Pith. sign in

REVIEW 3 major objections 6 minor 66 references

Where You Go is Who You Are: Behavioral Theory-Guided LLMs for Inverse Reinforcement Learning

T0 review · 3 major / 6 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 This paper claims that inversely following the Theory of Planned Behavior — inferring intentions with LLM-guided inverse reinforcement learning, then reasoning through beliefs — predicts sociodemographic attributes from mobility data far…

desk verdict A useful two-stage LLM+IRL pipeline for demographic inference from GPS, but the TPB-causal story is not yet supported; ask for code, confidence intervals, and a latent-control ablation. read the letter →

arxiv 2505.17249 v1 pith:AIWIA4S5 submitted 2025-05-22 cs.AI cs.LG

Yuran Sun , Susu Xu , Chenguang Wang , Xilei Zhao This is my paper
classification cs.AIcs.LG
keywords sociodemographicinferenceinversereinforcementlearninglargelanguagemodelsTheoryofPlannedBehaviorhumanmobilitytravelGPStrajectoriescognitivechainreasoning
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

Big GPS trajectory datasets usually lack demographic labels, and prior attempts to infer gender, age, or income from movement have been inaccurate. SILIC tries to fix this by running the Theory of Planned Behavior in reverse: it first uses LLM-guided inverse reinforcement learning to turn each person's daily activity sequence into a 31-dimensional reward vector representing latent intentions, then prompts an LLM to reason from those intentions through attitude, subjective norm, and perceived behavioral control to a demographic label. On the 2017 Puget Sound household travel survey, the framework reports accuracy of 85.5% for gender and 86.3% for age, a relative gain of about 31% over the strongest baselines, with much better recovery of rare classes such as adults over 65. If the result holds, mobility data alone could be enriched with demographic attributes for transportation planning, policy analysis, and behavioral simulation.

What carries the argument

The machinery is a linear reward function $R(s)=\theta^\top \phi(s)$ over a 31-dimensional state-feature encoding (activity type, hour, first-activity flag, activity count), with $\theta$ inferred by maximum-entropy IRL. The LLM plays two mechanical roles: it initializes $\theta$ from travel diaries to narrow the ill-posed reward space, and, at each iteration, it inspects the top-$K$ mismatched states and emits a $\{-1,0,1\}$ update direction for each weight, blended with the maximum-entropy gradient. The second stage, Cognitive Chain Reasoning, is a prompt structure that forces the LLM to reason through the three TPB belief constructs — attitude, subjective norm, perceived behavioral control — before emitting the demographic label, combining the reward weights with external context such as urbanicity and transit access.

What would settle it

Give SILIC a sample where participants filled out TPB belief questionnaires alongside their GPS diaries. If the LLM-inferred attitudes, subjective norms, and perceived behavioral control do not correlate with the self-reported ones while demographic accuracy stays high, then the theory-guided mechanism is not what is doing the predictive work; the gains would be attributable to the LLM's learned associations.

Watch

Extended reading notes

Core claim

The paper's central claim is that demographic attributes are recoverable from mobility because behavior is generated by a psychological pathway, and that pathway can be walked backwards with the right machinery. The claim is that the TPB chain — background factors shape beliefs, beliefs shape intentions, intentions shape behavior — can be inverted: observed trajectories first yield reward weights (intentions) via an IRL procedure whose initialization and updates are guided by an LLM, and then a Cognitive Chain Reasoning prompt makes the LLM infer the three TPB beliefs from those weights before predicting the demographic label, with contextual variables supplied to disambiguate groups. The reported evidence is that this two-stage inversion outperforms classical classifiers, zero-shot LLMs, and LLMs with plain chain-of-thought on gender, age, employment, and income prediction, with the largest gains in classes the baselines nearly miss.

Load-bearing premise

The whole framework hinges on unobservable intermediate variables: that the 31 reward weights encode genuine intentions and that an LLM can read the three TPB beliefs out of them — yet the paper's own limitations section concedes these internal constructs have no ground truth for direct validation.

Editorial extensions

If this is right

  • Trajectory datasets that currently lack demographic labels could be enriched at scale, giving transportation planners and policy analysts access to attributes that surveys provide only for small samples.
  • Underrepresented groups such as adults over 65 and retired or unemployed individuals would be far better captured than with conventional classifiers, reducing bias in downstream equity analyses.
  • The same inverse TPB pipeline transfers to other behavioral prediction tasks where latent intentions mediate observable choices, since the paper frames the framework as applicable beyond transportation.
  • Learned reward vectors double as interpretable behavioral profiles, so the framework can feed behaviorally grounded agent-based simulations with individual-level intention estimates.

Reading between the lines

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

  • If the accuracy gains actually come from the LLM's prior associations between diary patterns and demographics, the method may degrade when applied to regions or populations whose routines differ from the LLM's training distribution; the paper's regional-specificity limitation is consistent with this risk.
  • A cleaner test of the TPB mechanism would replace the belief-inference step with direct prediction from the same 31 reward weights plus context: if accuracy does not drop, the three belief constructs add nothing beyond prompt structure.
  • The framework sharpens a privacy concern: passive GPS traces are already sensitive, and a method that reads demographics from them at 85%+ accuracy changes what 'anonymous' trajectory data should be assumed to protect.
  • A natural extension is to use SILIC-inferred demographics to label synthetic trajectory generators, then measure whether downstream transportation models trained on the enriched synthetic data match those trained on survey-labeled data.
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

3 major / 6 minor

Summary. The paper introduces SILIC, a two-stage framework for inferring sociodemographic attributes from GPS-style travel diaries. Stage 1 uses GPT-4o to initialize and iteratively update reward weights in an inverse reinforcement learning (IRL) model, claiming to capture latent travel intentions. Stage 2 applies a Cognitive Chain Reasoning (CCR) prompt that first maps the learned reward weights to Theory of Planned Behavior (TPB) constructs (attitude, subjective norm, perceived behavioral control) and then predicts attributes such as gender and age, augmented by contextual features. Experiments on the 2017 Puget Sound Regional Council Household Travel Survey report large gains over classical ML baselines and zero-shot GPT-4o, e.g., gender accuracy 0.855 vs. 0.653 for GPT-4o, with ablations in Tables 3 and 4 showing that both the LLM-guided IRL components and the CCR module contribute to performance.

Significance. If the central claim holds, the paper would be a meaningful step toward combining psychological theory, LLM priors, and IRL for behaviorally grounded attribute inference from passive mobility data. The methodological pipeline is clearly described, the ablation study is helpful, and the use of a real household travel survey with multiple target attributes is a strength. The paper also includes explicit prompt templates, which support reproducibility. However, the significance of the work as a test of TPB-based inverse inference is currently limited because the experimental design does not isolate the TPB mechanism from generic LLM compression and arbitrary latent labels; the paper's own Limitations section concedes that the inferred internal variables lack ground-truth validation.

major comments (3)
  1. [§3.3, §4.3, Table 4] The headline claim that the accuracy gain comes from inversely following the TPB is not supported by the experiments. The CCR ablation compares IRL+CCR against IRL+Inference and IRL+CoT, but CoT also uses multi-step reasoning, so the comparison does not control for the number of reasoning steps or the semantic content of the intermediate constructs. A matched control with the same number of steps and the same reward-weight input but with theory-neutral labels (e.g., 'preferences' or 'constraints') is needed. Without such a control, the improvement could be due to the LLM's better use of compressed diary information, not to the specific TPB constructs. The unimplemented validation is acknowledged in the Limitations section, where the authors state that 'these internal variables lack ground truth for direct validation.'
  2. [§3.2, Eq. (1), Fig. 5] The reward initialization prompt (Fig. 5) asks GPT-4o to read the individual's multi-day travel diaries and output the 31 reward weights. Since those same diaries contain demographic cues (activity timing, activity types, trip chaining), the learned θ is a lossy LLM encoding of the diary text. The later CCR stage then decodes θ back into demographics. This pipeline could be acting as a two-stage compression and reconstruction of the diary, rather than as a theoretically mediated inference. To support the claim that θ captures latent intentions rather than diaristic demographic cues, the authors should include an experiment where θ is initialized from a non-diaristic source or where the diaries are stripped of demographic correlates, and show the accuracy drop.
  3. [§3.2, Eqs. (5)–(7), Table 3] The IRL component is validated only through KL divergence and L1 distance between expert and learner state visitation distributions. Those metrics measure behavioral cloning quality, not cognitive fidelity. Because IRL is ill-posed (as the authors note in §1), many reward functions can match the same expert behavior, and the paper does not demonstrate that the specific θ learned by SILIC is uniquely tied to intentions or TPB beliefs. An independent measure of intention—for example, survey self-reports or stated preference data—is needed, or at least a sensitivity analysis showing that the downstream demographic prediction is robust across reward functions that achieve the same KL/L1 fit.
minor comments (6)
  1. [§3.1] Typographical errors: 'each dairy captures' should be 'each diary captures'; 'This stage is designed' has an awkward capitalization; the Related Work section contains 'exploreed'.
  2. [§4.1] The test set size is not stated explicitly. With 617 qualified individuals and an 80/20 split, the test set is approximately 123 individuals, which is small for the reported accuracy differences; please report the exact test-set counts and add confidence intervals or significance tests.
  3. [Appendix A.6, Eq. (12)] The F1 formula contains a typo: 'P revisionc + Recallc' should be 'Precisionc + Recallc'.
  4. [Appendix A.1] The LLM sampling parameters (e.g., temperature, top-p) are not reported for GPT-4o. Since the method relies on LLM outputs for initialization and updates, these details are important for reproducibility.
  5. [Appendix A.4] The filtering criteria reduce the original 6,254 participants to 617; a flow diagram showing attrition at each step would improve transparency and help readers assess generalizability.
  6. [References] Some references are incomplete, such as 'Sun et al.' without a year, and several entries lack full author lists or publication venues; please standardize the bibliography.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the demographic labels are not used to fit the IRL reward weights, the CCR prompt, or any other trainable component, so the headline prediction is not a fitted value; the TPB-mechanism concern is a validity gap the paper itself concedes.

full rationale

The claimed derivation chain is: observed trajectories are converted to diaries; an LLM initializes individual reward weights from those diaries; maximum-entropy IRL plus LLM-guided updates refine the weights using only expert/learner state-visitation mismatches; and CCR prompts an LLM to infer TPB belief constructs from the weights and then predict demographics. At no point are the demographic labels used as inputs to the reward initialization, the IRL updates, or the CCR prompt. The reward weights are fit per individual from that individual's own trajectories without label supervision, and predictions are made on held-out individuals without using their labels. Therefore the 30.93% gender accuracy improvement is an empirical outcome relative to baselines, not an identity or a fitted parameter renamed as a prediction. The main weakness is that the intermediate 'belief' constructs (attitude, subjective norm, perceived behavioral control) are never independently validated; the paper explicitly acknowledges this in its Limitations section: 'these internal variables lack ground truth for direct validation. Any misalignment in these inferred representations may propagate to downstream sociodemographic predictions.' That is a real threat to the causal claim that TPB structure, rather than the LLM's priors or the extra reasoning steps, drives the gain. However, a confound between mechanism and architecture is not circularity under the definitions used here: the input (trajectories) and output (demographics) are distinct, no fitted parameter equals the predicted label by construction, and no equation reduces to itself. The self-citations (e.g., Chen et al., 2025) are used as general motivation about LLM cognitive inference and contextual ambiguity, not as the load-bearing proof of the paper's result, so they do not raise the circularity score. The honest finding is no significant circularity.

Assumptions & free parameters 4 free parameters · 7 assumptions · 1 invented entities

The central framework rests on a psychological theory (TPB) and a set of LLM heuristics rather than on new physical entities. The main free parameters are IRL hyperparameters; the main unverified assumptions concern the validity of the TPB inversion and the correspondence between reward weights and real mental states.

free parameters (4)
  • alpha (IRL learning rate) = 2
    Hand-set hyperparameter in Eq. 6; the claimed performance depends on it.
  • lambda_LLM (LLM blending weight) = 0.002
    Hand-set in Eq. 6; controls how much LLM update directions affect reward weights.
  • K (number of top mismatched states) = 30
    Prompt input size for LLM reward updates; chosen by the authors (Appendix A.1).
  • epsilon/epsilon' (convergence thresholds) = 1e-4
    Stopping criteria for value iteration and IRL, set in Appendix A.1.
assumptions (7)
  • domain assumption Theory of Planned Behavior correctly describes the causal pathway from background factors to travel behavior.
    Section 3 and Figure 1; the whole inverse reasoning is built on TPB as a valid model of travel decisions, following Ajzen (1991, 2020).
  • domain assumption A linear reward R(s)=theta^T phi(s) over the coarse state features (hour, activity type, first flag, activity count) captures the latent intentions relevant to sociodemographics.
    Eq. 1 and Section 3.1; the MDP state space omits trip distance and other features, and the paper's Limitations acknowledge this.
  • ad hoc to paper LLM-generated reward initializations and -1/0/1 update directions are behaviorally meaningful and improve IRL convergence.
    Section 3.2 and prompts in Appendix B; this is a heuristic assumption not derived from theory or validated against ground-truth rewards.
  • standard math Maximum entropy IRL (Ziebart et al. 2008) is an appropriate stochastic policy model for human travel decisions.
    Eqs. 2-3; standard method from cited literature.
  • domain assumption Individual-specific transition probabilities estimated from one person's observed activity sequences are sufficient for the MDP.
    Section 3.1; with only 2-7 days of GPS data, transition estimates are sparse.
  • domain assumption Contextual attributes and inferred beliefs are sufficient to identify sociodemographic labels.
    Section 3.3 and CCR prompt (Figure 7); for income and employment, contextual features like housing type may dominate.
  • domain assumption LLMs can faithfully perform ToM-based inference of psychological beliefs from reward weights.
    Section 3.3; relies on contested LLM theory-of-mind abilities; paper cites both supporting and skeptical works (Street 2024, Petrov et al. 2024).
invented entities (1)
  • Latent TPB belief constructs (attitude, subjective norm, perceived behavioral control) inferred from reward weights
    purpose: Intermediate variables in CCR that purportedly mediate between learned reward weights and sociodemographic predictions.
    The paper provides no independent measurement of these constructs; it only asks the LLM to produce them from reward vectors. The Limitations concede they 'lack ground truth for direct validation.' While the constructs originate from Ajzen's TPB, their instantiation and inference inside SILIC are postulated ad hoc.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Where You Go is Who You Are: Behavioral Theory-Guided LLMs for Inverse Reinforcement Learning." pith.science (2026). https://pith.science/paper/AIWIA4S5

@misc{pith2026250517249,
  author       = {Pith},
  title        = {Pith review of: Where You Go is Who You Are: Behavioral Theory-Guided LLMs for Inverse Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AIWIA4S5}},
  note         = {Machine review of arXiv:2505.17249}
}
read the original abstract

Big trajectory data hold great promise for human mobility analysis, but their utility is often constrained by the absence of critical traveler attributes, particularly sociodemographic information. While prior studies have explored predicting such attributes from mobility patterns, they often overlooked underlying cognitive mechanisms and exhibited low predictive accuracy. This study introduces SILIC, short for Sociodemographic Inference with LLM-guided Inverse Reinforcement Learning (IRL) and Cognitive Chain Reasoning (CCR), a theoretically grounded framework that leverages LLMs to infer sociodemographic attributes from observed mobility patterns by capturing latent behavioral intentions and reasoning through psychological constructs. Particularly, our approach explicitly follows the Theory of Planned Behavior (TPB), a foundational behavioral framework in transportation research, to model individuals' latent cognitive processes underlying travel decision-making. The LLMs further provide heuristic guidance to improve IRL reward function initialization and update by addressing its ill-posedness and optimization challenges arising from the vast and unstructured reward space. Evaluated in the 2017 Puget Sound Regional Council Household Travel Survey, our method substantially outperforms state-of-the-art baselines and shows great promise for enriching big trajectory data to support more behaviorally grounded applications in transportation planning and beyond.

Figures

Figures reproduced from arXiv: 2505.17249 by the authors.

Figure 1
Figure 1. Overview of our proposed framework. We inversely follow the Theory of Planned Behavior (TPB) to [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Methodological Framework. are first converted into structured travel diaries that record detailed information on daily trips, such as activity types and departure times. These diaries are then represented as sequential activity patterns and modeled within a Markov Decision Process (MDP) framework. We subsequently introduce an LLM-guided inverse reinforcement learning (IRL) framework, in which the LLM provides heuris… view at source ↗
Figure 3
Figure 3. F1 score variation across feature selection [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: F1 score variation across feature selection [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: Prompt used to initialize reward weights for IRL [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]
Figure 6
Figure 6. Figure 6: Prompt template used to generate reward update directions from the LLM. [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 7
Figure 7. Figure 7: CCR prompt for predicting sociodemographic attributes from IRL reward weights and contextual attributes [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

66 extracted references · 46 canonical work pages

  1. [1]

    Stephen Adams, Tyler Cody, and Peter A Beling. 2022. A survey of inverse reinforcement learning. Artificial Intelligence Review, 55(6):4307--4346

    work page 2022

  2. [2]

    I Ajzen. 1991. The theory of planned behavior. Organizational Behavior and Human Decision Processes, 50(2):179--211

    work page 1991

  3. [3]

    Icek Ajzen. 2020. The theory of planned behavior: Frequently asked questions. Human behavior and emerging technologies, 2(4):314--324

    work page 2020

  4. [4]

    Gulsah Akar, Kelly J Clifton, and Sean T Doherty. 2012. Redefining activity types: Who participates in which leisure activity? Transportation research part A: policy and practice, 46(8):1194--1204

    work page 2012

  5. [5]

    Hassan Ali, Philipp Allgeuer, and Stefan Wermter. 2024. Comparing apples to oranges: Llm-powered multimodal intention prediction in an object categorization task. In International Conference on Social Robotics, pages 292--306. Springer

    work page 2024

  6. [6]

    Maryam Amirizaniani, Elias Martin, Maryna Sivachenko, Afra Mashhadi, and Chirag Shah. 2024. Can llms reason like humans? assessing theory of mind reasoning in llms for open-ended questions. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, pages 34--44

    2024

  7. [7]

    Ali Bakhtiari, Hamid Mirzahossein, Navid Kalantari, and Xia Jin. 2023. Inferring socioeconomic characteristics from travel patterns. Journal of Regional and City Planning, 34(1):122--136

    work page 2023

  8. [8]

    Dibyanayan Bandyopadhyay, Soham Bhattacharjee, and Asif Ekbal. 2025. Thinking machines: A survey of llm based reasoning strategies. arXiv preprint arXiv:2503.10814

    arXiv 2025

Show all 66 references
  1. [9]

    Haoyang Cao, Samuel Cohen, and Lukasz Szpruch. 2021. Identifiability in inverse reinforcement learning. Advances in Neural Information Processing Systems, 34:12362--12373

    2021

  2. [10]

    Ruxiao Chen, Chenguang Wang, Yuran Sun, Xilei Zhao, and Susu Xu. 2025. From perceptions to decisions: Wildfire evacuation decision prediction with behavioral theory-informed llms. arXiv preprint arXiv:2502.17701

    2025 arXiv

  3. [11]

    Xiao Chen, Tao Pei, Ci Song, Hua Shu, Sihui Guo, Xi Wang, Yaxi Liu, and Jie Chen. 2024. Coupling human mobility and social relationships to predict individual socioeconomic status: A graph neural network approach. Transactions in GIS, 28(5):1412--1438

    2024

  4. [12]

    Kun Chu, Xufeng Zhao, Cornelius Weber, Mengdi Li, and Stefan Wermter. 2023. Accelerating reinforcement learning of robotic manipulations via feedback from large language models. arXiv preprint arXiv:2311.02379

    2023 arXiv

  5. [13]

    Faraz Dadgostari, Samarth Swarup, Stephen Adams, Peter Beling, and Henning S Mortveit. 2022. Identifying correlates of emergent behaviors in agent-based simulation models using inverse reinforcement learning. In 2022 Winter Simulation Conference (WSC), pages 322--333. IEEE

    2022

  6. [14]

    Kanishk Gandhi, Jan-Philipp Fr \"a nken, Tobias Gerstenberg, and Noah Goodman. 2023. Understanding social reasoning in language models with language models. Advances in Neural Information Processing Systems, 36:13518--13529

    2023

  7. [15]

    Anindya Ghose, Beibei Li, and Siyuan Liu. 2019. Mobile targeting using customer trajectory patterns. Management Science, 65(11):5027--5049

    2019

  8. [16]

    Yuling Gu, Oyvind Tafjord, Hyunwoo Kim, Jared Moore, Ronan Le Bras, Peter Clark, and Yejin Choi. 2024. Simpletom: Exposing the gap between explicit tom inference and implicit tom application in llms. arXiv preprint arXiv:2410.13648

    2024

  9. [17]

    Masahiko Haraguchi, Akihiko Nishino, Akira Kodaka, Maura Allaire, Upmanu Lall, Liao Kuei-Hsien, Kaya Onda, Kota Tsubouchi, and Naohiko Kohtake. 2022. Human mobility data and analysis for urban resilience: A systematic review. Environment and Planning B: Urban Analytics and Cit...

    2022

  10. [18]

    Roy Hayes, Jonathan Bao, Peter Beling, and Barry Horowitz. 2011. Use of inverse reinforcement learning for identity prediction. In Selected Papers and Presentations Presented at MODSIM World 2010 Conference Expo

    2011

  11. [19]

    Christian Holz-Rau and Joachim Scheiner. 2019. Land-use and transport planning--a field of complex cause-impact relationships. thoughts on transport growth, greenhouse gas emissions and the built environment. Transport Policy, 74:127--137

    2019

  12. [20]

    Jong Wook Kim and Beakcheol Jang. 2024. Privacy-preserving generation and publication of synthetic trajectory microdata: A comprehensive survey. Journal of Network and Computer Applications, page 103951

    2024

  13. [21]

    Thomas Koch and Elenna Dugundji. 2020. A review of methods to model route choice behavior of bicyclists: inverse reinforcement learning in spatial context and recursive logit. In Proceedings of the 3rd ACM SIGSPATIAL International Workshop on GeoSpatial Simulation, pages 30--37

    2020

  14. [22]

    Minae Kwon, Sang Michael Xie, Kalesha Bullard, and Dorsa Sadigh. 2023. Reward design with language models. arXiv preprint arXiv:2303.00001

    2023 arXiv

  15. [23]

    Seungpil Lee, Woochang Sim, Donghyeon Shin, Wongyu Seo, Jiwon Park, Seokki Lee, Sanha Hwang, Sejin Kim, and Sundong Kim. 2024. Reasoning abilities of large language models: In-depth analysis on the abstraction and reasoning corpus. ACM Transactions on Intelligent Systems and T...

    2024

  16. [24]

    Dmitry Leykin, Limor Aharonson-Daniel, and Mooli Lahad. 2016. Leveraging social computing for personalized crisis communication using social media. PLoS currents, 8:ecurrents--dis

    2016

  17. [25]

    Huao Li, Yu Quan Chong, Simon Stepputtis, Joseph Campbell, Dana Hughes, Michael Lewis, and Katia Sycara. 2023. Theory of mind for multi-agent collaboration via large language models. arXiv preprint arXiv:2310.10701

    2023 arXiv

  18. [26]

    Yuebing Liang, Shenhao Wang, Jiangbo Yu, Zhan Zhao, Jinhua Zhao, and Sandy Pentland. 2025. Analyzing sequential activity and travel decisions with interpretable deep inverse reinforcement learning. arXiv preprint arXiv:2503.12761

    2025 arXiv

  19. [27]

    Shan Liu, Ya Zhang, Zhengli Wang, Xiang Liu, and Hai Yang. 2025 a . Personalized origin--destination travel time estimation with active adversarial inverse reinforcement learning and transformer. Transportation Research Part E: Logistics and Transportation Review, 193:103839

    2025

  20. [28]

    Siyuan Liu, Miguel Araujo, Emma Brunskill, Rosaldo Rossetti, Joao Barros, and Ramayya Krishnan. 2013. Understanding sequential decisions via inverse reinforcement learning. In 2013 IEEE 14th International Conference on Mobile Data Management, volume 1, pages 177--186. IEEE

    2013

  21. [29]

    Yepeng Liu, Xuandong Zhao, Dawn Song, and Yuheng Bu. 2025 b . Dataset protection via watermarked canaries in retrieval-augmented llms. arXiv preprint arXiv:2502.10673

    2025 arXiv

  22. [30]

    Yifan Liu, Xishun Liao, Haoxuan Ma, Brian Yueshuai He, Chris Stanford, and Jiaqi Ma. 2024. Human mobility modeling with limited information via large language models. arXiv preprint arXiv:2409.17495

    2024 arXiv

  23. [31]

    Yiru Liu, Yudi Li, Guoyang Qin, Ye Tian, and Jian Sun. 2022. Understanding the behavioral effect of incentives on departure time choice using inverse reinforcement learning. Travel Behaviour and Society, 29:113--124

    2022

  24. [32]

    Massimiliano Luca, Gianni Barlacchi, Bruno Lepri, and Luca Pappalardo. 2021. A survey on deep learning for human mobility. ACM Computing Surveys (CSUR), 55(1):1--44

    2021

  25. [33]

    Yecheng Jason Ma, William Liang, Guanzhi Wang, De-An Huang, Osbert Bastani, Dinesh Jayaraman, Yuke Zhu, Linxi Fan, and Anima Anandkumar. 2023. Eureka: Human-level reward design via coding large language models. arXiv preprint arXiv:2310.12931

    2023 arXiv

  26. [34]

    Alberto Maria Metelli, Filippo Lazzati, and Marcello Restelli. 2023. Towards theoretical understanding of inverse reinforcement learning. In International Conference on Machine Learning, pages 24555--24591. PMLR

    2023

  27. [35]

    Alberto Maria Metelli, Giorgia Ramponi, Alessandro Concetti, and Marcello Restelli. 2021. Provably efficient learning of transferable rewards. In International Conference on Machine Learning, pages 7665--7676. PMLR

    2021

  28. [36]

    National Renewable Energy Laboratory . 2025. Transportation secure data center. https://www.nrel.gov/tsdc. Accessed May 13, 2025

    2025

  29. [37]

    Andrew Y Ng, Stuart Russell, and 1 others. 2000. Algorithms for inverse reinforcement learning. In Icml, volume 1, page 2

    2000

  30. [38]

    Thuy Ngoc Nguyen, Kasturi Jamale, and Cleotilde Gonzalez. 2024. Predicting and understanding human action decisions: Insights from large language models and cognitive instance-based learning. In Proceedings of the AAAI Conference on Human Computation and Crowdsourcing, volume ...

    2024

  31. [39]

    Tomohiro Okubo, Akihiro Kobayashi, Daisuke Kamisaka, and Akinori Morimoto. 2024. Estimation of route-choice behavior along lrt lines using inverse reinforcement learning. Inventions, 9(6):118

    2024

  32. [40]

    Matthias Orlikowski, Jiaxin Pei, Paul R \"o ttger, Philipp Cimiano, David Jurgens, and Dirk Hovy. 2025. Beyond demographics: Fine-tuning large language models to predict individuals' subjective text perceptions. arXiv preprint arXiv:2502.20897

    2025 arXiv

  33. [41]

    Nikolay B Petrov, Gregory Serapio-Garc \' a, and Jason Rentfrow. 2024. Limited ability of llms to simulate human psychological behaviours: a psychometric analysis. arXiv preprint arXiv:2405.07248

    2024 arXiv

  34. [42]

    Shaikh Shakeela, N Sai Shankar, P Mohan Reddy, T Kavya Tulasi, and M Mahesh Koneru. 2021. Optimal ensemble learning based on distinctive feature selection by univariate anova-f statistics for ids. International Journal of Electronics and Telecommunications, pages 267--275

    2021

  35. [43]

    Winnie Street. 2024. Llm theory of mind and alignment: Opportunities and risks. arXiv preprint arXiv:2405.08154

    2024 arXiv

  36. [44]

    Improving llm generation with inverse and forward alignment: Reward modeling, prompting, fine-tuning, and inference-time optimization

    Hao Sun, Thomas Pouplin, Nicol \'a s Astorga, Tennison Liu, and Mihaela van der Schaar. Improving llm generation with inverse and forward alignment: Reward modeling, prompting, fine-tuning, and inference-time optimization. In The First Workshop on System-2 Reasoning at Scale, ...

  37. [45]

    Hao Sun and Mihaela van der Schaar. 2024. Inverse-rlignment: Inverse reinforcement learning from demonstrations for llm alignment. arXiv preprint arXiv:2405.15624

    2024 arXiv

  38. [46]

    Lindia Tjuatja, Valerie Chen, Tongshuang Wu, Ameet Talwalkwar, and Graham Neubig. 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

  39. [47]

    Census Bureau

    U.S. Census Bureau . 2017. https://catalog.data.gov/dataset/tiger-line-shapefile-2017-2010-nation-u-s-2010-census-urban-area-national TIGER/Line Shapefile, 2017, 2010 nation, U.S., 2010 Census Urban Area National . Accessed: 2025-05-13

    2017

  40. [48]

    Census Bureau

    U.S. Census Bureau . 2025. U.s. census bureau. https://www.census.gov/. Accessed: 2025-05-13

    2025

  41. [49]

    Department of Transportation

    Federal Highway Administration U.S. Department of Transportation. 2018. https://nhts.ornl.gov/assets/2017_USTravelProfile.pdf 2017 national household travel survey: U.s. travel profile . Technical report, Federal Highway Administration. Accessed: 2025-05-18

    2018

  42. [50]

    Environmental Protection Agency

    U.S. Environmental Protection Agency . 2025. Smart location mapping. https://www.epa.gov/smartgrowth/smart-location-mapping. Accessed: 2025-05-13

    2025

  43. [51]

    Washington State Geospatial Open Data Portal . 2025. https://geo.wa.gov/ Washington state geospatial open data portal . Accessed: 2025-05-13

    2025

  44. [52]

    Lun Wu, Liu Yang, Zhou Huang, Yaoli Wang, Yanwei Chai, Xia Peng, and Yu Liu. 2019. Inferring demographics from human trajectories and geographical context. Computers, Environment and Urban Systems, 77:101368

    2019

  45. [53]

    Xiefeng Wu. 2024. From reward shaping to q-shaping: Achieving unbiased learning with llm-guided knowledge. arXiv preprint arXiv:2410.01458

    2024 arXiv

  46. [54]

    Takahiro Yabe, Kota Tsubouchi, Akihito Sudo, and Yoshihide Sekimoto. 2016. Estimating evacuation hotspots using gps data: What happened after the large earthquakes in kumamoto, japan. In Proc. of the 5th International Workshop on Urban Computing, volume 81, pages 1--5

    2016

  47. [55]

    Bin Zhang, Soora Rasouli, and Tao Feng. 2024 a . Social demographics imputation based on similarity in multi-dimensional activity-travel pattern: A two-step approach. Travel Behaviour and Society, 37:100843

    2024

  48. [56]

    Yadong Zhang, Shaoguang Mao, Tao Ge, Xun Wang, Adrian de Wynter, Yan Xia, Wenshan Wu, Ting Song, Man Lan, and Furu Wei. 2024 b . Llm as a mastermind: A survey of strategic reasoning with large language models. arXiv preprint arXiv:2404.01230

    2024 arXiv

  49. [57]

    Lei Zhao, Mengdi Wang, and Yu Bai. 2023. Is inverse reinforcement learning harder than standard reinforcement learning? a theoretical perspective. arXiv preprint arXiv:2312.00054

    2023 arXiv

  50. [58]

    Xilei Zhao, Yiming Xu, Ruggiero Lovreglio, Erica Kuligowski, Daniel Nilsson, Thomas J Cova, Alex Wu, and Xiang Yan. 2022. Estimating wildfire evacuation decision and departure timing using large-scale gps data. Transportation research part D: transport and environment, 107:103277

    2022

  51. [59]

    Zhan Zhao and Yuebing Liang. 2023. A deep inverse reinforcement learning approach to route choice modeling with context-dependent rewards. Transportation Research Part C: Emerging Technologies, 149:104079

    2023

  52. [60]

    Yuan Zhong, Nicholas Jing Yuan, Wen Zhong, Fuzheng Zhang, and Xing Xie. 2015. You are where you go: Inferring demographic attributes from location check-ins. In Proceedings of the eighth ACM international conference on web search and data mining, pages 295--304

    2015

  53. [61]

    Pei Zhou, Aman Madaan, Srividya Pranavi Potharaju, Aditya Gupta, Kevin R McKee, Ari Holtzman, Jay Pujara, Xiang Ren, Swaroop Mishra, Aida Nematzadeh, and 1 others. 2023. How far are large language models from agents with theory-of-mind? arXiv preprint arXiv:2310.03051

    2023 arXiv

  54. [62]

    Lei Zhu, Jeffrey Gonder, and Lei Lin. 2017. Prediction of individual social-demographic role based on travel behavior variability using long-term gps data. Journal of Advanced Transportation, 2017(1):7290248

    2017

  55. [63]

    Yuanshao Zhu, Yongchao Ye, Ying Wu, Xiangyu Zhao, and James Yu. 2023. Synmob: Creating high-fidelity synthetic gps trajectory dataset for urban mobility analysis. Advances in Neural Information Processing Systems, 36:22961--22977

    2023

  56. [64]

    Brian D Ziebart, Andrew L Maas, J Andrew Bagnell, Anind K Dey, and 1 others. 2008. Maximum entropy inverse reinforcement learning. In Aaai, volume 8, pages 1433--1438. Chicago, IL, USA

    2008

  57. [65]

    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...

  58. [66]

    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 gl...

Pith tools

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