Throwing Away Maps and Dijkstra: How AMAP's 4B Model Self-Learns Transit Route Planning (TransitLM)

Forum topic · 小凯 · 2026-05-23

Summary

A deep-dive analysis of TransitLM, a research project from AMAP (Alibaba) showing that a 4B-parameter text generation model trained on 13+ million real transit route-planning logs can perform bus and subway route planning without any map data, routing engine (Dijkstra/A*/RAPTOR), or manually maintained topology. The system registers 120,845 station IDs as unique tokens to prevent hallucination, converts planning into autoregressive JSON generation, and was evaluated on three tasks (optimal, preference-aware, and diversified route generation) across Beijing, Shanghai, Shenzhen, and Chengdu. The 4B model achieved 71-73.7% exact match, far exceeding hundred-billion-scale general LLMs like Gemini-3.1-Pro (40.2%), and under GPS-only input—no text hints—it retained 97% connectivity and 98.5% station grounding, showing emergent spatial representation learned purely from text. Joint training across the three tasks showed no negative transfer, and scaling CPT data from 6.25% to 100% steadily improved exact match from 49.9% to 71%. The article argues this demonstrates a paradigm shift: data itself becomes the map.

Key points

Thesis: Traditional transit route planning requires three costly pillars—map data, routing algorithms (Dijkstra/A*/RAPTOR pipelines), and manual configuration. TransitLM (AMAP/Alibaba) asks whether a single text generation model can replace all three by learning from real user travel logs.
Dataset: 12,945,264 route-planning sessions plus 880K station and 148K line descriptions (13.97M+ records), covering Beijing, Shanghai, Shenzhen, and Chengdu; 120,845 unique stations, 13,666 bus/subway lines, 20+ billion tokens, average record length 2,377 Chinese characters. Notably, 30.5% of routes mix modes (taxi/bike first/last mile).
Two-resource design: A CPT corpus (13.9M textualized routes) for domain-adaptive continued pretraining to internalize network topology, and TransitBench, an SFT benchmark with 3 tasks × 30K train / 10K test.
Anti-hallucination design: All 120,845 station IDs are registered as individual tokens, so the model cannot emit nonexistent stations and learns station-level topology directly.
Map-free generation: Input is origin/destination GPS coordinates plus query intent; output is a complete JSON route (line and station sequences with transfer markers, total distance/time/fare, first/last-mile mode and distance). No map, no routing engine, no coordinate-to-station lookup.

Standout result: emergent spatial grounding

Under a GPS-only ablation (no text hints at all):

| Model | Connectivity | Station grounding | Exact match | |---|---|---|---| | Gemini-3.1-Pro (general LLM) | — | 79.9% | 17.7% | | DeepSeek-V4-Pro (general LLM) | — | <<1% | <<1% | | Qwen3-4B (TransitLM) | 97.0% | 98.5% | 71.0% | | 4B-Joint (joint training) | 97.9% | 98.9% | 73.7% |

General LLMs nearly collapse; the 4B specialized model is almost lossless—evidence that spatial topology emerges from textual route descriptions, not memorization of coordinate-station pairs.

Main benchmark results (10,000 samples)

ORG (optimal route generation): 4B-Joint achieves 97.9% connectivity, 73.7% exact match, 1.30% MAPE.
PRG (preference-aware): 95.3% connectivity, 52.6% exact match; four preferences (subway-first, bus-first, fewer transfers, fastest).
DRG (diverse routes, 3 per query): 97.1% connectivity, 67.2% exact match.
General LLMs had an easier input format (predicting only endpoints, not full sequences) yet still lag badly—Gemini-3.1-Pro at 40.2% exact match on ORG.
No negative transfer: joint training across the three tasks matches or beats single-task models everywhere.
Scaling law: exact match rises 49.9% → 71.0% as CPT data goes 6.25% → 100%; connectivity saturates early (94% at 6.25%), showing topology is learned first, fine matching later.

Why it works: from symbolic to neural reasoning

Traditional planning models the world as a graph and searches it; TransitLM models it as a text distribution and samples from it. Frequent co-occurrence patterns ("board at A, ride 3 stops to B, transfer"), shared transfer hubs, and GPS-station co-occurrence let the model encode spatial relationships statistically in its parameters. The infrastructure shifts from map databases to travel logs themselves—launching in a new city means collecting logs, not building maps.

Limitations raised by the analysis

1. Only 4 Chinese cities, single-platform (AMAP) data; cross-city zero-shot transfer untested. 2. Static route structure—no real-time delays, closures, or congestion. 3. Text-only; no visual map input; simplified timetables. 4. Evaluation focuses on structural correctness, not comfort/crowding. 5. Exact match (REM) is judged against *user-chosen* routes, not necessarily optimal ones—71% may be the metric's ceiling rather than the model's.

Bottom line

TransitLM's real contribution is not "using AI for routing" but proving that maps are not required: with 13M+ real travel records, the data itself is the map, and the emergent spatial topology learned from it may be fresher and more behavior-accurate than manually maintained map databases.

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 (Alibaba Group), Beijing. Published 2026-05-21.