TransitLM introduces a large-scale dataset of over 13 million transit route planning records from four Chinese cities, enabling end-to-end map-free transit route generation.
论文原始摘要
Public transit route planning traditionally depends on structured map infrastructure and complex routing engines, and no existing dataset supports training models to bypass this dependency. We present TransitLM, a large-scale dataset of over 13 million transit route planning records from four Chinese cities covering 120,845 stations and 13,666 lines, released as a continual pre-training corpus and benchmark data for three evaluation tasks with complementary metrics. Experiments show that an LLM trained on TransitLM produces structurally valid routes at high accuracy and implicitly grounds arbitrary GPS coordinates to appropriate stations without any explicit mapping. These results demonstrate that transit route planning can be learned entirely from data, enabling end-to-end, map-free route generation directly from origin-destination information. The dataset and benchmark are available at https://huggingface.co/datasets/GD-ML/TransitLM, with evaluation code at https://github.com/HotTricker/TransitLM.
Paper Collector 中文速览
构建大规模公交数据集，实现无地图端到端路线生成
方法概述
构建覆盖4个城市、12万站点、1.3万线路的1300万条路线数据集，设计三个评估任务和互补指标，用LLM进行持续预训练，验证模型能生成结构有效路线并隐式完成GPS到站点映射
核心贡献
发布1300万条公交路线数据集，首次实现无需结构化地图和路由引擎的端到端公交路线生成，支持隐式GPS到站点映射
原始来源与核验范围
核验仅覆盖论文正文中的主张、证据片段和参考文献关系。NGJOO 未独立运行作者代码、重做实验或验证真实部署效果；页面中的可复现性分数是文档完整度评估，不是复现实验结果。
核心问题
Can large language models perform end-to-end, map-free transit route generation from origin-destination information alone, without relying on external map infrastructure or routing engines?
核心方法
The authors constructed TransitLM from Amap navigation platform logs covering Beijing, Shanghai, Shenzhen, and Chengdu, extracting over 12.9 million planning sessions with 120,845 stations and 13,666 transit lines. They trained Qwen3 models (0.6B-4B) using a two-stage pipeline: continual pre-training on 13.9 million textual route descriptions followed by supervised fine-tuning on three benchmark tasks with 30,000 training examples each.
方法组件
- Classical routing algorithms including Dijkstra, A*, RAPTOR, and Connection Scan Algorithm require structured map infrastructure and real-time schedule data.
- Recent LLM-based approaches like LLM-A*, GridRoute, MapBench, MapTrace, ReasonMap, and TraveLLM still depend on external map data or graph inputs.
- No existing method has achieved end-to-end, map-free transit route generation from origin-destination information.
论点验证
可复现性评估
较低可复现性 (0%)
缺失的复现细节
- No code repository available - implementation details cannot be verified or reused
- No dataset available - the TransitLM dataset is not publicly released
- Hyperparameters only referenced in Appendix E (not accessible) - missing learning rates, batch sizes, CPT epochs, warmup steps, optimizer settings, weight decay
- Random seeds not specified for reproducibility of training runs
- Exact data splits not detailed - only mentions SFT data from 'separate time period' without specifics
- Dataset size and composition not specified - number of samples for CPT and SFT stages unknown
- Data preprocessing steps not documented - how raw transit data was converted to training sequences
- Specific PPU accelerator model and compute requirements not detailed
- Evaluation metrics implementation details not provided
- Geographic coverage and transit system details of the dataset not specified
局限与证据边界
现有分析产物未识别到作者明确写出的局限。这不等于论文没有局限；请结合原文、代码和独立复现实验判断。
本分析由 PDF 阅读助手 自动生成，仅供参考，不构成学术评审意见。验证结论和可复现性评估基于论文文本自动分析，可能存在偏差。原始论文请参阅 arXiv。
分析时间：2026-05-23T01:20:39+00:00 · 数据来源：Paper Collector