Reproduce “TransitLM: A Large-Scale Dataset and Benchmark for Map-Free Transit Route Generation” on OpenResearch

About this paper

Hanyu Guo, Jiedong Yang, Chao Chen, Longfei Xu, Kaikui Liu, Xiangxiang Chu

Researchers from AMAP, Alibaba Group introduced TransitLM, a large-scale dataset comprising 13.9 million public transit planning records, and TransitBench, a three-task benchmark, to enable map-free, end-to-end transit route generation using large language models. Models trained on this data achieved over 93% route connectivity and up to 71% route exact match, demonstrating implicit spatial grounding and effective generalization across varied planning objectives.

Key insights

• Large language models can effectively learn to generate structurally valid and connected public transit routes directly from historical planning data, removing the dependency on explicit map-based routing engines.
• Models implicitly acquire spatial grounding capabilities, enabling them to resolve raw GPS coordinates to appropriate transit stations and understand network topology without explicit geographic databases or textual cues.
• A single, jointly trained model demonstrates strong generalization across diverse route planning objectives (e.g., optimal, preference-aware, multi-route generation) without negative transfer, suggesting the feasibility of unified transit planning solutions.

Summary by alphaXiv.