TransitLM: A Large-Scale Dataset and Benchmark for Map-Free Transit Route Generation

2026-05-22 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

TransitLM is a new, large-scale dataset and benchmark designed for map-free public transit route generation using Large Language Models (LLMs). Released by AMAP, Alibaba Group, it comprises over 13 million transit route planning records from four Chinese cities (Beijing, Shanghai, Shenzhen, Chengdu), covering 120,845 stations and 13,666 lines. This dataset, available as a continual pre-training corpus and benchmark data, enables LLMs to learn transit route planning entirely from data, bypassing traditional map infrastructure and complex routing engines. Experiments with Qwen3 models demonstrate that LLMs trained on TransitLM produce structurally valid routes with high accuracy, implicitly ground arbitrary GPS coordinates to appropriate stations, and generalize across optimal, preference-aware, and multi-route planning objectives. The dataset and evaluation code are publicly available on Hugging Face and GitHub.

Key takeaway

For machine learning engineers developing geospatial AI or transit planning systems, TransitLM offers a validated approach to achieve map-free route generation. You should consider leveraging domain-adaptive continual pre-training on large-scale trajectory data to internalize complex network topology, reducing reliance on external map APIs. This method enables LLMs to implicitly ground GPS coordinates and generalize across diverse planning objectives, offering a robust, self-contained solution for urban mobility applications.

Key insights

End-to-end, map-free transit route generation is feasible by training LLMs on large-scale trajectory data, enabling implicit spatial grounding.

Principles

• Rich trajectory data can replace map-based routing engines.
• Implicit spatial grounding emerges from data alone.
• Domain-specific data is critical, not just model scale.

Method

Continual pre-training on 13.9 million textual route descriptions and static network data, followed by supervised fine-tuning on optimal, preference-aware, and multi-route generation tasks using Qwen3 models.

In practice

• Extend LLM vocabulary with dedicated station ID tokens.
• Use dual-stage CPT and SFT for domain adaptation.
• Jointly train for diverse planning objectives.

Topics

• Public Transit
• Route Planning
• Large Language Models
• Geospatial AI
• Dataset Benchmarking
• Continual Pre-training

Code references

• HotTricker/TransitLM

Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• TransitLM: A Large-Scale Dataset and Benchmark for Map-Free Transit Route Generation — TransitLM enables LLMs to learn map-free transit route generation directly from data, implicitly grounding GPS coordinates to stations.
• TravelEval: A Comprehensive Benchmarking Framework for Evaluating LLM-Powered Travel Planning Agents — TravelEval offers a holistic, simulation-based benchmark revealing LLMs' multi-dimensional travel planning weaknesses.
• Textual Supervision Enhances Geospatial Representations in Vision-Language Models — Textual supervision in VLMs significantly improves implicit geospatial representation learning compared to vision-only models.
• CityTrajBench: A Unified Benchmark for City-Scale Vehicle Trajectory Generation — CityTrajBench unifies urban trajectory generation evaluation, revealing multi-objective quality and no single model dominance.
• FareedKhan-dev / train-llm-from-scratch — Building a Transformer LLM from scratch reveals practical challenges and opportunities for model scaling and domain-specific fine-tuning.
• Guiding LLM-Based Human Mobility Simulation with Mobility Measures from Shared Data — Guiding LLM-based mobility simulations with population-level measures improves collective behavior realism and efficiency.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.