PaperlayerView exact revision on arXiv

Source-linked AI summary

Valuing Time in Silicon: Can Large Language Models Replicate Human Value of Travel Time

Yingnan Yan, Tianming Liu, Yafeng Yin

arXiv:2507.22244v2econ.GNq-fin.EC

TL;DR

Whether LLMs replicate the underlying values and context-sensitive decisions of human travelers remains uncertain. The study estimates the VOT of three LLMs using a factorial experiment and compares their behavior with humans. It finds substantial similarities, alongside model-specific differences that call for validation and calibration.

  • Problem

    Transportation research has rarely examined whether LLMs share travelers’ underlying values, maintain value consistency across contexts, or respond to different time-cost trade-offs.

    Source: §2 · p. 5

  • Method

    The study uses a full factorial design and prompt engineering based on a human VOT study to estimate three LLMs’ valuations and sensitivities, then compares them with humans.

    Source: §1 · p. 3, §5 · p. 20

  • Results

    GPT-4o and Claude-Sonnet-4 show VOT ranges similar to human data, while Gemini-2.5-pro exceeds typical human benchmarks; all three show human-aligned patterns across travel purposes and income.

    Source: §5 · p. 20

  • Takeaways & Limitations

    LLMs offer a non-perfect basis for preliminary screening, hypothesis generation, and synthetic travel data, but their outputs should be validated against target populations before deployment.

    Source: §5 · p. 21

    LLM context sensitivities diverge from humans in some dimensions, particularly income elasticity, making further calibration important for reliable transportation-planning use.

    Source: §5 · p. 21

Abstract

from arXiv · show

As a key advancement in artificial intelligence, large language models (LLMs) are set to transform transportation systems. While LLMs offer the potential to simulate human travelers in future mixed-autonomy transportation systems, their behavioral fidelity in complex scenarios remains largely unconfirmed by existing research. This study addresses this gap by conducting a comprehensive analysis of the value of travel time (VOT) of three popular LLMs. We employ a full factorial experimental design to systematically examine LLMs' sensitivities to various transportation contexts, including the choice setting, travel purpose, and socio-demographic factors. Our results reveal a high degree of behavioral similarity between LLMs and humans. Some LLMs exhibit an aggregate VOT similar to that of humans, and all tested models demonstrate human-like sensitivity to travel purpose, income, and the time-cost trade-off ratios of the alternatives. Furthermore, the behavioral patterns of LLMs are highly consistent across varied contexts. However, while the behavior of every single model is highly robust, we also find some heterogeneity across models regarding the magnitude and direction of sensitivity to travel contexts and income elasticity. Overall, this study provides a foundational benchmark for the future development of LLMs as proxies for human travelers, demonstrating their robust decision-making capabilities while cautioning that misaligned magnitudes of economic trade-offs between humans and LLMs necessitate rigorous validation and additional conditioning of LLMs before their application.

1 Introduction

LLMs may serve as simulated travelers, but whether their underlying decision values align with humans and remain sensitive to travel contexts is unclear. This study compares three LLMs with human travel-time valuation and finds substantial behavioral similarity alongside model-specific differences.

1 Introduction

  • LLMs are proposed as simulated travelers because they can mimic human-generated behavior and adapt decisions to specific situations, but transportation choices involve spatial, mode-availability, and time-cost constraints.Potential uses include autonomous agents and synthetic travel data for hypothesis generation, policy evaluation, and agent-based modeling.Source: §1 · p. 2
  • Existing transportation research largely tests whether LLMs reproduce travelers’ final choices, leaving their underlying values and sensitivity to context and alternative trade-offs largely unexplored.Evidence from other domains is mixed: LLMs sometimes resemble humans but also show deviations and population-representation gaps.Source: §1 · p. 3
  • The three LLMs show substantial VOT similarity to humans, but their context sensitivities—especially income elasticity—differ in magnitude across models.The study presents this comparison as a first systematic analysis of LLM travel-time valuation.Source: §1 · p. 3
  • The study uses factorial design and prompt engineering based on a human VOT study to estimate three LLMs’ VOT and compare their values and context sensitivities with humans.The paper focuses on VOT as a cost-benefit metric and as a time-cost trade-off test of contextual responsiveness.Source: §1 · p. 3

2 Literature Review

Prior work finds human-like behavior from LLMs across transportation and other fields, but also reports mismatches and leaves their underlying travel values and contextual decision patterns unresolved.

2 Literature Review

  • Transportation studies report that LLMs can reproduce travel outcomes, but whether they use the underlying values guiding human decisions remains largely unexplored.Prior work covers destination, activity, trajectory, and mode choices; one study examined ethical dilemmas in autonomous vehicles.Source: §2 · p. 4
  • Studies of LLMs’ economic behavior find human-like strategies and biases alongside departures from human patterns, including stronger reciprocity, fairness, or rationality in some contexts.The reviewed work also reports sensitivity to question wording and status quo bias, while specific choices may still differ from humans.Source: §2 · p. 4
  • Psychology and sociology research reports human-like personalities and survey responses, but also missing human response biases, demographic representation gaps, and less within-group diversity.These findings span personality assessments, instructed attitudes, surveys, and political opinions.Source: §2 · p. 5
  • The review identifies gaps in evidence on LLMs’ VOT, consistency across travel contexts, and responses to alternatives with different time-cost trade-off ratios.These gaps matter because trade-off ratios influence human travel choices and inform traveler-value analysis and planning applications.Source: §2 · p. 5

3 Methodology

The study treats LLMs as stated-preference respondents, varies travel and demographic contexts in a full factorial design, estimates their value of travel time, and compares it with human benchmarks.

3 Methodology

  • The three-phase method designs the stated-preference survey, collects LLM responses by prompting, then estimates VOT and compares it with established human values.The study selects GPT-4o, Gemini-2.5-pro, and Claude-Sonnet-4.Source: §3 · p. 5, Figure 1 · p. 6, §3.2 · p. 9

3.1 Factorial design of SP contexts

  • The full factorial design varies six factors—choice setting, travel purpose, income, gender, age, and education—across 768 combinations.Choice alternatives follow Calfee et al. (2001): 13 alternatives per choice set, with two related choice sets in each package; packages vary average time-cost trade-off ratios.Source: §3.1 · pp. 6, 8, 9

3.2 LLM Response Collection

  • GPT-4o alone completed the full factorial across six choice packages; the other two models were queried only with Choice Package 1, while 60 responses were collected per factor combination for each model and choice set.A unified prompt assigned a sociodemographic profile, specified alternatives, requested rankings of 13 options without repeated ranks and reasons, and used temperature 1.Source: §3.2 · pp. 9–10

3.3 VOT Calculation and Analysis

  • A rank-ordered logit estimates preferences from alternative rankings; VOT is calculated as the marginal rate of substitution between travel cost and time, then regressed on income and traveler characteristics.The log-income coefficient is interpreted as income elasticity.Source: §3.3 · pp. 10–11

4 Results

Across the tested models, VOT often follows human-like patterns, but its magnitude and income sensitivity differ from human estimates and vary by model. Those patterns remain broadly stable across travel contexts and choice settings.

4.1 The Aggregated VOT of LLMs

  • GPT-4o’s mean VOT ($7.92/h) lies within the inflation-adjusted human range ($6.02/h–$10.56/h), whereas Claude-Sonnet-4’s ($5.80/h) is below it; nevertheless, both models’ VOTs exhibit ranges similar to humans.Gemini-2.5-pro’s VOT falls within the human range only at the lowest income level; its VOTs at higher income levels are much greater than human results.Source: §4.1 · pp. 12–13

4.2 Sensitivity to Travel Context

  • All three models show statistically significant positive income coefficients and negative personal- and leisure-travel coefficients, matching human patterns; high-school education coefficients are also negative and significant.These coefficients indicate higher VOT with higher assigned income, for business rather than personal or leisure travel, and for higher rather than lower education.Source: §4.2 · p. 13, Table 3 · p. 14
  • Across models, VOT rises with income and is higher for commuting and business than leisure and personal travel, while the commute–business ordering differs by model.Higher education consistently corresponds to higher VOT, whereas age and gender distinctions vary across models.Source: §4.2 · pp. 13–14, Figure 4 · p. 15, Figure 5 · p. 16

4.3 Sensitivity to Choice Setting

  • Across choice packages, GPT-4o’s VOT rises with income and with the alternatives’ average trade-off ratio; its response magnitude falls within reported human ranges.Human VOT increases by 60–233% when the trade-off level doubles and by 198–340% when it quintuples; GPT-4o’s response falls within these ranges across income levels.Source: §4.3 · p. 15, (b) Gender Group Analysis · p. 17, Figure 7 · p. 17
  • GPT-4o retains consistent sensitivity across choice packages: income, personal and leisure travel, and gender retain coefficient signs and significance, while some other effects are weaker.Age-group and commuting coefficients are significant in only some estimations; significant coefficients for other variables retain their signs.Source: (b) Gender Group Analysis · pp. 17–19, Table 4 · p. 18
  • GPT-4o’s income elasticities remain below 0.3 across choice packages, below human levels and close across travel purposes, indicating weaker income sensitivity.The authors conclude that calibration remains necessary across choice-package designs to match the magnitude of human economic trade-offs.Source: (b) Gender Group Analysis · p. 19, Figure 8 · p. 20

5 Conclusion

The study compares three LLMs’ value of travel time (VOT) with human benchmarks, finding human-like patterns alongside important model-specific differences. Its conclusions support cautious, validated use rather than treating any model as a universal proxy for travelers.

5 Conclusion

  • The experiment treats GPT-4o, Gemini-2.5-pro, and Claude-Sonnet-4 as survey respondents, assigning personas and travel contexts and analyzing responses with rank-ordered logit and linear regression.The full factorial design varies trip purpose and choice settings with different time-cost trade-offs, and the study compares model consistency and sensitivities with humans.Source: §5 · p. 20
  • GPT-4o and Claude-Sonnet-4 show VOT ranges similar to human data, while Gemini-2.5-pro significantly exceeds typical human benchmarks.The authors therefore call for suitable model selection and careful behavioral validation before applications such as synthetic data generation.Source: §5 · p. 20
  • All three LLMs show stable increases in VOT with income and higher VOT for business travel and commuting, with purpose and income effects consistent across choice settings.These patterns align with observations on human travelers and are described as coherent rather than random.Source: §5 · pp. 20–21
  • For GPT-4o, time-cost trade-off ratios positively affect VOT, and the authors suggest using LLMs to generate initial hypotheses where human data are costly or impractical to collect.This sensitivity extends beyond socio-demographic factors to a decision-making principle.Source: §5 · p. 21
  • LLM income elasticities diverge from humans across travel purposes, limiting the models’ ability to capture travelers’ economic sensitivity to income.The authors identify further calibration as critical for behavioral fidelity in transportation planning.Source: §5 · p. 21
  • Different LLMs can show opposite sensitivities and align with conflicting human subgroups, so no single model represents the full heterogeneity of travelers.The authors recommend validating each model against its target population before deployment.Source: §5 · p. 21
  • The comparison is limited to three closed-source LLMs and U.S. human data, while the human benchmark may predate shifts in travel behavior despite monetary inflation adjustment.The authors caution that behavioral change may constrain generalization of the comparative baselines and recommend validation and additional conditioning before quantitative applications.Source: §5 · pp. 21–22

Declaration of Generative AI and AI-assisted technologies in the writing process

The authors used Google Gemini for language editing, then reviewed and edited the content and retained responsibility for the publication.

Declaration of Generative AI and AI-assisted technologies in the writing process

Related papers

Loading 2507.22244v2…