Aligning LLM with human travel choices: a persona-based embedding learning approach

Manzi Li; Tianming Liu; Yafeng Yin

REVIEW 4 major objections 5 minor 68 references

Aligning LLM with human travel choices: a persona-based embedding learning approach

T0 review · 4 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 A persona-loading embedding trained with a Monte-Carlo stochastic EM algorithm aligns a frozen LLM with human travel choices well enough to beat MNL and existing LLM baselines on both aggregate mode shares and individual choices.

desk verdict A genuinely novel alignment idea, but the headline comparison is compromised by a record-level train/test split that lets the same traveler appear in both training and test. read the letter →

arxiv 2505.19003 v1 pith:JZY2WUQG submitted 2025-05-25 cs.AI

Tianming Liu , Manzi Li , Yafeng Yin This is my paper

classification cs.AI

keywords traveldemandmodelinglargelanguagemodelalignmentpersona-basedpromptingembeddinglearningdiscretechoicemodepredictionSwissmetrostochasticexpectation-maximization

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

The paper tries to establish that a frozen LLM can be aligned with human travel choice behavior without fine-tuning, using a persona-based prompting scheme where the persona itself is learned from data. The key claim is that this method beats both classic discrete choice models (multinomial logit) and existing lightweight LLM alignment methods (zero-shot, few-shot, and same-group persona loading) on the Swissmetro dataset, achieving the lowest Jensen–Shannon divergence from true mode shares and the highest macro and weighted F1 scores for individual predictions. If true, it offers an accessible, low-cost route to LLM-based travel behavior simulation that uses only typical survey data: a small detailed panel for persona inference and a larger sparse sample for training the loading function. The authors further claim the learned embeddings are interpretable, revealing behavior clusters by age, gender, and car-versus-rail usage.

What carries the argument

The central object is the persona loading function $P(Z_k|d_i) = p(s_{i,k}, s_{i,-k})$, where $s_{i,k}$ is a cosine similarity between learnable embeddings $e_i = e(d_i;\beta)$ of socio-demographics, and $p$ is a weighted softmax with temperature $\lambda = 40/3$. The estimation engine is a Monte-Carlo stochastic EM algorithm (Algorithm 1) that alternates an E-step sampling $L$ personas per training record and an M-step updating $\beta$ through a regularized weighted log-likelihood (Equation 21). This machinery converts the intractable problem of optimizing a prompt generator over open-ended text into a tractable maximum-likelihood problem over a low-dimensional embedding space, while keeping the LLM's weights frozen.

What would settle it

Re-run the persona inference step multiple times for the same 250 Swissmetro respondents and check whether the inferred 1–10 ratings are stable; then train the loading function with personas randomly permuted across respondents. If permuted personas achieve the same Jensen–Shannon divergence (0.021) and F1 scores as the original ones, or if repeated inference changes the learned embeddings materially, the claim that the personas carry the travelers' true preference information would be refuted.

Watch

Extended reading notes

Core claim

The central discovery is that the LLM alignment problem can be split into two learnable stages. First, an expert LLM infers a textual persona for each respondent in a detailed dataset, using that respondent's socio-demographics and all of their observed context–choice pairs, rating the traveler's values on time, cost, flexibility, habit, comfort, and trip purpose on a 1–10 scale. Second, a persona loading function maps any new traveler's socio-demographics through learnable embeddings to a distribution over these inferred personas, using cosine similarity and a weighted softmax; the embedding parameters are fit by a Monte-Carlo stochastic EM algorithm that maximizes the probability that the frozen LLM, prompted with a sampled persona, reproduces observed choices. On a 400-record test set from Swissmetro, the method achieves a Jensen–Shannon divergence of 0.021 against ground-truth mode shares, versus 0.483 for MNL, 0.216 for zero-shot LLM, and 0.044 for same-group persona loading, and it records the highest macro and weighted F1 scores, while the learned embeddings show distinct clusters by age, gender, and mode-user group.

Load-bearing premise

The framework assumes that the personas inferred by an LLM from a traveler's demographics and observed choices are valid labels of that traveler's economic preferences and behavioral traits, so the loading function is trained toward a target that may be noisy, arbitrary, or biased by the LLM's own priors.

Editorial extensions

If this is right

Travel demand models can be built from typical survey data without fine-tuning an LLM; the only requirements are a small panel of repeated choices for persona inference and a sparse set of choice records for training the loading function.
Aggregate mode-share forecasts improve sharply: the method's predicted shares (4.0% train, 51.7% Swissmetro, 44.3% car) are much closer to ground truth (6.0%, 53.3%, 40.7%) than those of MNL or any LLM baseline tested.
The learned embeddings double as an interpretable segmentation tool, exposing behavior clusters such as elders split by car-versus-rail usage without requiring a separate clustering model.
Alignment costs shift from training to inference design, making LLM-based travel behavior simulation feasible with off-the-shelf API models rather than supervised fine-tuning hardware.
Because the persona basis is trained only on a detailed panel, the framework is compatible with fusing data from multiple choice contexts, although the paper itself demonstrates only the Swissmetro mode-choice case.

Reading between the lines

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

An implication the authors leave implicit is that the same two-stage machinery should transfer to other discrete choice settings, such as route choice or departure time, because the embedding space is not tied to the Swissmetro alternatives.
A testable consequence not explored in the paper: the learned loading function should transfer across choice contexts better than MNL coefficients, since embeddings encode latent behavioral similarity rather than alternative-specific utilities.
The paper does not validate the personas as true psychological states; one could run an ablation in which personas are randomly permuted across respondents and check whether the reported JSD and F1 scores collapse, which would reveal whether the persona labels themselves carry the predictive signal.

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.

Desk Editor's Note

read the letter

Colleague,

The one thing to know: the paper's central idea—learning a persona-loading function from behavioral embeddings to condition a frozen LLM—is new and worth engaging with. The method is clearly laid out, and the framing around typical travel demand data (detailed panel D_h, sparse general D_l) is sensible. I also credit the authors for explicitly targeting resource-efficient alignment rather than SFT.

But the strong empirical claim ('significantly outperformed') is not supported as reported. Section 5.1 builds D_l and D_t by sampling individual choice records from the same pool of respondents, not by sampling respondents. With nine records per respondent, many of the 754 respondents left after D_h will contribute records to both D_l and D_t. The proposed method trains its loading function on D_l, so those test records come from people whose other choices were used to tune the model. That is leakage: individual tastes persist across a person's records, so the model can partially memorize the test respondents. The same-group persona baseline does not train on D_l at all, so the comparison is biased in the wrong direction. The paper reports no respondent-level disjointness check, no cluster-robust errors, nothing. This is the load-bearing issue, not the persona validity concern (which is real but secondary).

Beyond that, the evaluation is a single split with no error bars or significance tests, hyperparameters are hand-set without sensitivity analysis, and no code or exact prompts are provided, so the results aren't independently checkable. The interesting interpretation of learned embeddings (Section 6.2) is anecdotal without stability checks.

I want to be fair: the method itself is thoughtful, the stochastic EM is a reasonable way to cut LLM calls, and the idea of borrowing behavioral similarity across socio-demographic groups is promising. The paper deserves a serious referee, but the authors need to redo the split at respondent level and report cluster-level uncertainty before any comparison is credible.

For the reading group I'd say maybe—the method is worth discussing, but the paper is best used to teach evaluation pitfalls. I wouldn't cite the empirical results until they're fixed.

Recommendation: send to peer review with a clear demand for a respondent-level split and proper uncertainty quantification.

Referee Report

4 major / 5 minor

Summary. The paper proposes a persona-based prompt-conditioning framework for aligning a frozen LLM with human travel choice data. Personas are inferred from a small detailed dataset D_h (Eq. 4); a persona-loading function, parameterized by a linear embedding of socio-demographic variables and a softmax over cosine similarities (Eqs. 17-18), is estimated on D_l via a Monte-Carlo stochastic EM algorithm (Algorithm 1). The framework is evaluated on the Swissmetro dataset, comparing aggregate-mode-share JSD and individual-level macro/weighted F1 against MNL, zero-shot, few-shot, and same-group persona baselines (Table 3), and the learned embeddings are interpreted as behavioral clusters (Section 6.2).

Significance. If the reported results were valid, the framework would be a meaningful, resource-efficient alternative to supervised fine-tuning for LLM-based travel behavior simulation, with the additional benefit of interpretable socio-demographic embeddings. The paper's strengths include a clearly specified stochastic EM procedure, the use of an off-the-shelf LLM without fine-tuning, evaluation on a real stated-preference dataset, and comparison against several relevant baselines. However, the significance is conditional on resolving the evaluation-validity issues described below, particularly the record-level train/test construction and the absence of uncertainty quantification.

major comments (4)

[Section 5.1] The train/test construction samples individual choice records rather than respondents: D_l is formed by randomly sampling 200 records and D_t by sampling 400 records from the remaining records. Because each respondent contributes nine records, the same respondent can appear in both D_l and D_t; a rough calculation suggests that on the order of 60-80 of the 400 test records come from respondents whose other records were used for training. Since the persona-loading function is trained on D_l to map (d, X) to Y, these overlapping respondents make test labels correlated with training labels and can inflate both the individual-level F1 and aggregate-share accuracy reported in Table 3. The comparison is especially unfair to the same-group persona baseline, which does not train on D_l. The authors should re-run the experiment with respondent-disjoint splits and report the number of overlapping respondents, or provide cluster-level standard errors.
[Section 6.1, Table 3] The central performance claim rests on a single split with one run per model. No standard errors, confidence intervals, or significance tests are provided, and the margins over the same-group persona baseline are modest (JSD 0.021 vs 0.044; macro F1 0.556 vs 0.542; weighted F1 0.683 vs 0.657). The word 'significantly' in the abstract and in Section 6.1 is therefore not supported. Please provide repeated random splits, bootstrap confidence intervals, or a paired significance test over test records (or over respondents).
[Section 5.2, Eqs. (19)-(21)] The method's performance depends on several hand-set hyperparameters: the softmax temperature lambda=40/3, the easy-sample downweighting alpha_e=0.5, the regularization strength alpha_m=0.4, and the initial Monte-Carlo sample size L0 and its increment. No sensitivity analysis is reported, and the choices are not derived from data or from a theoretical criterion. Given the small absolute differences in Table 3, the reported gains could be specific to this configuration. Please report sensitivity over a grid of these values, or justify them with a validation-based selection procedure.
[Section 4.1, Eq. (4)] The persona inference step uses the same GPT-4o model that later performs the simulated choices, and the inferred personas are treated as ground-truth labels for the persona-loading function without any validation. If the personas are arbitrary or reflect the LLM's own priors rather than human preferences, the learned loading function may align the simulator with itself rather than with humans. Please validate the personas (e.g., against established attitude scales), test stability across repeated LLM inference runs, or report a sanity check using a different LLM for persona inference.

minor comments (5)

[Section 6.1] The prose comparison with MNL states macro F1 'from 0.393 to 0.556' and weighted F1 'from 0.527 to 0.683', but Table 3 reports MNL macro F1 = 0.474 and weighted F1 = 0.606; the prose should match the table.
[Section 6.1] For the few-shot LLM comparison, the text reports a JSD decrease 'from 0.216 to 0.021' and a weighted F1 increase 'from 0.543 to 0.683'; these values are the zero-shot numbers, while Table 3 gives few-shot JSD = 0.108 and weighted F1 = 0.594. Please correct the comparison.
[Algorithm 1] The initial Monte-Carlo sample size L0, the increment rule, and the convergence criterion are not specified, which prevents exact reproduction of the estimation procedure.
[Section 6.2] The text refers to 'age above 65' and later to 'age above 66' inconsistently; Table 1 defines the final age category as 'above 65'. Please harmonize the terminology.
[Figures 7 and 8] In the manuscript version provided to me, the text labels in Figures 7 and 8 appear as garbled 'uni000...' tokens, making the embedding values and cluster labels unreadable. Please ensure the figure text is rendered correctly in the final version.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the persona-loading model is trained on D_l and evaluated on a separately constructed D_t, so the reported gains are measured, not fitted.

full rationale

The central claim (Section 6.1, Table 3) is an empirical comparison on Swissmetro. The derivation chain is: persona inference (Eq. 4) uses D_h to create a persona basis; the loading function (Eq. 10) is estimated by maximizing the D_l likelihood (Eqs. 12-13, Algorithm 1); test metrics (JSD, macro/weighted F1) are computed on D_t, which is sampled only after D_h and D_l are removed. No equation defines the test prediction in terms of the test labels. The learned embedding parameters are fitted to D_l, and the D_t predictions are generated by the frozen LLM with sampled personas, so the reported 0.021 JSD and F1 scores are not fitted values renamed as predictions. Self-citations to Liu et al. (2024c) are used to set the zero-shot/few-shot baselines and for an interpretive aside about GPT-4's travel-time weighting; they are not load-bearing for the proposed framework's derivation. The persona inference by the same GPT-4o model used for simulation could create self-consistent biases, but this is an empirical validity concern, not a logical circularity. Likewise, the record-level (rather than respondent-level) construction of D_l and D_t creates a potential data-leakage/correctness threat, but leakage is a statistical artifact, not a definitional equivalence; it does not make the test prediction equal to a training input by construction. The paper is self-contained against external benchmarks (MNL, zero-shot, few-shot, same-group persona), and no uniqueness theorem or author-imported ansatz is required to force the result. Accordingly, no circular step is identified.

Assumptions & free parameters 5 free parameters · 6 assumptions · 2 invented entities

The framework rests on several assumptions: that LLM-inferred personas are valid, that embedding similarity reflects behavioral affinity, and that the softmax loading form is appropriate. The free parameters are mostly hand-set hyperparameters plus the learned embeddings. The personas and embedding space are invented latent constructs without independent external evidence.

free parameters (5)

lambda (softmax temperature) = 40/3
Set by hand in Section 5.2 to control the sharpness of the persona loading distribution; not tuned or justified.
alpha_e (easy-sample downweighting) = 0.5
Set by hand in Section 5.2, Equation 19, to reduce weight of all-correct observations.
alpha_m (regularization strength) = 0.4
Set by hand in Section 5.2, Equation 21, to balance variance regularization.
embedding parameters beta_1 to beta_4 = Learned values shown in Figure 7
Model parameters fitted to D_l via the stochastic EM algorithm; these form the core of the persona loading function.
initial Monte-Carlo sample size L0 and increment = Not reported
Algorithm 1 starts with L0 and increments by 1 each iteration; the paper does not give the value, which affects training behavior.

assumptions (6)

domain assumption An LLM conditioned with a persona inferred from observed choices simulates human travel choices more accurately than direct prompting.
Core premise of the framework; only tested on one dataset (Swissmetro).
domain assumption Behavioral similarity between socio-demographic groups is captured by cosine similarity between learned embeddings.
Equations 7 to 10 assume this without independent validation.
domain assumption The expert LLM's persona inference (Equation 4) produces valid labels of traveler preferences.
Personas are LLM-generated, not ground truth; validity is assumed and not checked.
ad hoc to paper The softmax form of the persona loading function (Equation 10) is appropriate.
No derivation from behavioral theory; chosen for computational convenience.
ad hoc to paper The modified EM weights (Equation 19) and regularization (Equation 21) improve generalization.
Heuristic corrections without theoretical justification.
domain assumption D_h and D_l are drawn from the same population.
Sampled from the same Swissmetro dataset, but in general this needs support.

invented entities (2)

persona
purpose: Latent description of a traveler's economic preferences and behavioral traits used to condition LLM prompts
Personas are generated by the same LLM (Equation 4) and are not validated against any external measure.
behavioral embedding space
purpose: Latent space of socio-demographic groups used to compute similarity for persona loading
The embedding function e(d; beta) is learned from data and not tied to an observable quantity.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Aligning LLM with human travel choices: a persona-based embedding learning approach." pith.science (2026). https://pith.science/paper/JZY2WUQG

@misc{pith2026250519003,
  author       = {Pith},
  title        = {Pith review of: Aligning LLM with human travel choices: a persona-based embedding learning approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JZY2WUQG}},
  note         = {Machine review of arXiv:2505.19003}
}

read the original abstract

The advent of large language models (LLMs) presents new opportunities for travel demand modeling. However, behavioral misalignment between LLMs and humans presents obstacles for the usage of LLMs, and existing alignment methods are frequently inefficient or impractical given the constraints of typical travel demand data. This paper introduces a novel framework for aligning LLMs with human travel choice behavior, tailored to the current travel demand data sources. Our framework uses a persona inference and loading process to condition LLMs with suitable prompts to enhance alignment. The inference step establishes a set of base personas from empirical data, and a learned persona loading function driven by behavioral embeddings guides the loading process. We validate our framework on the Swissmetro mode choice dataset, and the results show that our proposed approach significantly outperformed baseline choice models and LLM-based simulation models in predicting both aggregate mode choice shares and individual choice outcomes. Furthermore, we showcase that our framework can generate insights on population behavior through interpretable parameters. Overall, our research offers a more adaptable, interpretable, and resource-efficient pathway to robust LLM-based travel behavior simulation, paving the way to integrate LLMs into travel demand modeling practice in the future.

Figures

Figures reproduced from arXiv: 2505.19003 by the authors.

**Figure 2.** Illustration of the persona inference process [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗

**Figure 3.** Illustration of the persona loading function [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Illustration of the model estimation process [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]

**Figure 5.** Figure 5: Comparison of predicted mode shares proposed method shows the least. The other LLM-based benchmarks produce similar underestimations for this mode. Regarding the Swissmetro mode, the MNL model overestimates its share while underestimating that of the car. The zero-sho…

**Figure 6.** Figure 6: Comparison of confusion matrices for all models [PITH_FULL_IMAGE:figures/full_fig_p024_6.png]

**Figure 7.** Figure 7: , which signal the positional proximity of the embedding projections in their respective dimensions. 40 30 20 10 0 10 20 30 40 Learned embedding values Female Male age<=24 age_25-39 age_40-54 age_55-65 age>=66 income under 50k income 50-100k income over 100k train user…

**Figure 8.** Figure 8: Clusters of socio-demographic groups in the embedding space [PITH_FULL_IMAGE:figures/full_fig_p026_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

68 extracted references · 36 canonical work pages

[1]

L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al

Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al. (2023). Gpt-4 technical report. arXiv preprint arXiv:2303.08774

arXiv 2023
[2]

S., Chien, J., D \' az, M., El-Sayed, S., Pittman, J., Mohamed, S., and McKee, K

Agnew, W., Bergman, A. S., Chien, J., D \' az, M., El-Sayed, S., Pittman, J., Mohamed, S., and McKee, K. R. (2024). The illusion of artificial inclusion. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems , pages 1--12

work page 2024
[3]

V., Arriaga, R

Aher, G. V., Arriaga, R. I., and Kalai, A. T. (2023). Using large language models to simulate multiple humans and replicate human subject studies. In International Conference on Machine Learning , pages 337--371. PMLR

work page 2023
[4]

R., Liu, R., Richardson, S

Anthis, J. R., Liu, R., Richardson, S. M., Kozlowski, A. C., Koch, B., Evans, J., Brynjolfsson, E., and Bernstein, M. (2025). Llm social simulations are a promising research method. arXiv preprint arXiv:2504.02234

arXiv 2025
[5]

The Claude 3 model family: Opus , Sonnet , Haiku

Anthropic (2024). The Claude 3 model family: Opus , Sonnet , Haiku . Available at https://www-cdn.anthropic.com/de8ba9b01c9ab7cbabf5c33b80b7bbc618857627/Model_Card_Claude_3.pdf

work page 2024
[6]

P., Busby, E

Argyle, L. P., Busby, E. C., Fulda, N., Gubler, J. R., Rytting, C., and Wingate, D. (2023). Out of one, many: Using language models to simulate human samples. Political Analysis , 31(3):337--351

work page 2023
[7]

L., and Pereira, F

Arkoudi, I., Krueger, R., Azevedo, C. L., and Pereira, F. C. (2023). Combining discrete choice models and neural networks through embeddings: Formulation, interpretability and performance. Transportation research part B: methodological , 175:102783

work page 2023
[8]

Beck, T., Schuff, H., Lauscher, A., and Gurevych, I. (2024). Sensitivity, performance, robustness: Deconstructing the effect of sociodemographic prompting. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers) , pages 2589--2615

work page 2024

Show all 68 references

[9]

Bierlaire, M., Axhausen, K., and Abay, G. (2001). The acceptance of modal innovation: The case of swissmetro. In Swiss transport research conference

2001
[10]

K., \'E ltet o , N., et al

Binz, M., Akata, E., Bethge, M., Br \"a ndle, F., Callaway, F., Coda-Forno, J., Dayan, P., Demircan, C., Eckstein, M. K., \'E ltet o , N., et al. (2024). Centaur: a foundation model of human cognition. arXiv preprint arXiv:2410.20268

2024 arXiv
[11]

D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al

Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. (2020). Language models are few-shot learners. Advances in neural information processing systems , 33:1877--1901

2020
[12]

Chamieh, I., Zesch, T., and Giebermann, K. (2024). Llms in short answer scoring: Limitations and promise of zero-shot and few-shot approaches. In Proceedings of the 19th workshop on innovative use of nlp for building educational applications (bea 2024) , pages 309--315

2024
[13]

Chen, C., He, Y., Wang, H., Chen, J., and Luo, Q. (2024a). Delayptc-llm: Metro passenger travel choice prediction under train delays with large language models. arXiv preprint arXiv:2410.00052

2024 arXiv
[14]

Chen, J., Wang, X., Xu, R., Yuan, S., Zhang, Y., Shi, W., Xie, J., Li, S., Yang, R., Zhu, T., et al. (2024b). From persona to personalization: A survey on role-playing language agents. arXiv preprint arXiv:2404.18231

2024 arXiv
[15]

Chen, R., Wang, C., Sun, Y., Zhao, X., and Xu, S. (2025). From perceptions to decisions: Wildfire evacuation decision prediction with behavioral theory-informed llms. arXiv preprint arXiv:2502.17701

2025 arXiv
[16]

X., Shan, Y., and Zhong, S

Chen, Y., Liu, T. X., Shan, Y., and Zhong, S. (2023). The emergence of economic rationality of gpt. Proceedings of the National Academy of Sciences , 120(51):e2316205120

2023
[17]

V., Yang, S., Shah, D., Hu, J., and Rogers, T

Chuang, Y.-S., Studdiford, Z., Nirunwiroj, K., Goyal, A., Frigo, V. V., Yang, S., Shah, D., Hu, J., and Rogers, T. T. (2024). Beyond demographics: Aligning role-playing llm-based agents using human belief networks. arXiv preprint arXiv:2406.17232

2024 arXiv
[18]

S., Bryan, C

Demszky, D., Yang, D., Yeager, D. S., Bryan, C. J., Clapper, M., Chandhok, S., Eichstaedt, J. C., Hecht, C., Jamieson, J., Johnson, M., et al. (2023). Using large language models in psychology. Nature Reviews Psychology , 2(11):688--701

2023
[19]

Fan, C., Chen, J., Jin, Y., and He, H. (2024). Can large language models serve as rational players in game theory? a systematic analysis. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 38, pages 17960--17967

2024
[20]

Feng, J., Du, Y., Zhao, J., and Li, Y. (2024). Agentmove: Predicting human mobility anywhere using large language model based agentic framework. arXiv preprint arXiv:2408.13986

2024 arXiv
[21]

Gao, C., Lan, X., Li, N., Yuan, Y., Ding, J., Zhou, Z., Xu, F., and Li, Y. (2024). Large language models empowered agent-based modeling and simulation: A survey and perspectives. Humanities and Social Sciences Communications , 11(1):1--24

2024
[22]

and Singh, A

Goli, A. and Singh, A. (2023). Can llms capture human preferences? arXiv preprint arXiv:2305.02531

2023 arXiv
[23]

Gong, L., Lin, Y., Lu, Y., Han, X., Liu, Y., Guo, S., Lin, Y., and Wan, H. (2025). Mobility-llm: Learning visiting intentions and travel preference from human mobility data with large language models. Advances in Neural Information Processing Systems , 37:36185--36217

2025
[24]

C., Christakis, N

Grossmann, I., Feinberg, M., Parker, D. C., Christakis, N. A., Tetlock, P. E., and Cunningham, W. A. (2023). Ai and the transformation of social science research. Science , 380(6650):1108--1109

2023
[25]

Hagendorff, T., Fabi, S., and Kosinski, M. (2023). Human-like intuitive behavior and reasoning biases emerged in large language models but disappeared in chatgpt. Nature Computational Science , 3(10):833--838

2023
[26]

Haj-Yahia, S., Mansour, O., and Toledo, T. (2025). Incorporating domain knowledge in deep neural networks for discrete choice models. Transportation Research Part C: Emerging Technologies , 171:105014

2025
[27]

C., Ben-Akiva, M., and Zegras, C

Han, Y., Pereira, F. C., Ben-Akiva, M., and Zegras, C. (2022). A neural-embedded discrete choice model: Learning taste representation with strengthened interpretability. Transportation Research Part B: Methodological , 163:166--186

2022
[28]

L., Horton, J

Hansen, A. L., Horton, J. J., Kazinnik, S., Puzzello, D., and Zarifhonarvar, A. (2024). Simulating the survey of professional forecasters. Available at SSRN

2024
[29]

Hensher, D. A. and Greene, W. H. (2003). The mixed logit model: the state of practice. Transportation , 30:133--176

2003
[30]

Horton, J. J. (2023). Large language models as simulated economic agents: What can we learn from homo silicus? Technical report, National Bureau of Economic Research

2023
[31]

and Collier, N

Hu, T. and Collier, N. (2024). Quantifying the persona effect in llm simulations. arXiv preprint arXiv:2402.10811

2024 arXiv
[32]

J., Jung, S.-g., and Salminen, J

Jansen, B. J., Jung, S.-g., and Salminen, J. (2023). Employing large language models in survey research. Natural Language Processing Journal , 4:100020

2023
[33]

V., Heldt, T., and Johansson, P

Johansson, M. V., Heldt, T., and Johansson, P. (2006). The effects of attitudes and personality traits on mode choice. Transportation Research Part A: Policy and Practice , 40(6):507--525

2006
[34]

Kim, J., Kovach, M., Lee, K.-M., Shin, E., and Tzavellas, H. (2024). Learning to be homo economicus: Can an llm learn preferences from choice. arXiv preprint arXiv:2401.07345

2024 arXiv
[35]

S., Reid, M., Matsuo, Y., and Iwasawa, Y

Kojima, T., Gu, S. S., Reid, M., Matsuo, Y., and Iwasawa, Y. (2022). Large language models are zero-shot reasoners. Advances in neural information processing systems , 35:22199--22213

2022
[36]

Korinek, A. (2023). Language models and cognitive automation for economic research. Technical report, national Bureau of economic Research

2023
[37]

T., Pereira, F

Lahoz, L. T., Pereira, F. C., Sfeir, G., Arkoudi, I., Monteiro, M. M., and Azevedo, C. L. (2023). Attitudes and latent class choice models using machine learning. Journal of choice modelling , 49:100452

2023
[38]

Li, A., Chen, H., Namkoong, H., and Peng, T. (2025). Llm generated persona is a promise with a catch. arXiv preprint arXiv:2503.16527

2025 arXiv
[39]

Li, X., Huang, F., Lv, J., Xiao, Z., Li, G., and Yue, Y. (2024). Be more real: Travel diary generation using llm agents and individual profiles. arXiv preprint arXiv:2407.18932

2024 arXiv
[40]

Liu, A., Feng, B., Xue, B., Wang, B., Wu, B., Lu, C., Zhao, C., Deng, C., Zhang, C., Ruan, C., et al. (2024a). Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437

2024 arXiv
[41]

Liu, S., Cao, N., Chen, Y., Jiang, Y., and Cong, G. (2024b). nextlocllm: next location prediction using llms. arXiv preprint arXiv:2410.09129

2024
[42]

Liu, T., Li, M., and Yin, Y. (2024c). Can large language models capture human travel behavior? evidence and insights on mode choice. Avaliable at SSRN 4937575

2024
[43]

Liu, T., Yang, J., and Yin, Y. (2024d). Toward llm-agent-based modeling of transportation systems: A conceptual framework. arXiv preprint arXiv:2412.06681

2024 arXiv
[44]

Lu, Y., Huang, J., Han, Y., Bei, B., Xie, Y., Wang, D., Wang, J., and He, Q. (2025). Beyond believability: Accurate human behavior simulation with fine-tuned llms. arXiv preprint arXiv:2503.20749

2025 arXiv
[45]

and Seo, T

Ma, Y. and Seo, T. (2025). Incorporating graph neural network into route choice model. arXiv preprint arXiv:2503.02315

2025
[46]

S., O'Brien, J., Cai, C

Park, J. S., O'Brien, J., Cai, C. J., Morris, M. R., Liang, P., and Bernstein, M. S. (2023). Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology , pages 1--22

2023
[47]

S., Zou, C

Park, J. S., Zou, C. Q., Shaw, A., Hill, B. M., Cai, C., Morris, M. R., Willer, R., Liang, P., and Bernstein, M. S. (2024a). Generative agent simulations of 1,000 people. arXiv preprint arXiv:2411.10109

2024 arXiv
[48]

S., Schoenegger, P., and Zhu, C

Park, P. S., Schoenegger, P., and Zhu, C. (2024b). Diminished diversity-of-thought in a standard large language model. Behavior Research Methods , pages 1--17

2024
[49]

Paulssen, M., Temme, D., Vij, A., and Walker, J. L. (2014). Values, attitudes and travel behavior: a hierarchical latent variable mixed logit model of travel mode choice. Transportation , 41:873--888

2014
[50]

Pereira, F. C. (2019). Rethinking travel behavior modeling representations through embeddings. arXiv preprint arXiv:1909.00154

2019 arXiv
[51]

Sifringer, B., Lurkin, V., and Alahi, A. (2020). Enhancing discrete choice models with representation learning. Transportation Research Part B: Methodological , 140:236--261

2020
[52]

K., and Sahoo, J

Song, Y., Wang, T., Cai, P., Mondal, S. K., and Sahoo, J. P. (2023). A comprehensive survey of few-shot learning: Evolution, applications, challenges, and opportunities. ACM Computing Surveys , 55(13s):1--40

2023
[53]

G., Fung, Y

Sun, C., Yang, K., Reddy, R. G., Fung, Y. R., Chan, H. P., Small, K., Zhai, C., and Ji, H. (2024). Persona-db: Efficient large language model personalization for response prediction with collaborative data refinement. arXiv preprint arXiv:2402.11060

2024 arXiv
[54]

M., Hauth, A., et al

Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.-B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A. M., Hauth, A., et al. (2023). Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805

2023 arXiv
[55]

Tjuatja, L., Chen, V., Wu, T., Talwalkwar, A., and Neubig, G. (2024). Do llms exhibit human-like response biases? a case study in survey design. Transactions of the Association for Computational Linguistics , 12:1011--1026

2024
[56]

Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.-A., Lacroix, T., Rozi \`e re, B., Goyal, N., Hambro, E., Azhar, F., et al. (2023). Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971

2023 arXiv
[57]

Tseng, Y.-M., Huang, Y.-C., Hsiao, T.-Y., Chen, W.-L., Huang, C.-W., Meng, Y., and Chen, Y.-N. (2024). Two tales of persona in llms: A survey of role-playing and personalization. arXiv preprint arXiv:2406.01171

2024 arXiv
[58]

Vij, A., Carrel, A., and Walker, J. L. (2013). Incorporating the influence of latent modal preferences on travel mode choice behavior. Transportation Research Part A: Policy and Practice , 54:164--178

2013
[59]

Wang, J., Jiang, R., Yang, C., Wu, Z., Onizuka, M., Shibasaki, R., and Xiao, C. (2024a). Large language models as urban residents: An llm agent framework for personal mobility generation. arXiv preprint arXiv:2402.14744

2024 arXiv
[60]

Wang, Q., Wang, S., Zheng, Y., Lin, H., Zhang, X., Zhao, J., and Walker, J. (2024b). Deep hybrid model with satellite imagery: How to combine demand modeling and computer vision for travel behavior analysis? Transportation Research Part B: Methodological , 179:102869

2024
[61]

Wang, S., Mo, B., and Zhao, J. (2020a). Deep neural networks for choice analysis: Architecture design with alternative-specific utility functions. Transportation Research Part C: Emerging Technologies , 112:234--251

2020
[62]

Wang, S., Mo, B., and Zhao, J. (2021a). Theory-based residual neural networks: A synergy of discrete choice models and deep neural networks. Transportation research part B: methodological , 146:333--358

2021
[63]

Wang, S., Wang, Q., Bailey, N., and Zhao, J. (2021b). Deep neural networks for choice analysis: A statistical learning theory perspective. Transportation Research Part B: Methodological , 148:60--81

2021
[64]

Wang, S., Wang, Q., and Zhao, J. (2020b). Deep neural networks for choice analysis: Extracting complete economic information for interpretation. Transportation Research Part C: Emerging Technologies , 118:102701

2020
[65]

and Farooq, B

Wong, M. and Farooq, B. (2021). Reslogit: A residual neural network logit model for data-driven choice modelling. Transportation Research Part C: Emerging Technologies , 126:103050

2021
[66]

Yang, Y., Zhang, W., Lin, H., Liu, Y., and Qu, X. (2024). Applying masked language model for transport mode choice behavior prediction. Transportation Research Part A: Policy and Practice , 184:104074

2024
[67]

J., and Ma, Y

Zhai, Y., Tong, S., Li, X., Cai, M., Qu, Q., Lee, Y. J., and Ma, Y. (2023). Investigating the catastrophic forgetting in multimodal large language models. arXiv preprint arXiv:2309.10313

2023 arXiv
[68]

Ziems, C., Held, W., Shaikh, O., Chen, J., Zhang, Z., and Yang, D. (2024). Can large language models transform computational social science? Computational Linguistics , 50(1):237--291

2024

Pith tools

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

[1] [1]

L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al

Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al. (2023). Gpt-4 technical report. arXiv preprint arXiv:2303.08774

arXiv 2023

[2] [2]

S., Chien, J., D \' az, M., El-Sayed, S., Pittman, J., Mohamed, S., and McKee, K

Agnew, W., Bergman, A. S., Chien, J., D \' az, M., El-Sayed, S., Pittman, J., Mohamed, S., and McKee, K. R. (2024). The illusion of artificial inclusion. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems , pages 1--12

work page 2024

[3] [3]

V., Arriaga, R

Aher, G. V., Arriaga, R. I., and Kalai, A. T. (2023). Using large language models to simulate multiple humans and replicate human subject studies. In International Conference on Machine Learning , pages 337--371. PMLR

work page 2023

[4] [4]

R., Liu, R., Richardson, S

Anthis, J. R., Liu, R., Richardson, S. M., Kozlowski, A. C., Koch, B., Evans, J., Brynjolfsson, E., and Bernstein, M. (2025). Llm social simulations are a promising research method. arXiv preprint arXiv:2504.02234

arXiv 2025

[5] [5]

The Claude 3 model family: Opus , Sonnet , Haiku

Anthropic (2024). The Claude 3 model family: Opus , Sonnet , Haiku . Available at https://www-cdn.anthropic.com/de8ba9b01c9ab7cbabf5c33b80b7bbc618857627/Model_Card_Claude_3.pdf

work page 2024

[6] [6]

P., Busby, E

Argyle, L. P., Busby, E. C., Fulda, N., Gubler, J. R., Rytting, C., and Wingate, D. (2023). Out of one, many: Using language models to simulate human samples. Political Analysis , 31(3):337--351

work page 2023

[7] [7]

L., and Pereira, F

Arkoudi, I., Krueger, R., Azevedo, C. L., and Pereira, F. C. (2023). Combining discrete choice models and neural networks through embeddings: Formulation, interpretability and performance. Transportation research part B: methodological , 175:102783

work page 2023

[8] [8]

Beck, T., Schuff, H., Lauscher, A., and Gurevych, I. (2024). Sensitivity, performance, robustness: Deconstructing the effect of sociodemographic prompting. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers) , pages 2589--2615

work page 2024