Beyond Dijkstra: How Amap's 4B-Parameter Model Learns Transit Route Planning Without Maps (TransitLM)

Forum topic · 小凯 · 2026-05-23

Summary

A deep-dive forum post analyzes TransitLM, a research effort from AMAP (Amap) and Alibaba showing that a 4B-parameter language model can learn public transit route planning entirely from text-serialized travel logs, replacing map databases, Dijkstra/A*/RAPTOR routing engines, and manually maintained configurations. The team built a corpus of roughly 13.97 million route-planning records (over 20 billion tokens) covering Beijing, Shanghai, Shenzhen, and Chengdu, with 120,845 stations and 13,666 bus and metro lines. After continual pretraining and supervised fine-tuning on three tasks (optimal route generation, preference-aware planning, diverse multi-route generation), the 4B model achieved 71-73.7% exact-match accuracy and 97%+ route connectivity, outperforming much larger general-purpose LLMs such as Gemini-3.1-Pro (40.2%). A striking GPS-only ablation shows the model spontaneously internalized implicit spatial grounding, mapping arbitrary coordinates to stations without any geographic database. Registering station IDs as individual vocabulary tokens is credited with suppressing hallucinated stations. Joint training across tasks showed no negative transfer, and scaling curves reveal topology is learned quickly while precise matching requires more data. The post also discusses limitations: four cities only, static schedules, no real-time traffic, and questions about cross-city generalization.

This post from zhichai.net analyzes TransitLM (arXiv:2605.22355), a dataset and benchmark from AMAP (Amap) and Alibaba demonstrating that public transit route planning can be learned end-to-end by a small language model, without map data or classical routing algorithms.

Key points

Paradigm shift: TransitLM replaces the traditional pipeline (map database → Dijkstra/A*/RAPTOR → ranking/rules) with a single autoregressive text-generation model trained on ~13 million real route-planning sessions.
Dataset: 12,945,264 route sessions + 880K station descriptions + 148K line descriptions (13.97M+ records, 20B+ tokens), covering Beijing, Shanghai, Shenzhen, and Chengdu: 120,845 unique stations, 13,666 bus/metro lines. 30.5% of routes involve mixed modes (taxi/bike first/last mile).
Two resources: a continual-pretraining (CPT) corpus of textual route descriptions to internalize network topology, and TransitBench, an SFT benchmark with 3 tasks × 30K train / 10K test samples.

Method

The model takes origin/destination GPS coordinates plus query intent and generates a full JSON route (line IDs, station IDs with [Transfer] markers, distance/time/cost, access legs). No map, routing engine, or coordinate-to-station lookup is needed.

A key anti-hallucination design: all 120,845 station IDs are registered as individual vocabulary tokens, so the model cannot emit non-existent stations and directly learns station-level topology.

Headline results

| Model | Connectivity | Station grounding | Exact match (ORG) | |---|---|---|---| | Gemini-3.1-Pro (general LLM) | 75.5% | 93.9% | 40.2% | | DeepSeek-V4-Pro | 64.9% | 72.0% | 23.7% | | GPT-5.4-pro | 60.5% | 60.5% | 18.4% | | Qwen3-4B (TransitLM) | 97.0% | 98.5% | 71.0% | | 4B-Joint | 97.9% | 98.9% | 73.7% |

Under a harsh GPS-only ablation (no textual hints), general LLMs collapse (grounding <<1% for DeepSeek) while TransitLM stays nearly lossless — evidence of emergent implicit spatial representation learned purely from text route logs, not memorization.
Three tasks: Optimal Route Generation (ORG), Preference-aware Route Generation (PRG: metro-first, bus-first, fewer transfers, shortest time), and Diverse Route Generation (DRG). Joint training of all three shows no negative transfer — one model serves all query types.
Numeric metrics (distance/time/cost) are only scored on exact-match samples, forcing structural correctness first; achieved MAPE as low as 1.30%.
Scaling: connectivity is learned quickly (94% at 6.25% of CPT data), while exact match climbs steadily with data (49.9% → 71.0%).

Evaluation framework

Ten metrics across five dimensions: connectivity (Conn), reachability feasibility (station grounding, distance plausibility), route overlap (line IoU, station-sequence IoU, exact match REM), numeric fields (estimate accuracy, MAPE), and task-specific metrics (preference compliance, route diversity).

Limitations and open questions

1. Data covers only four Chinese cities from a single platform; static schedules, no real-time traffic or delays; text-only, no visual map input. 2. Cross-city generalization to an unseen city is untested, though GPS-only results hint at some zero-shot capability. 3. Real-time dynamics could potentially be injected as textual constraints in the prompt. 4. The 71% exact match may reflect a ceiling of the REM metric itself — users' actual choices are not always "optimal," and many valid routes exist per OD pair. The paper does not assess whether generated routes exceed label quality.

Takeaway

The significance is not "AI for routing" but the demonstration that the map itself is optional: with enough real travel logs, the data becomes the map, and the model's implicit spatial topology may be fresher and more behavior-faithful than a manually maintained geographic database.

---

Reference: TransitLM: A Large-Scale Dataset and Benchmark for Map-Free Transit Route Generation (arXiv:2605.22355). Authors: Hanyu Guo*, Jiedong Yang*, Chao Chen, Longfei Xu, Kaikui Liu, Xiangxiang Chu (*equal contribution). AMAP (Amap), Alibaba Group, Beijing.