355 by the authors.

Figure 1
Figure 1. An illustration of LLM-simulated human-centric sensors. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. An illustration of the framework design. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Demonstration analysis on RAG and ICL. LLMs deliver promising simulation results. As shown in [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Input information feature analysis results. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Output reasoning analysis in terms of (A) distance (where the x-axis is the predicted MMI, and y-axis is [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Output reasoning analysis in terms of different input features with (A) Qwen-2.5-32B for the Napa case, [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Spatial distribution of predicted MMI at zip code level: (A) The 2014 Napa earthquake, and (B) the 2019 [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. An Agentic Interface for End-to-End Probabilistic Seismic Hazard and Risk Analysis

    physics.geo-ph 2026-06 conditional novelty 6.0 of 10

    An MCP-based LLM-orchestrated server provides end-to-end European seismic hazard and risk calculations, reproducing official ESHM20 475-year spectra to a median of about 5%.

Reference graph

Works this paper leans on

58 extracted references · 44 canonical work pages · cited by 1 Pith paper

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    https://www.census.gov/programs-surveys/acs/ American community survey (acs)

    2022. https://www.census.gov/programs-surveys/acs/ American community survey (acs)

    work page 2022

  4. [4]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, and 1 others. 2023. https://arxiv.org/abs/2303.08774 Gpt-4 technical report . ArXiv preprint, abs/2303.08774

    arXiv 2023

  5. [5]

    Temitope Akinboyewa, Huan Ning, M Naser Lessani, and Zhenlong Li. 2024. Automated floodwater depth estimation using large multimodal model for rapid flood mapping. Computational Urban Science, 4(1):12

    work page 2024

  6. [6]

    did you feel it?

    Gail M. Atkinson and David J. Wald. 2007. https://doi.org/10.1785/gssrl.78.3.362 “did you feel it?” intensity data: A surprisingly good measure of earthquake ground motion . Seismological Research Letters, 78(3):362--368

    work page doi:10.1785/gssrl.78.3.362 2007

  7. [7]

    Jan-Micha Bodensohn, Ulf Brackmann, Liane Vogel, Anupam Sanghi, and Carsten Binnig. 2025. Unveiling challenges for llms in enterprise data engineering. arXiv preprint arXiv:2504.10950

    work page arXiv 2025

  8. [8]

    Angelo Cardellicchio, Sergio Ruggieri, Valeria Leggieri, and Giuseppina Uva. 2023. https://doi.org/10.1016/j.prostr.2023.01.250 A machine learning framework to estimate a simple seismic vulnerability index from a photograph: the vulma project . Procedia Structural Integrity, 44:1956--1963. XIX ANIDIS Conference, Seismic Engineering in Italy

    work page doi:10.1016/j.prostr.2023.01.250 2023

Show all 58 references
  1. [9]

    Gregory G Deierlein, Frank McKenna, Adam Zsarn \'o czay, Tracy Kijewski-Correa, Ahsan Kareem, Wael Elhaddad, Laura Lowes, Matthew J Schoettler, and Sanjay Govindjee. 2020. A cloud-enabled application framework for simulating regional-scale impacts of natural hazards on the bui...

    2020

  2. [10]

    Ivan Frigerio, Stefania Ventura, Daniele Strigaro, Matteo Mattavelli, Mattia De Amicis, Silvia Mugnano, and Mario Boffi. 2016. A gis-based approach to identify the spatial variability of social vulnerability to seismic hazard in italy. Applied geography, 74:12--22

    2016

  3. [11]

    Yanjun Gao, Skatje Myers, Shan Chen, Dmitriy Dligach, Timothy A Miller, Danielle Bitterman, Matthew Churpek, and Majid Afshar. 2024. When raw data prevails: Are large language model embeddings effective in numerical data representation for medical machine learning applications...

    2024 arXiv

  4. [12]

    Google . 2025. Google Maps Platform . https://developers.google.com/maps. Accessed: 2025-05-19

    2025

  5. [13]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, and 1 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948

    2025 arXiv

  6. [14]

    Shibo Hao, Yi Gu, Haodi Ma, Joshua Jiahua Hong, Zhen Wang, Daisy Zhe Wang, and Zhiting Hu. 2023. Reasoning with language model is planning with world model. arXiv preprint arXiv:2305.14992

    2023 arXiv

  7. [15]

    Haley Hostetter, MZ Naser, Xinyan Huang, and John Gales. 2024. Large language models in fire engineering: An examination of technical questions against domain knowledge. arXiv preprint arXiv:2403.04795

    2024 arXiv

  8. [16]

    Iunio Iervolino. 2023. Implications of gmpe’s structure for multi-site seismic hazard. Soil Dynamics and Earthquake Engineering, 172:108022

    2023

  9. [17]

    Rebecca Louise Jones, Debarati Guha-Sapir, and Sandy Tubeuf. 2022. Human and economic impacts of natural disasters: can we trust the global data? Scientific data, 9(1):572

    2022

  10. [18]

    Moustafa Moufid Kassem, Fadzli Mohamed Nazri, and Ehsan Noroozinejad Farsangi. 2020. The seismic vulnerability assessment methodologies: A state-of-the-art review. Ain Shams Engineering Journal, 11(4):849--864

    2020

  11. [19]

    Maja Kucharczyk and Chris H Hugenholtz. 2021. Remote sensing of natural hazard-related disasters with small drones: Global trends, biases, and research opportunities. Remote Sensing of Environment, 264:112577

    2021

  12. [20]

    Geon Lee, Wenchao Yu, Kijung Shin, Wei Cheng, and Haifeng Chen. 2025. Timecap: Learning to contextualize, augment, and predict time series events with large language model agents. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 18082--18090

    2025

  13. [21]

    Bowen Li, Zhaoyu Li, Qiwei Du, Jinqi Luo, Wenshan Wang, Yaqi Xie, Simon Stepputtis, Chen Wang, Katia Sycara, Pradeep Ravikumar, and 1 others. 2024 a . Logicity: Advancing neuro-symbolic ai with abstract urban simulation. Advances in Neural Information Processing Systems, 37:69...

    2024

  14. [22]

    Dawei Li, Bohan Jiang, Liangjie Huang, Alimohammad Beigi, Chengshuai Zhao, Zhen Tan, Amrita Bhattacharjee, Yuxuan Jiang, Canyu Chen, Tianhao Wu, and 1 others. 2024 b . From generation to judgment: Opportunities and challenges of llm-as-a-judge. arXiv preprint arXiv:2411.16594

    2024

  15. [23]

    Lingyao Li, Michelle Bensi, and Gregory Baecher. 2023. Exploring the potential of social media crowdsourcing for post-earthquake damage assessment. International Journal of Disaster Risk Reduction, 98:104062

    2023

  16. [24]

    Lingyao Li, Michelle Bensi, Qingbin Cui, Gregory B Baecher, and You Huang. 2021. Social media crowdsourcing for rapid damage assessment following a sudden-onset natural hazard event. International Journal of Information Management, 60:102378

    2021

  17. [25]

    Zhong-Zhi Li, Duzhen Zhang, Ming-Liang Zhang, Jiaxin Zhang, Zengyan Liu, Yuxuan Yao, Haotian Xu, Junhao Zheng, Pei-Jie Wang, Xiuyi Chen, and 1 others. 2025. From system 1 to system 2: A survey of reasoning large language models. arXiv preprint arXiv:2502.17419

    2025 arXiv

  18. [26]

    Zihui Ma, Lingyao Li, Yujie Mao, Yu Wang, Olivia Grace Patsy, Michelle T Bensi, Libby Hemphill, and Gregory B Baecher. 2024. Surveying the use of social media data and natural language processing techniques to investigate natural disasters. Natural Hazards Review, 25(4):03124003

    2024

  19. [27]

    Rafaela Martelo, Kimia Ahmadiyehyazdi, and Ruo-Qian Wang. 2024. Towards democratized flood risk management: An advanced ai assistant enabled by gpt-4 for enhanced interpretability and public engagement. arXiv preprint arXiv:2403.03188

    2024 arXiv

  20. [28]

    Devin F McPhillips, Julie A Herrick, Sean Ahdi, Alan K Yong, and Scott Haefner. 2020. Updated compilation of vs30 data for the united states. (No Title)

    2020

  21. [29]

    Federico Mori, Amerigo Mendicelli, Massimiliano Moscatelli, Gino Romagnoli, Edoardo Peronace, and Giuseppe Naso. 2020. A new vs30 map for italy based on the seismic microzonation dataset. Engineering Geology, 275:105745

    2020

  22. [30]

    Morgan P Moschetti, Brad T Aagaard, Sean K Ahdi, Jason Altekruse, Oliver S Boyd, Arthur D Frankel, Julie Herrick, Mark D Petersen, Peter M Powers, Sanaz Rezaeian, and 1 others. 2024. The 2023 us national seismic hazard model: Ground-motion characterization for the conterminous...

    2024

  23. [31]

    Hakan T Otal, Eric Stern, and M Abdullah Canbaz. 2024. Llm-assisted crisis management: Building advanced llm platforms for effective emergency response and public collaboration. In 2024 IEEE Conference on Artificial Intelligence (CAI), pages 851--859. IEEE

    2024

  24. [32]

    Frederik Ramm and Jochen Topf. 2010. OpenStreetMap: Die freie Weltkarte nutzen und mitgestalten. Lehmanns Media

    2010

  25. [33]

    James Requeima, John Bronskill, Dami Choi, Richard Turner, and David K Duvenaud. 2024. Llm processes: Numerical predictive distributions conditioned on natural language. Advances in Neural Information Processing Systems, 37:109609--109671

    2024

  26. [34]

    Isma \"e l Riedel, Philippe Gu \'e guen, Mauro Dalla Mura, Erwan Pathier, Thomas Leduc, and Jocelyn Chanussot. 2015. Seismic vulnerability assessment of urban environments in moderate-to-low seismic hazard regions using association rule learning and support vector machine meth...

    2015

  27. [35]

    Argho Sarkar, Tashnim Chowdhury, Robin Roberson Murphy, Aryya Gangopadhyay, and Maryam Rahnemoonfar. 2023. Sam-vqa: Supervised attention-based visual question answering model for post-disaster damage assessment on remote sensing imagery. IEEE Transactions on Geoscience and Rem...

    2023

  28. [36]

    Soudabeh Taghian Dinani, Doina Caragea, and Nikesh Gyawali. 2023. Disaster tweet classification using fine-tuned deep learning models versus zero and few-shot large language models. In International Conference on Data Management Technologies and Applications, pages 73--94. Springer

    2023

  29. [37]

    Geological Survey

    U.S. Geological Survey . 1989. The modified mercalli intensity (mmi) scale. https://pubs.usgs.gov/gip/earthq4/severitygip.html. Accessed: 2024-05-18

    1989

  30. [38]

    Geological Survey

    U.S. Geological Survey . 2014 a . Did you feel it? -- community internet intensity map for 2014 napa earthquake. https://earthquake.usgs.gov/earthquakes/eventpage/nc72282711/dyfi/responses. Accessed: 2024-05-18

    2014

  31. [39]

    Geological Survey

    U.S. Geological Survey . 2014 b . M 6.0 - south napa earthquake, california. https://earthquake.usgs.gov/earthquakes/eventpage/nc72282711/executive. Accessed: 2024-05-18

    2014

  32. [40]

    Geological Survey

    U.S. Geological Survey . 2019 a . Did you feel it? -- community internet intensity map for 2019 ridgecrest earthquake. https://earthquake.usgs.gov/earthquakes/eventpage/ci38457511/dyfi/intensity. Accessed: 2024-05-18

    2019

  33. [41]

    Geological Survey

    U.S. Geological Survey . 2019 b . M 7.1 - 2019 ridgecrest earthquake sequence. https://earthquake.usgs.gov/earthquakes/eventpage/ci38457511/executive. Accessed: 2024-05-18

    2019

  34. [42]

    David J Wald, Bruce C Worden, Vincent Quitoriano, and Kris L Pankow. 2006. Shakemap manual. Technical Manual, users guide, and software guide Version

    2006

  35. [43]

    Chenguang Wang, Davis Engler, Xuechun Li, James Hou, David J Wald, Kishor Jaiswal, and Susu Xu. 2024 a . Near-real-time earthquake-induced fatality estimation using crowdsourced data and large-language models. International Journal of Disaster Risk Reduction, 111:104680

    2024

  36. [44]

    Gelan Wang, Yu Liu, Shukai Liu, Ling Zhang, and Liqun Yang. 2025. Remflow: Rag-enhanced multi-factor rainfall flooding warning in sponge airports via large language model. International Journal of Machine Learning and Cybernetics, pages 1--21

    2025

  37. [45]

    Yue Wang, Tianfan Fu, Yinlong Xu, Zihan Ma, Hongxia Xu, Bang Du, Yingzhou Lu, Honghao Gao, Jian Wu, and Jintai Chen. 2024 b . Twin-gpt: digital twins for clinical trials via large language model. ACM Transactions on Multimedia Computing, Communications and Applications

    2024

  38. [46]

    Gwen White and Sadie Liptak. 2025. Small business continuity and disaster recovery plans using ai and chatgpt

    2025

  39. [47]

    Lionel Wong, Gabriel Grand, Alexander K Lew, Noah D Goodman, Vikash K Mansinghka, Jacob Andreas, and Joshua B Tenenbaum. 2023. From word models to world models: Translating from natural language to the probabilistic language of thought. arXiv preprint arXiv:2306.12672

    2023 arXiv

  40. [48]

    Jiannan Xiang, Tianhua Tao, Yi Gu, Tianmin Shu, Zirui Wang, Zichao Yang, and Zhiting Hu. 2023. Language models meet world models: Embodied experiences enhance language models. Advances in neural information processing systems, 36:75392--75412

    2023

  41. [49]

    Yuwei Yan, Qingbin Zeng, Zhiheng Zheng, Jingzhe Yuan, Jie Feng, Jun Zhang, Fengli Xu, and Yong Li. 2024. Opencity: A scalable platform to simulate urban activities with massive llm agents. arXiv preprint arXiv:2410.21286

    2024 arXiv

  42. [50]

    Hang Yang, Hao Chen, Hui Guo, Yineng Chen, Ching-Sheng Lin, Shu Hu, Jinrong Hu, Xi Wu, and Xin Wang. 2024. Llm-medqa: Enhancing medical question answering through case studies in large language models. arXiv preprint arXiv:2501.05464

    2024 arXiv

  43. [51]

    Shiping Yang, Jie Wu, Wenbiao Ding, Ning Wu, Shining Liang, Ming Gong, Hengyuan Zhang, and Dongmei Zhang. 2025. Quantifying the robustness of retrieval-augmented language models against spurious features in grounding data. arXiv preprint arXiv:2503.05587

    2025 arXiv

  44. [52]

    Shukang Yin, Chaoyou Fu, Sirui Zhao, Yunhang Shen, Chunjiang Ge, Yan Yang, Zuwei Long, Yuhan Dai, Yongdong Luo, Haoyu Cao, and 1 others. 2024. Sparrow: Data-efficient video-llm with text-to-image augmentation. arXiv preprint arXiv:2411.19951

    2024 arXiv

  45. [53]

    Chen Yu and Zhiguo Wang. 2024. Multimodal social sensing for the spatio-temporal evolution and assessment of nature disasters. Sensors, 24(18):5889

    2024

  46. [54]

    Qian Yu, Chaofeng Wang, Frank McKenna, Stella X Yu, Ertugrul Taciroglu, Barbaros Cetiner, and Kincho H Law. 2020. Rapid visual screening of soft-story buildings from street view images using deep learning classification. Earthquake Engineering and Engineering Vibration, 19:827--838

    2020

  47. [55]

    Yiyao Yu, Yuxiang Zhang, Dongdong Zhang, Xiao Liang, Hengyuan Zhang, Xingxing Zhang, Ziyi Yang, Mahmoud Khademi, Hany Awadalla, Junjie Wang, and 1 others. 2025. Chain-of-reasoning: Towards unified mathematical reasoning in large language models via a multi-paradigm perspective...

    2025 arXiv

  48. [56]

    Chenhui Zhang and Sherrie Wang. 2024. Good at captioning bad at counting: Benchmarking gpt-4v on earth observation data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7839--7849

    2024

  49. [57]

    Hengyuan Zhang, Yanru Wu, Dawei Li, Sak Yang, Rui Zhao, Yong Jiang, and Fei Tan. 2024. Balancing speciality and versatility: a coarse to fine framework for supervised fine-tuning large language model. In Findings of the Association for Computational Linguistics ACL 2024, pages...

    2024

  50. [58]

    Zirui Zhao, Wee Sun Lee, and David Hsu. 2023. Large language models as commonsense knowledge for large-scale task planning. Advances in Neural Information Processing Systems, 36:31967--31987

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.