2026-09-03

7 arXiv categories· 96 journals· 2656 candidates — 4 worth your time· 116 without an open abstract

Headline

A personalized travel planning benchmark and agent, built from historical behavior trajectories

Today, together

tag shift
no tag ran above its 30-day average
on the same shoulders
5 of the day's papers cite Accessibility evaluation of land-use and transport strategies: review and research directions (Karst Geurs, Bert van Wee 2003); 5 cite Travel demand and the 3Ds: Density, diversity, and design (Robert Cervero, Kara M. Kockelman 1997) counted over the 111 of this day's 120 named papers whose reference lists we have — standing on the same older work is not the same as being about the same thing
first appearance
Fondazione Bruno Kessler none of them on a paper we published in the 83 days before this issue

Two counts, not one counted twice: every row here but “on the same shoulders” counts the papers published above, and that one counts every paper this day named — those and the ones under Also published today, whose abstracts we could not read.

The Behavior2Trip benchmark was established using 11,400 instances from a major Chinese online travel platform to enable personalized travel planning through user behavior trajectories. Another study developed a simulation paradigm that integrates physical mobility flows with social behavioral adaptation to evaluate the impacts of urban traffic regulations. Finally, the TEMPLAR Wales dataset compiled 3,757 georeferenced settlement records across Wales to link a reproducible settlement frame with lexical detections based on Welsh place-name elements.

codedatapreprint

Zihao Cheng, Yingyu Shan, Hongru Wang, et al.
arXiv cs.CL

Data Management and Algorithms · Constraint Satisfaction and Optimization · Spatial Cognition and Navigation · reinforcement learning

Behavior2Trip: Towards Personalized Travel Planning via User Behavior Trajectory

The researchers established a new task, Behavior-Aware Travel Planning, and built the Behavior2Trip benchmark using 11,400 instances from a major Chinese online travel platform, where each instance averages 39.8 past user behaviors across 14 attributes. They developed B2T-Agent, a reinforcement learning-based agent that uses behavior trajectories, external retrieval tools, and an internal memory module. While GPT-4.1 achieved a full-constraint pass rate of only 0.5% on the hardest tasks within the benchmark, the Qwen3-8B model trained with B2T-Agent outperformed all baselines and demonstrated strong generalization on the external TravelPlanner benchmark.

Why it matters — It shifts personalized travel planning away from relying on active, burdensome user inputs or multi-turn clarifications, proving instead that rich, latent preferences can be extracted directly from historical behavior trajectories.

Caveat: The benchmark data is sourced from a single Chinese online travel platform, which may reflect specific regional travel habits and platform-specific user behaviors.

arXiv · pdf · code

published: Transport Policy

Arianna Burzacchi, Marco Pistore
Transport Policy · was arXiv

Fondazione Bruno Kessler

Traffic control and management · Transportation Planning and Optimization · Transportation and Mobility Innovations · Fondazione Bruno Kessler

Evaluating impacts of traffic regulations in complex mobility systems using scenario-based simulations

The researchers developed a simulation paradigm for ex-ante evaluation of urban traffic regulations by integrating a physical layer of mobility flows and emissions with a social layer of behavioral adaptation. The framework was tested using real-world data to compare a baseline scenario against multiple what-if scenarios representing broad traffic restriction schemes.

Why it matters — It provides a structured method to simulate both the direct physical impacts and the indirect behavioral responses of drivers before a policy is implemented, reducing the uncertainty of complex urban traffic interventions.

Caveat: The framework's predictive accuracy depends heavily on the behavioral assumptions encoded as model parameters, which may not fully capture real-world human responses.

Transport Policy · doi

datapreprint

Oktay Karakuş, Can Eyupoglu
arXiv cs.LG

Geographic Information Systems Studies · Historical Studies of British Isles · Landscape and Cultural Studies

TEMPLAR Wales: A georeferenced environmental and toponymic dataset of Welsh settlements

The TEMPLAR Wales dataset compiles 3,757 settlement records across Wales, linking a reproducible settlement frame to 1,350 lexical detections across 1,294 settlements based on 24 Welsh place-name elements. It integrates settlement-level environmental attributes including river and coastal proximity, elevation, terrain context, land cover, and woody cover across multiple spatial scales. The resource is structured as four relational tables accompanied by a field-level data dictionary and source-provenance register.

Why it matters — It provides a standardized, georeferenced foundation for quantitative toponymy and linguistic geography by explicitly separating computational lexical detections from verified etymologies, and contemporary environmental measurements from historical landscape reconstructions.

Caveat: The dataset uses contemporary environmental measurements and computational lexical matches which should not be treated as verified historical landscape reconstructions or definitive etymologies.

arXiv · pdf

preprint

Toshiya Kitahara, Ryu Shirakami, Koh Takeuchi, et al.
arXiv cs.AI

Traffic Prediction and Management Techniques · Human Mobility and Location-Based Analysis · Traffic control and management · Tokyo

CASTANET: Causality-Aware Spatio-Temporal Adversarial Network Using Traffic Incident Effects

The researchers developed CASTANET, a deep learning framework that integrates spatio-temporal graph neural networks with causal treatment effect estimation to predict non-periodic traffic congestion caused by sudden incidents. Evaluated on real-world traffic and accident records from Tokyo, the model reduces root mean square error (RMSE) by 4.0% overall compared to the best baseline, by 10.1% during incident-conditioned scenarios, and by up to 14.55% under severe congestion.

Why it matters — It establishes a method to explicitly incorporate sparse, highly biased incident records into traffic forecasting while mitigating selection bias, enabling accurate predictions of sudden, non-periodic congestion that standard deep learning models fail to anticipate.

Caveat: The model's performance was evaluated using accident records from a single city, Tokyo, which served as the proxy for all traffic incidents.

arXiv · pdf

Still cited

Michael Batty (2008), The Size, Scale, and Shape of Cities 1 of today's items cite it · 18 of 3240 in the archive stand on it

Also published today

These appeared today in journals we track. Their abstracts are not openly available, so we cannot summarise them.