TransitLM: A Large-Scale Dataset and Benchmark for Map-Free Transit Route Generation
Hanyu Guo, Jiedong Yang, Chao Chen, Longfei Xu, Kaikui Liu, Xiangxiang Chu
arXiv:2605.22355 · 179▲ · cs.CL, cs.AI, cs.LG
View on arXiv →Premium readers get an interactive explainer for this paper — a figure you can poke at, not just read.
What it is
TransitLM is a dataset of over 13 million transit route planning records (from Amap logs across four Chinese cities, 120,845 stations) plus a three-task benchmark, released as a continual pre-training corpus and supervised fine-tuning data. The authors validate it by continually pre-training then fine-tuning Qwen3 models (0.6B to 4B) that generate complete transit routes end-to-end from origin-destination GPS coordinates, with each station ID registered as a dedicated vocabulary token so routes cannot be hallucinated via character composition.
Why it matters
It shows a language model can internalize transit network topology well enough to produce structurally valid routes directly from coordinates, without a map database or routing engine API at inference time. For practitioners this means route generation with no external tool-calling latency, network dependency, or usage quotas, though it requires retraining whenever the network topology changes and each city expansion grows the vocabulary linearly.
Practical takeaway
If you build routing features, watch for the pattern of encoding discrete entities (stations, IDs) as dedicated tokens plus domain continual pre-training to suppress hallucination. Note the hard scaling wall the authors flag: covering all of China would need roughly 1.8 million station tokens, so this approach does not obviously generalize without vocabulary compression.
Key result
The 4B jointly-trained model reached 71.0% Route Exact Match and 97% connectivity on Optimal Route Generation (10K samples, 4 cities), versus 40.2% Exact Match for the best general-purpose LLM (Gemini-3.1-Pro) even when that baseline was given an easier boarding/alighting-only output format. In GPS-only input (all text removed), the domain model held at 70.4% Exact Match while general LLMs collapsed to under 1%, though the benchmark measures structural overlap against a single production routing engine's outputs, not real-trip validation.
Jargon
Subscribe
Get the next issue.
Free. One email a week. Unsubscribe any time: no account, no dark patterns.