Researchers from Chulalongkorn University and Ahsanullah University of Science and Technology have developed an integrated computational framework designed to assist resource-constrained transit agencies in evaluating public bus services. The system, named TransitSim-LLM, combines structured data diagnostics from the General Transit Feed Specification (GTFS) with the reasoning capabilities of large language models (LLMs) and the operational simulation of the Simulation of Urban Mobility (SUMO) platform. Published in Frontiers in Future Transportation, the study demonstrates how these distinct analytical components can be linked to identify potential service deficiencies and assess the simulated consequences of proposed interventions.
The primary motivation for the study addresses the challenge faced by transit agencies in translating raw service data into actionable planning decisions. Conventional transit performance evaluation often requires specialized expertise, reliable operational data, and substantial analytical resources, which may be scarce in developing countries. While GTFS provides a standardized representation of transit routes, stops, and schedules, data availability alone does not ensure effective planning. The research team sought to create a workflow that connects schedule-derived diagnostics with intervention generation and performance assessment, thereby reducing the technical barrier to comprehensive service evaluation in Southeast Asian urban contexts.
The methodology involves five analytical phases, beginning with the transformation of GTFS feeds into a structured diagnostic vector. This vector comprises seven dimensions, including service frequency, temporal distribution, spatial coverage, stop spacing, route efficiency, fare structure, and service intensity. These indicators characterize planned service provision rather than observed passenger demand. The LLM then interprets these indicators to identify potential service deficiencies and produces a prioritized issue set based on potential passenger impact, fixability, and data confidence. This prioritization guides the generation of machine-readable intervention scenarios, which are subsequently validated against five technical criteria before entering the simulation stage.
For the case study, the team analyzed a GTFS feed obtained from Malaysia’s national open data portal in January 2025. The feed contained 21 routes, 1,856 scheduled weekday trips, and 271 unique stops. The diagnostic analysis revealed significant variations in service frequency across the network. Route J32 had the shortest mean headway of 6 minutes, while route J200 had the longest at 60 minutes. Under the study-defined thresholds, nine routes were classified as having good service, ten as moderate, and two as poor. The network also exhibited high structural dependence on individual routes, with 76.8% of stops served by only one distinct route, indicating potential vulnerability to disruptions.
The LLM diagnostic module identified four principal planning concerns: low scheduled frequency on selected routes, limited temporal variation, high stop-level dependence on individual routes, and closely spaced consecutive stops. Frequency deficiencies on routes J200 and J42 received the highest priority due to their infrequent departure opportunities. Based on these findings, two intervention scenarios were formulated for simulation. Scenario 1 emphasized frequency enhancement for selected low-frequency services, while Scenario 2 applied a more temporally focused service adjustment. Both scenarios were evaluated against the original GTFS configuration using a common SUMO network constructed from OpenStreetMap data.
The simulation results highlighted the operational trade-offs associated with service enhancements. Under Scenario 1, completed vehicle trips increased by 5.3% to 1,952, and the reported waiting-time indicator decreased by 6.2% to 172.6 seconds. However, these improvements came with a 5.8% increase in total vehicle distance to 19,300 kilometers and a slight reduction in mean speed to 21.8 km/h. Scenario 2 produced smaller changes, with a 1.7% increase in completed trips and a 2.0% increase in vehicle distance. The findings illustrate that greater simulated service output can coincide with additional vehicle activity and changes in operating speed, necessitating multi-indicator assessment rather than reliance on a single performance measure.
To assess the suitability of openly available language models for these tasks, the researchers benchmarked four LLMs: Llama 3.3 70B, Llama 3.1 8B, Llama 4 Scout 17B, and Qwen3 32B. The three Llama models achieved diagnostic quality scores of 85 out of 100 under the study-defined rubric, while Qwen3 32B scored lower. The Llama 4 Scout 17B model demonstrated the lowest response latency at approximately 0.4 seconds. In a repeated-run assessment, the Llama 3.3 70B model consistently identified the principal frequency-related issue in all five runs, although secondary topics varied. The LLM-generated intervention scenarios also achieved higher rubric-based quality scores than rule-based and random alternatives, suggesting that context-aware reasoning contributes to more specific and actionable planning responses.
The study emphasizes that technical scenario validity does not equate to real-world implementation feasibility. The simulation environment was constructed using shortest-path algorithms on an OpenStreetMap-derived network, which may differ from actual bus trajectories. Furthermore, eight of the 271 GTFS stops could not be matched to the simulation network, and the model was not independently calibrated against observed traffic counts or vehicle trajectories. The waiting-time indicator extracted from the simulation is a vehicle-oriented metric and does not represent actual passenger waiting time unless passenger arrivals are explicitly modeled. These limitations underscore the need for observed passenger-demand data and operational feasibility assessments before any real-world implementation of the proposed interventions.
The framework offers a systematic approach for examining alternative transit service configurations and communicating their operational trade-offs. By integrating structured transit data, language-based analytical reasoning, and microsimulation, TransitSim-LLM provides a traceable workflow for evidence-informed planning. The study contributes to the literature by formalizing the process through mathematical indicators and computational algorithms, providing empirical evidence on the diagnostic performance of openly available LLMs, and demonstrating the framework in a Southeast Asian public bus network. The results suggest that such integrated tools can support transit agencies in navigating the complexities of service evaluation and intervention assessment.
Despite its methodological contributions, the study has several limitations that should be considered when interpreting the findings. The framework primarily relies on static GTFS data, which describe scheduled services rather than observed passenger demand or actual operating conditions. Consequently, the diagnostic indicators cannot independently establish experienced service reliability or demand-related deficiencies. The evaluation of LLM performance relied on study-defined rubrics, and the resulting scores do not constitute independent ground-truth validation of diagnostic accuracy. Additionally, the framework was demonstrated using a single public bus network, and its generalizability to other transit systems remains to be established through broader empirical applications.
Future research directions include incorporating observed passenger demand, origin-destination information, and real-time operational data to strengthen the empirical basis of service evaluation. Improving network-matching procedures and incorporating observed traffic and transit operating data for simulation calibration would also enhance the correspondence between simulated and real-world operations. Integrating fleet and resource constraints, economic appraisal, and multi-objective optimization could further examine the balance between service improvements and operational requirements. These advancements would strengthen the empirical validity, operational relevance, and practical applicability of TransitSim-LLM as an evidence-informed transit planning tool.
The development of TransitSim-LLM represents a significant step toward making advanced transit analysis more accessible to resource-constrained agencies. The study highlights the potential of artificial intelligence to support decision-making in public transportation, provided that its outputs are interpreted with an understanding of their limitations and the need for further validation.
Subject of Research: Integrated computational framework for GTFS-driven transit service evaluation using LLMs and SUMO microsimulation
Article Title: TransitSim-LLM: a closed-loop framework for GTFS-driven transit service evaluation using open-source LLMs and SUMO microsimulation
Article References: Hassan, M., & Turjo, T. D. (2026). TransitSim-LLM: a closed-loop framework for GTFS-driven transit service evaluation using open-source LLMs and SUMO microsimulation. Frontiers in Future Transportation, 7, Article 1961606. https://doi.org/10.3389/ffutr.2026.1961606
Image Credits: AI Generated
DOI: 10.3389/ffutr.2026.1961606
Keywords: transit planning, large language models, GTFS, SUMO microsimulation, public transportation, Johor Bahru, service evaluation, computational framework, TransitSim-LLM, closed-loop, framework, GTFS-driven
News Source: Denise Maddox. (November 10, 2026). AI Framework Assesses Transit Service Changes in Johor Bahru. Scienmag.
