I study large language models in the pipelines that produce official travel
statistics: the data behind mode-share figures, activity-based demand models, and infrastructure
investment cases. My research documents a failure mode with real policy stakes. LLMs will "repair" data to
satisfy whatever an automated verifier asserts, even when the verifier itself is wrong, and the
result carries a verified label it did not earn.
The constructive half of the work builds automated checks that flag only what the
data can actually prove, not artifacts of how it was measured, so repair and generation loops are
never asked to fix what the data cannot show is broken. Five connected studies carry this line of work
from measurement to policy use. My dissertation asks what a sound check actually changes as the LLM
moves from repairing observed travel diaries to generating daily activity schedules, the core
representation behind activity-based travel demand models.
5 papersone connected line of work, from measuring the field's trade-offs to policy use
961 → 42studies screened into an evidence-scored map of AI travel models
16,997person-days from the 2022 US National Household Travel Survey in the activity scheduler's input table
When a city travels: activity start times by purpose (Daejeon 2021).Where it travels: Daejeon zones and demand in PTV Visum.
News
Presented Team TUPA's mid-term research plan at the KIRD Global R&D Challenge evaluation in Daejeon, after a second expert mentoring review. The plan asks whether one LLM-based activity-scheduling method can forecast travel demand in both Korea and Australia as autonomous vehicles arrive, including trips that older adults, people with disabilities, and non-drivers now forgo.
The two manuscripts on verification soundness for LLM data repair passed the first phase of double-blind review and advanced to the second.
Preparing a manuscript on language-model completion of household travel diaries for the Transportation Research Part A special issue on generative AI, with Meng Li, Tengfeng Lin, Junyuk Lee, and Inhi Kim.
Team TUPA completed its first expert mentoring review and KIRD's two-day workshop on R&D calls and proposal writing in Daejeon.
Two manuscripts on verification soundness for LLM data repair entered double-blind peer review. Details will appear here after notification.
Team TUPA began the KIRD Global R&D Challenge with a two-day opening training in Daejeon on team roles and research-fund rules.
Revised manuscript of the AI-ABM trilemma review resubmitted to the International Journal of Urban Sciences.
Selected as principal investigator of Team TUPA in the Global R&D Challenge of the Korea Institute of Research and Development (KIRD), one of two lab teams funded at a nationwide competition rate of 11:1, for work on evolving LLM-based agent mobility planners for the autonomous-driving era (lab news).
Outstanding Paper Award at the Korea ITS Society 2026 Spring Conference in Jeju, for neurosymbolic repair of structurally inconsistent household travel surveys (lab news).
Newest first. Scroll for earlier news.
Research
The interpretability, fidelity, and efficiency map: 42 AI-enhanced activity-based models scored on three axes (from the review below). Most cluster toward fidelity.
Foundations
An Evidence-Based Review of AI-Enhanced Activity-Based Models: Trade-offs Between Interpretability, Fidelity, and Efficiency
Kervin Joshua Lucas, Inhi Kim
Int. J. of Urban Sciences · second revision in progress
42 scored studies, one map: AI travel models pile up on predictive
fidelity, while behavioral interpretability and computational efficiency go chronically unmeasured.
Abstract & key numbers
Progress in AI-enhanced activity-based travel models is constrained by an under-articulated
three-way trade-off among behavioral interpretability, predictive fidelity, and computational
efficiency. From a structured review (961 records screened to 42 core studies), every study is
scored on falsifiable indicators and placed on a constant-sum trilemma map (I + F + E = 20), a
visualization device whose headline pattern holds under three alternative normalizations. Quality
is use-case-dependent: progress reads as movement within target regions, not toward a universal
optimum.
961 → 42screening funnel to core scored studies
I + F + E = 20constant-sum trilemma map; pattern robust to three alternative normalizations
Diagnose · Construct
Two manuscripts, details withheld
Titles, venue, and results withheld until the venue reaches a decision
Under double-blind review
The diagnostic and constructive core of the dissertation: one paper
shows that verifier-grounded LLM repair complies with flawed automated checks; the other builds checks
that flag only what the data can prove. Both are listed in full once the venue reaches a decision,
expected late 2026.
Diagnose
Prompt effects, evidence scaffolding and sample distortion in language-model completion of household travel diaries
Transportation Research Part A: Policy and Practice · special issue on generative AI · in preparation
Can a language model fill the gaps in a travel diary? We hide recorded
values in real household travel diaries from Greater Melbourne and Daejeon, such as a trip's destination
area or arrival time, check what the models put back, and ask whether answering only some cases shifts
what the completed survey says about travel.
Policy use
Can LLM Activity Schedulers Be Trusted for Transport Policy Counterfactuals?
Kervin Joshua Lucas, Inhi Kim
Transportation Research Part A: Policy and Practice · special issue on generative AI · planned
Can a generated schedule carry a policy conclusion? Whether LLM
activity schedulers are trustworthy enough to answer counterfactual policy questions, building on
the verifier-soundness line in the work above.
Scope
Extends the verified-LLM thread of the papers above from repairing observed travel records to
generating the schedules a policy question would be asked of. Planned for the special issue on
generative AI in transport policy and applications, and currently at the scoping stage.
Dissertation
Research question (working direction). When an LLM pipeline
that repairs travel data or generates daily activity schedules is gated by an automated check, what
does the check's soundness change? The chapters ask what an unsound check destroys, what a sound one
can certify, how it shifts the balance of validity, fidelity, coverage, and cost in generated
schedules, and where the pipeline breaks when its assumptions change. A bounded or negative answer is
an acceptable result.
Dissertation chapters by stage, focus, and status
Ch.
Stage
Focus
Status
2
Foundations
The interpretability, fidelity, and efficiency trade-off in AI travel models
IJUS, second revision in progress
3
Diagnose
How unsound verifiers corrupt LLM repair; automated editing of survey travel diaries
in review; in preparation
4
Construct
Provably sound verification
manuscript, in review
5
Apply
When a check improves an LLM scheduler of a person's day (NHTS 2022)
in progress, pilot stage
6
Stress-test
Where the pipelines break under changed assumptions, missing inputs, and intervention prompts
planned
Chapter 5 design
An LLM proposes a full-day activity schedule from survey-style persona attributes, and a checker
audits it in a propose, check, and revise loop. The chapter compares a conventional rule checker
with a sound checker built on Chapter 4, and sets both against a conditional deep generative model and
simple baselines trained on the same public diaries. It measures schedule validity, fidelity to the
observed diaries, acceptance coverage, and cost. A certificate states only that the named rules
hold; it makes no claim about behavioral fidelity. Data: the 2022 NextGen National Household Travel
Survey, 16,997 person-days and 31,074 trips from 7,893 households.
Publications
Peer-reviewed journal and conference work. Full list and live citation counts on
Google Scholar.
Strategic transit route recommendation considering multi-trip feature desirability using a logit model with optimal travel-time analysis
M.A. Guillermo, M.C. Rivera, K.J.C. Lucas, R.S. Concepcion II, A.A. Bandala, et al. · Journal of Advanced Computational Intelligence and Intelligent Informatics 26(6) · 2022 · 7 citations · DOI
Korean conferences
Recursive neurosymbolic optimization for the semantic repair of structurally inconsistent household travel surveys
K.J. Lucas, J. Lee, I. Kim · Korea ITS Society 2026 Spring Conference, Jeju (한국 ITS 학회 학술대회) · 2026 · Outstanding Paper Award
Prediction models for travel mode and trip purpose using operator networks to advance activity-based models
K.J. Lucas, I. Kim · Korean Society of Transportation Conference (대한교통학회 학술대회) · 2025
여행 연쇄 예측에서 시공간적 영향 평가 (Evaluating spatiotemporal influence on trip-chain prediction)
K.J. Lucas, J. Lee, S. Paek, I. Kim · Korea ITS Society 2025 Spring Conference (한국 ITS 학회 2025 춘계학술대회) · 2025
Developing a framework for a complete transport forecasting system: a literature review
K.J. Lucas, I. Kim · Korea ITS Society Conference (한국 ITS 학회 학술대회) · 2024
International conferences
Recursive neurosymbolic optimization for the semantic repair of structurally inconsistent household travel surveys using LLM
K.J. Lucas, Y. Liu, T. Lin, I. Kim · 15th Asia-Pacific Conference on Transportation and the Environment (APTE 2026), Jeju · 2026 · talk
Transportation Infrastructure Impacts Calculator (TIIC): an infrastructure assessment tool
A.E.C. Ruiz, N.R. Roxas Jr, K.I.D.Z. Roquel, K.D.S. Yu, A.M. Fillone, K.J.C. Lucas · Transportation Research Procedia 82 · 2025 · DOI
Trip generation in the Philippines: analysis of traditional, geospatial, and AI-based models
K.J. Lucas, A. Fillone, I. Kim · Suwon ITS Asia-Pacific Forum · 2025
Evaluating spatiotemporal influence on trip-chain prediction
K.J. Lucas, J. Lee, S. Paek, I. Kim · Suwon ITS Asia-Pacific Forum · 2025
What if Metro Manila developed a comprehensive rail transit network?
J.R.F. Regidor, A.M. Fillone, D.S. Aloc, K.J.C. Lucas · Proceedings of the Eastern Asia Society for Transportation Studies (EASTS), vol. 11 · 2017 · 9 citations · PDF
Philippine conferences
Metro Manila transportation network: big data analytics and applications
A. Fillone, E. Dadios, N. Roxas, M. Era, K. Lucas, R. Abad, K. Roquel · Sustainable Mobility Research Unit · 2020 · 2 citations
GIS for better public transportation and transit
M.C. Paringit, K.J. Lucas, M. Cutora · De La Salle University Research Congress, Manila · 2019 · 1 citation
Honors & grants
Principal Investigator grant, Global R&D Challenge2026
Korea Institute of Research and Development (KIRD)
Team TUPA, one of two lab teams selected at a nationwide competition rate of 11:1.
Outstanding Paper Award2026
Korea ITS Society 2026 Spring Conference · Jeju
For neurosymbolic repair of structurally inconsistent household travel surveys.
Other work
Earlier applied studies that shaped this direction, including trip generation under
spatial heterogeneity in Aklan, Philippines, comparing classical, geospatial (GWR), and
operator-learning (DeepONet) models.
Education
PhD, MobilitySince 2024
Cho Chun Shik Graduate School of Mobility, KAIST · Daejeon, South Korea
Third-year candidate, expected 2028. Travel demand modeling and trustworthy AI for activity-based models.
MSc, Civil Engineering2017–2020
De La Salle University · Manila, Philippines
Thesis: effects of traffic congestion pricing schemes on travel behavior in Bonifacio Global City, Taguig.
BSc, Civil Engineering2009–2014
De La Salle University · Manila, Philippines
Experience
Principal Investigator, Team TUPA2026
Global R&D Challenge, Korea Institute of Research and Development (KIRD)
Funded independent project “Development of Evolving Agent-based Mobility Planners in Autonomous Driving Era using LLM”, with Tengfeng Lin, Yonghui Liu, and Donghyun Kwon. The program trains teams to plan international joint research through training sessions, expert mentoring, and mid-term and final evaluations; the team's plan covers LLM-based activity scheduling for travel-demand forecasting that transfers between Korea and Australia.
Assistant Professorial Lecturer2021–2023
De La Salle University · Part-time
Taught transportation engineering and transport-modeling software courses.
Science Research Specialist2015–2022
Civil Engineering Department, De La Salle University · Full-time
Transportation-planning data collection and analysis; modeled Metro Manila road and traffic networks for transport studies; supported teaching and departmental research.
