Research map / Architectures and efficiency / RNNs and normalization

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

NeurIPS 2026 · Architectures and efficiency › RNNs and normalization · approach: Differentiable optimization layers · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces TransitLM, a 13M-record transit route planning dataset from four Chinese cities, plus a three-task benchmark, and shows that Qwen3 models continually pre-trained and SFT-tuned on it generate map-free transit routes from OD GPS coordinates.

Problem

Transit route planning normally requires structured map infrastructure and routing engines, and no existing dataset supports training LLMs to generate complete, station-level transit routes end-to-end from origin-destination information.

Approach

Treat route planning logs from a production navigation platform as a text corpus with station IDs registered as single tokens, then continually pre-train and fine-tune small Qwen3 models so they implicitly ground GPS coordinates to stations and generate connected routes without any map or routing engine.

Key results

• Qwen3-4B reaches 97.0% connectivity and 71.0% Route Exact Match on Optimal Route Generation; the joint 4B-Joint variant reaches 97.9% and 73.7%.
• Best general-purpose LLM (Gemini-3.1-Pro) reaches only 75.5% connectivity and 40.2% Route Exact Match on a more lenient station-only output setting.
• Under GPS-only input, Qwen3-4B keeps 70.4% Route Exact Match (vs 71.0% with text), while general-purpose LLMs drop to under 1% for most models.

Surprising findings

• General-purpose LLMs' connectivity rises under GPS-only input (e.g., DeepSeek-V4 from 64.9% to 80.3%) while station grounding collapses from 72.0% to 16.8%.
• Even 6.25% of the CPT data yields 94.0% connectivity, showing connectivity is learned quickly while exact route matching needs more data.

Limitations

• Covers only four Chinese cities from a single platform (Amap).
• Captures only static route structures without real-time dynamics.
• Only Qwen3 models up to 4B were tested; larger models were not explored.

Experimental setup

Datasets: TransitLM CPT corpus TransitLM SFT data

Benchmarks: TransitBench (Optimal Route Generation, Preference-Aware Planning, Multi-Route Generation)

Metrics: Connectivity Station Grounding Distance Plausibility Line Overlap Station Sequence Overlap Route Exact Match Estimation Accuracy MAPE Preference Compliance Route Diversity

Baselines: GPT-5.4 DeepSeek-V4 Gemini-3.1-Pro Claude-Opus-4.6 Qwen3.6-Plus Doubao-Seed-2.0-Pro Qwen3-0.6B Qwen3-1.7B

Builds on: Domain-adaptive continual pre-training DeepSpeed ZeRO-3

Closest papers by meaning

• SeoulMMOD: A Large-Scale Multimodal Origin-Destination Flow Benchmark (NeurIPS 2026 · similarity 0.83)
• Semantic Routing via Autoregressive Modeling (NeurIPS 2024 · similarity 0.83)
• Satellites Reveal Mobility: A Commuting Origin-destination Flow Generator for Global Cities (NeurIPS 2025 · similarity 0.83)
• GTG: Generalizable Trajectory Generation Model for Urban Mobility (AAAI 2025 · similarity 0.83)
• MobTA: Bus-Conditioned Zero-Shot Trajectory Generation via Task Arithmetic (NeurIPS 2026 · similarity 0.83)
• BO4Mob: Bayesian Optimization Benchmarks for High-Dimensional Urban Mobility Problem (NeurIPS 2025 · similarity 0.82)