Pith. sign in

REVIEW 3 major objections 5 minor 43 references

Empirically grounding LLM agents in census, time-use, and spatial data makes them statistically credible simulators of population behavior during disruptions, more than doubling the captured share of observed heatwave adaptation compared wi

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 17:56 UTC pith:UIUDECWG

load-bearing objection Worth refereeing for the external heatwave validation, but the normal-day numbers are circular and the survey's representativeness needs checking. the 3 major comments →

arxiv 2607.17437 v1 pith:UIUDECWG submitted 2026-07-19 cs.AI

Empirical Grounding Improves the Realism of LLM Agents Simulating Human Behavior During Disruptions

Chen Xia , Zexi Kuang , Yuqing Hu This is my paper
classification cs.AI
keywords LLM agentsempirical groundinghuman behavior simulationheatwave adaptationactivity-based modelingtime-use dataagent-based simulationexternal validation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Empirically grounding LLM agents in real demographic profiles, baseline daily routines from time-use surveys, and urban spatial context can make them statistically credible simulators of population behavior—both on ordinary days and during a heatwave. The paper tests this against an ungrounded baseline and an independent household survey conducted during the July 2024 Philadelphia heatwave. On normal days, grounding lifts the mean correlation with observed activity profiles from 0.528 to 0.912 and cuts mean squared error from 0.066 to 0.008; under heatwave conditions, correlation rises from 0.349 to 0.836 and error falls from 0.098 to 0.012. The grounded model captures 46.4% of the observed heatwave response amplitude, versus 20.6% for the ungrounded baseline. The point, for a sympathetic reader, is that empirical grounding does not replace LLM reasoning—it anchors it, giving the generative capacity a statistically realistic behavioral baseline from which to adapt.

Core claim

The central claim is that empirical grounding makes LLM agents statistically credible simulators of population behavior under disruption. Instead of letting agents reason from generic needs alone, the authors embed demographic profiles from the American Community Survey, baseline routines from the American Time Use Survey, and urban spatial context into agent initialization, memory, decision prompts, and activity execution. The empirical routines act as a behavioral baseline that agents are instructed to follow unless the environment creates a reason to deviate, so heatwave behavior is modeled as a context-sensitive deviation from a realistic routine. Validated against an independent survey

What carries the argument

The load-bearing mechanism is empirical grounding: embedding American Community Survey demographic profiles and American Time Use Survey daily routines into each agent's initialization, state memory, and decision prompts, with the urban map serving as a spatial constraint layer for activity execution. Baseline routines are generated from time-use diaries using a stochastic Markov-chain procedure with dwell-time constraints, so each demographic group gets a realistic hourly activity-probability profile. During decision-making, the activity-selection prompt weighs 'consistency with typical behavior'—the empirically derived activity probability for the current time slot—as the most important fa

Load-bearing premise

The whole validation rests on the independent heatwave survey—603 self-reported 24-hour recalls from the Philadelphia DMA—being an unbiased picture of what residents actually did during the heatwave; if that survey is skewed, the measured improvement of grounded over ungrounded agents is calibrated against a flawed benchmark.

What would settle it

Re-run the same grounded pipeline against passive mobility records (cellphone location or transit counts) for the same July 2024 heatwave days. If the grounded model's correlation with observed activity profiles no longer beats the ungrounded baseline's, the reported gains are specific to the survey instrument rather than to population behavior.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Empirically grounded LLM agents become a viable generative tool for disruption scenarios with little or no historical precedent, such as compound or novel hazards.
  • The improvement on normal-day routines means the same framework can be used to build baseline activity distributions for urban resilience and infrastructure planning.
  • Because the grounded model still captures only 46.4% of observed heatwave response amplitude, the results imply that realistic disruption simulation needs additional behavioral mechanisms beyond grounding—especially for institutionally scheduled activities and nonessential travel.
  • External validation against an independent survey is established as a necessary condition for using LLM agents as credible population simulators, not just for post hoc checking but as part of the modeling workflow.

Where Pith is reading between the lines

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

  • If the result generalizes, the main bottleneck for LLM-agent social simulation shifts from the language model's reasoning power to the quality of the empirical priors supplied to it; any representative activity-diary dataset could be substituted for the time-use survey.
  • The benchmark is a self-reported 24-hour recall survey, so a natural next test is to re-validate grounded agents against passive mobility records (phone location, transit counts) for the same heatwave; failure to reproduce the 0.836 correlation there would indicate the gain is partly survey-specific.
  • The largest residual deltas in work/study and travel/other point to a testable extension: adding modules for institutional schedules, work flexibility, and indoor thermal conditions should shrink exactly those gaps.
  • Grounding can be viewed as a form of regularization on LLM behavior; if so, the 'consistency with typical behavior' weight is a controllable knob for trading fidelity to the baseline against flexibility in novel scenarios.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes an empirically grounded LLM-agent framework that embeds ACS demographic profiles, ATUS-derived daily routines, and urban spatial context into the AgentSociety architecture, and compares it with an ungrounded baseline under normal conditions and during the July 2024 Philadelphia heatwave. Normal-day reconstruction is evaluated against ATUS-derived baseline activity profiles; heatwave behavior is validated against an independent household survey of 603 Philadelphia DMA respondents. The authors report substantial improvements from grounding: normal-day mean correlation rises from 0.528 to 0.912 and MSE falls from 0.066 to 0.008; heatwave correlation rises from 0.349 to 0.836 and MSE falls from 0.098 to 0.012; the grounded model captures 46.4% of the observed heatwave response amplitude versus 20.6% for the baseline. They conclude that empirical grounding makes LLM agents more statistically credible simulators of population behavior during disruptions.

Significance. If the results hold, the paper makes a useful methodological contribution: it offers a concrete template for combining empirical microdata with LLM flexibility and, importantly, validates the approach against an independent survey rather than only against the data used for grounding. The heatwave case is a relevant stress test, and the experimental design isolates the contribution of grounding while keeping the broader environment fixed. The use of multiple metrics (correlation, MSE, MAE, JSD) and the honest discussion of remaining behavioral gaps are strengths. However, the current evidentiary value is materially weakened by circularity in the normal-day test, insufficient documentation of the survey's representativeness, and a mismatched normal baseline in the adaptation-amplitude analysis. These issues must be addressed before the central claim is fully supported.

major comments (3)
  1. [SI Method S2; Methods 'Empirical grounding of baseline routines'; Table 1] The normal-day reconstruction result appears largely circular. In SI Method S2, the PlanBlock selection prompt explicitly includes 'the ATUS-derived activity probability distribution for the current time slot' and instructs the agent to follow its daily routine unless urgent needs arise. The same ATUS-derived profiles are then used as the benchmark for normal-day validation (Methods, 'Empirical grounding of baseline routines'; Table 1). The high correlation (0.912) and low MSE (0.008) therefore measure how faithfully the prompt-following mechanism reproduces the injected distribution, not whether the agent generates empirically realistic behavior from demographic and contextual cues alone. To support the claim that empirical grounding improves normal-day realism, the authors should validate on a held-out time-use sample not used in prompt construction, or at least ablate the injection of
  2. [Materials and Methods 'Independent heatwave survey benchmark'; Table 1 heatwave rows; Table S2] The heatwave benchmark is the core evidence for the paper's central claim, yet its representativeness is not established. The paper reports only that respondents were adults in the Philadelphia DMA and that 603 responses survived screening. It does not report recruitment channel, response rate, screening criteria, survey dates, or a demographic comparison against ACS benchmarks for the Philadelphia DMA. Self-selection into an online 24-hour recall survey during a heat advisory is plausibly associated with the outcome of interest (e.g., overrepresentation of people remaining at home, underrepresentation of those without internet or working outdoors). If so, the survey-derived heatwave profile is biased and the grounded model's closeness to it (corr 0.836 vs. 0.349) is miscalibrated. Please provide full recruitment details, response rate, screening criteria, and an ACS margin comparison; f
  3. [Results 'Empirical grounding improves representation of heatwave-induced behavioral adaptation'; Fig. 4; Table S3] The 'observed' heatwave response amplitude (mean change magnitude = 0.08156) is computed as the difference between the heatwave survey profiles and ATUS 2023 national normal profiles. This confounds heatwave effects with geographic (Philadelphia vs. national) and temporal (July 2024 vs. 2023) differences. The claim that the grounded model captures 46.4% of the observed response amplitude is therefore not a clean measure of heatwave adaptation. Please provide a Philadelphia-specific normal baseline (e.g., from survey respondents' recall of a typical day, if collected, or from the ATUS Philadelphia subsample if sample size permits) and re-express the amplitude ratio; at minimum, quantify the sensitivity of the 46.4% figure to alternative normal baselines.
minor comments (5)
  1. [Materials and Methods 'Independent heatwave survey benchmark'] Provide the exact survey administration dates and the corresponding National Weather Service heat warning dates, as well as the criteria used in 'quality screening' that yielded the 603 retained responses.
  2. [SI Tables S1, S2, S5] Clarify the mapping between the nine agent groups defined in Table S5 and the five validation groups used for heatwave analysis in Table S2, especially how hybrid vs. in-person work modes are merged into 'employed' groups.
  3. [Figures 2 and 3] Ensure that the figures are legible in the final version and that captions describe all panels; the current text references 'representative' comparisons but the figures contain multiple curves whose labels should be readable at print resolution.
  4. [References] Verify reference details, especially reference 5 (PNAS 123, e2538145123, 2026) and other 2026 citations, to ensure they are correct and not preprint placeholders.
  5. [Main Text, 'Empirical grounding improves representation...'] The sentence 'Full demographic-level change magnitudes and the category-level deltas are reported in SI Appendix, Tables S3 and S4' should read 'Tables S3 and S4' (already correct); check that table numbering is consistent throughout.

Circularity Check

1 steps flagged

Normal-day 'reconstruction' reduces to prompt-following because ATUS profiles are injected into the decision prompt and then used as the benchmark; the heatwave validation remains an independent external test.

specific steps
  1. self definitional [SI Method S2 ('Agent Architecture and Empirical Data Integration'); main-text Methods ('LLM agent simulation framework'); Results ('Empirical grounding improves reconstruction of baseline activity rou]
    "The grounded model differed from the baseline by explicitly including the empirical activity probability for the current time period in the decision context. ... The selection prompt incorporates ... the ATUS-derived activity probability distribution for the current time slot ... The prompt instructs the agent to follow its daily routines unless urgent needs or special circumstances arise."

    The normal-day validation compares simulated activity profiles against ATUS empirical profiles (Results reports mean correlation 0.912), but the same ATUS-derived activity probability distribution is inserted into the PlanBlock prompt at each decision step, and the prompt tells the agent to follow its daily routines unless urgent needs arise. The output is therefore conditioned on the benchmark it is scored against; the 0.912 correlation is substantially a prompt-following score, not an independent reconstruction of population behavior. The paper's own wording ('routine-preservation and reconstruction test') mitigates this, but the metric remains constructed from the target data.

full rationale

The central claim—that empirical grounding improves heatwave realism—is supported by an independent benchmark: the July 2024 Philadelphia heatwave survey was 'used only as an empirical validation benchmark, not as an input for initializing, calibrating, or prompting the agents,' so the heatwave correlations (0.836) and amplitude capture (46.4%) are not circular by construction. The normal-day result, however, compares model outputs to the same ATUS profiles that are embedded in the decision prompt, which is a genuine input-target overlap; the authors acknowledge it as a 'reconstruction test' rather than an external prediction. This partial circularity affects a supporting result, not the main external validation. No load-bearing self-citation chain is present: the Markov-chain citation to the authors' prior work is a methodological choice, and the AgentSociety baseline is externally cited. Score reflects one supporting prediction that reduces by construction, while the central heatwave finding retains independent content.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The paper introduces no new postulated entities. Its free parameters are design choices and empirical inputs. The most notable feature is the circularity between the normal-day validation target (ATUS) and the ATUS data injected into agent prompts. The central heatwave claim relies on the independent survey as ground truth, which is an unvalidated assumption about survey representativeness.

free parameters (5)
  • Need satisfaction thresholds = 0.2 (hunger, energy, safety); 0.3 (social)
    Hand-chosen thresholds in the NeedsBlock that trigger need prioritization; they shape activity decisions and are not derived from the validation data or a stated theory.
  • Plan acceptance score threshold = 0.5
    Hand-chosen threshold in the PlanBlock regeneration loop; affects how many plan candidates are accepted and thus the final activity distributions.
  • Repetition penalty horizon = 4 hours
    Recent activity history window used to penalize short-horizon cycling in guidance selection; chosen by hand.
  • Consistency-with-typical-behavior weight = not quantified ('weighted as the most important factor')
    The prompt weighting of the empirical ATUS distribution against other decision factors is described only qualitatively in SI Method S2, leaving a key tuning parameter undefined.
  • ATUS-derived Markov-chain transition and dwell-time distributions = estimated from 2023 ATUS microdata per 15-min interval and per demographic group
    Empirical inputs the simulation depends on; they are fitted to ATUS data, which is a legitimate data source, but they are fitted quantities and the normal-day validation compares against the same source, inducing circularity.
axioms (5)
  • domain assumption ATUS 2023 weekday routines approximate normal weekday activity in Philadelphia in July 2024
    SI Method S1 selects the 2023 ATUS as the most recent 'new normal' record and maps it to Philadelphia; whether national time-use data transfers to a specific city and year is assumed.
  • domain assumption The self-reported heatwave survey accurately reflects population activity during the heatwave
    The independent survey is the ground truth for heatwave validation; the paper acknowledges recall error and subjective interpretation risks in the Discussion.
  • domain assumption A first-order Markov chain with semi-Markov dwell times adequately generates daily activity sequences
    Used in Method S1 to construct baseline routines from ATUS; it is a simplifying modeling choice that shapes the empirical routines given to agents.
  • domain assumption LLM prompt-following faithfully translates empirical context into agent decisions
    The framework assumes that embedding empirical probability profiles in prompts causes agents to default to those routines; the normal-day test largely measures this behavior rather than independent validity.
  • standard math Standard statistical metrics (Pearson correlation, MSE, JSD) are valid measures of population-level activity realism
    These are standard tools for comparing probability profiles across time and activity categories; no unusual assumptions.

pith-pipeline@v1.3.0-alltime-deepseek · 13364 in / 11402 out tokens · 108773 ms · 2026-08-01T17:56:38.016765+00:00 · methodology

0 comments
read the original abstract

Large language model (LLM) agents offer a generative approach to simulating human behavior under conditions that may have few or no direct historical analogues, a common challenge in disaster and infrastructure-disruption planning. However, this generative capacity creates a validity problem: individually plausible agent reasoning may fail to reproduce empirical population behavior. We evaluate whether empirical grounding improves the statistical realism of LLM-agent simulations during disruptions. Specifically, we develop an empirically grounded LLM-agent framework that embeds demographic profiles from the American Community Survey, baseline routines from the American Time Use Survey, and urban spatial context into agent initialization, memory, decision prompts, and activity execution. An independent household survey conducted during the July 2024 Philadelphia heatwave is reserved as an external validation benchmark. Compared with an ungrounded LLM-agent baseline, the grounded model improved reconstruction of normal daily routines, increasing mean correlation with empirical activity profiles from 0.528 to 0.912 and reducing mean squared error from 0.066 to 0.008. Under heatwave conditions, the grounded model better reproduced survey-derived activity profiles, increasing mean correlation from 0.349 to 0.836 and reducing mean squared error from 0.098 to 0.012. The grounded model captured 46.4% of observed heatwave response amplitude, compared with 20.6% for the ungrounded baseline. These findings show that empirical grounding can make LLM agents more statistically credible simulators of population behavior while revealing remaining gaps in modeling human adaptation during disruptions.

Figures

Figures reproduced from arXiv: 2607.17437 by Chen Xia, Yuqing Hu, Zexi Kuang.

Figure 1
Figure 1. Figure 1: Framework for empirically grounded LLM-agent simulation and validation [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Empirical grounding improves reconstruction of normal-day activity routines [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Empirical grounding improves heatwave activity realism [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗

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

43 extracted references · 4 linked inside Pith

  1. [1]

    J. S. Park, J. C. O’Brien, C. J. Cai, M. R. Morris, P. Liang, M. S. Bernstein, Generative agents: Interactive simulacra of human behavior. Proc. 36th Annu. ACM Symp. User Interface Softw. Technol., Article 2, 1–22 (2023)

    2023

  2. [2]

    C. Gao, X. Lan, N. Li, Y. Yuan, J. Ding, Z. Zhou, F. Xu, Y. Li, Large language models empowered agent-based modeling and simulation: A survey and perspectives. Humanit. Soc. Sci. Commun. 11, 1259 (2024)

    2024

  3. [4]

    C. A. Bail, Can generative AI improve social science? Proc. Natl. Acad. Sci. U.S.A. 121, e2314021121 (2024)

    2024

  4. [5]

    Y. Xie, L. Liang, S. Li, Y. Lu, Z. Xiao, M. Shi, J. Huang, M. Wang, Y. Xie, Evaluating the statistical realism of LLM-generated social science data. Proc. Natl. Acad. Sci. U.S.A. 123, e2538145123 (2026)

    2026

  5. [6]

    Haraguchi, A

    M. Haraguchi, A. Nishino, A. Kodaka, M. Allaire, U. Lall, K.-H. Liao, K. Onda, K. Tsubouchi, N. Kohtake, Human mobility data and analysis for urban resilience: A systematic review. Environ. Plan. B Urban Anal. City Sci. 49, 1507–1535 (2022)

    2022

  6. [7]

    T. Yabe, N. K. W. Jones, P. S. C. Rao, M. C. Gonzalez, S. V. Ukkusuri, Mobile phone location data for disasters: A review from natural hazards and epidemics. Comput. Environ. Urban Syst. 94, 101777 (2022)

    2022

  7. [8]

    Reynaud, Y

    Q. Reynaud, Y. Haradji, F. Sempé, N. Sabouret, Using time-use surveys in multi agent based simulations of human activity. Proc. 9th Int. Conf. Agents Artif. Intell. 1, 67–77 (2017)

    2017

  8. [9]

    Larooij, P

    M. Larooij, P. Törnberg, Validation is the central challenge for generative social simulation: A critical review of LLMs in agent-based modeling. Artif. Intell. Rev. 59, 15 (2026)

    2026

  9. [10]

    Z. Wu, R. Peng, T. Ito, M. Onizuka, C. Xiao, LLM-based social simulations require a boundary. arXiv:2506.19806 [Preprint] (2025). 9

    Pith/arXiv arXiv 2025

  10. [11]

    R. C. Hubal, E. A. Cohen Hubal, Simulating patterns of life: More representative time-activity patterns that account for context. Environ. Int. 172, 107753 (2023)

    2023

  11. [12]

    S. G. Yücel, T. Schwanen, Heatwave adaptation conditioned by everyday life. Glob. Environ. Change 94, 103026 (2025)

    2025

  12. [13]

    Renninger, C

    A. Renninger, C. Cabrera-Arnau, Extreme heat reduces and reshapes urban mobility. PNAS Nexus 5, pgag078 (2026)

    2026

  13. [14]

    Batur, V

    I. Batur, V. O. Alhassan, M. V. Chester, S. E. Polzin, C. Chen, C. R. Bhat, R. M. Pendyala, Understanding how extreme heat impacts human activity-mobility and time use patterns. Transp. Res. Part D Transp. Environ. 136, 104431 (2024)

    2024

  14. [15]

    H. Tian, H. Cai, L. Hu, Y. Qiang, B. Zhou, M. Yang, B. Lin, Unveiling community adaptations to extreme heat events using mobile phone location data. J. Environ. Manage. 366, 121665 (2024)

    2024

  15. [16]

    N. S. Bedi, Q. H. Adams, J. J. Hess, G. A. Wellenius, The role of cooling centers in protecting vulnerable individuals from extreme heat. Epidemiology 33, 611–615 (2022)

    2022

  16. [17]

    A. J. Collins, M. Koehler, C. Lynch, Methods that support the validation of agent-based models: An overview and discussion. J. Artif. Soc. Soc. Simul. 27, 11 (2024)

    2024

  17. [18]

    Windrum, G

    P. Windrum, G. Fagiolo, A. Moneta, Empirical validation of agent-based models: Alternatives and prospects. J. Artif. Soc. Soc. Simul. 10, 8 (2007)

    2007

  18. [19]

    Available at https://www.bls.gov/tus/

    United States Department of Labor, Bureau of Labor Statistics, American Time Use Survey. Available at https://www.bls.gov/tus/. Accessed June 2026

    2026

  19. [20]

    Available at https://www.census.gov/programs-surveys/acs/data/data-via-api.html

    United States Census Bureau, American Community Survey data via API. Available at https://www.census.gov/programs-surveys/acs/data/data-via-api.html. Accessed June 2026

    2026

  20. [22]

    M Ouyang, Review on modeling and simulation of interdependent critical infrastructure systems. Reliab. Eng. Syst. Saf. 121, 43 to 60 (2014)

    2014

  21. [23]

    Bougie, N

    N. Bougie, N. Watanabe, CitySim: Modeling urban behaviors and city dynamics with large-scale LLM- driven agent simulation. Proc. 2025 Conf. Empir. Methods Nat. Lang. Process. Ind. Track, 215–229 (2025)

    2025

  22. [24]

    Bauer, Working from home as an adaptation strategy to heat: Comparing temperatures and workers’ assessments for 203 offices and 107 homes

    A. Bauer, Working from home as an adaptation strategy to heat: Comparing temperatures and workers’ assessments for 203 offices and 107 homes. Build. Environ. 272, 112680 (2025)

    2025

  23. [25]

    C. C. Hampo, L. H. Schinasi, S. Hoque, Surviving indoor heat stress in United States: A comprehensive review exploring the impact of overheating on the thermal comfort, health, and social economic factors of occupants. Heliyon 10, e25801 (2024)

    2024

  24. [26]

    Zscheischler, S

    J. Zscheischler, S. Westra, B. J. J. M. van den Hurk, S. I. Seneviratne, P. J. Ward, A. Pitman, A. AghaKouchak, D. N. Bresch, M. Leonard, T. Wahl, X. Zhang, Future climate risk from compound events. Nat. Clim. Change 8, 469–477 (2018)

    2018

  25. [27]

    Available at https://weatherspark.com

    Weather Spark, July 2024 Weather History in Philadelphia Pennsylvania, United States. Available at https://weatherspark.com. Accessed 6 June 2026

    2024

  26. [31]

    Simulation of daily activity patterns incorporating interactions within households: Algorithm overview and performance

    O. Pribyl, K. G. Goulias, “Simulation of daily activity patterns incorporating interactions within households: Algorithm overview and performance” in Progress in Activity-Based Analysis, H. J. P. Timmermans, Ed. (Elsevier, 2005), pp. 243–266

    2005

  27. [32]

    Available at https://www.openstreetmap.org

    OpenStreetMap contributors, OpenStreetMap. Available at https://www.openstreetmap.org. Accessed June 2026

    2026

  28. [33]

    Available at https://docs.safegraph.com/docs/places

    SafeGraph, Places data schema. Available at https://docs.safegraph.com/docs/places. Accessed June 2026

    2026

  29. [34]

    Zhang, W

    J. Zhang, W. Ao, J. Yan, C. Rong, D. Jin, W. Wu, Y. Li, MOSS: A large-scale open microscopic traffic simulation system. arXiv:2405.12520 [Preprint] (2024)

    Pith/arXiv arXiv 2024

  30. [35]

    Pearson, Note on regression and inheritance in the case of two parents

    K. Pearson, Note on regression and inheritance in the case of two parents. Proc. R. Soc. Lond. 58, 240–242 (1895)

  31. [36]

    R. J. Hyndman, A. B. Koehler, Another look at measures of forecast accuracy. Int. J. Forecast. 22, 679– 688 (2006)

    2006

  32. [37]

    T. Chai, R. R. Draxler, Root mean square error or mean absolute error? Arguments against avoiding RMSE in the literature. Geosci. Model Dev. 7, 1247–1250 (2014)

    2014

  33. [38]

    Agresti, Categorical Data Analysis, 3rd Ed

    A. Agresti, Categorical Data Analysis, 3rd Ed. (Wiley, Hoboken, NJ, 2013). 10

    2013

  34. [39]

    L. J. Cronbach, L. Furby, How we should measure “change”: Or should we? Psychol. Bull. 74, 68–80 (1970)

    1970

  35. [40]

    mandatory

    M. Larooij, P. Törnberg, Do large language models solve the problems of agent-based modeling? A critical review of generative social simulations. arXiv:2504.03274 [Preprint] (2025). 11 Figures and Tables Figure 1. Figure 1. Framework for empirically grounded LLM-agent simulation and validation 12 Figure 2. Figure 2. Empirical grounding improves reconstruc...

    Pith/arXiv arXiv 2025

  36. [41]

    J. I. Dingel, B. Neiman, How many jobs can be done at home? J. Public Econ. 189, 104235 (2020)

    2020

  37. [42]

    J. I. Giménez-Nadal, J. A. Molina, J. Velilla, Work time and well-being for workers at home: Evidence from the American Time Use Survey. Int. J. Manpow. 41, 184–206 (2020)

    2020

  38. [43]

    Eisenberg, A

    Y. Eisenberg, A. Hofstra, E. Twardzik, Quantifying active travel among people with disabilities in the US. Disabil. Health J. 17, 101615 (2024)

    2024

  39. [44]

    Přibyl, K

    O. Přibyl, K. G. Goulias, Simulation of daily activity patterns incorporating interactions within households: Algorithm overview and performance. Transp. Res. Rec. 1926, 135– 141 (2005)

    1926

  40. [45]

    C. Xia, Y. Hu, J. Chen, Community time-activity trajectory modeling based on Markov chain simulation and Dirichlet regression. Comput. Environ. Urban Syst. 100, 101933 (2023)

    2023

  41. [46]

    J. Chen, R. Adhikari, E. Wilson, J. Robertson, A. Fontanini, B. Polly, O. Olawale, Stochastic simulation of occupant-driven energy use in a bottom-up residential building stock model. Appl. Energy 325, 119890 (2022)

    2022

  42. [47]

    W. Wang, T. Osaragi, Generating and understanding human daily activity sequences using time-varying Markov chain models. Travel Behav. Soc. 34, 100711 (2024)

    2024

  43. [48]

    J. Piao, Y. Yan, J. Zhang, N. Li, J. Yan, X. Lan, Z. Lu, Z. Zheng, J. Y. Wang, D. Zhou, C. Gao, F. Xu, F. Zhang, K. Rong, J. Su, Y. Li, AgentSociety: Large-scale simulation of LLM-driven generative agents advances understanding of human behaviors and society. arXiv:2502.08691 [Preprint] (2025)

    Pith/arXiv arXiv 2025