MapTab (Metromap, Map-Only): leaderboard
Metric: Exact match accuracy (%) of the planned route on MapTab's 1,600 Metromap test queries (one per held-out origin-destination pair; city-disjoint test maps), given the map image only (criteria-free shortest-route queries): the station sequence must equal an optimal reference route (spelling similarity above 50%; transfer stations must be marked); malformed outputs count as wrong; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 21 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.5 Flash | 44.19 | #55 |
| 2 | GPT-5.5 Instant | 43.63 | #130 |
| 3 | Gemini 3 Flash (Preview) | 37.06 | #78 |
| 4 | Qwen 3 VL 32B (Thinking) | 13.31 | #287 (Qwen 3 VL 32B) |
| 5 | Doubao-Seed-1.6-251015 (Thinking) | 12.06 | #197 (Doubao-Seed-1.6-251015) |
| 6 | Qwen 3.6 35B A3B (Thinking) | 11.56 | #214 (Qwen 3.6 35B A3B) |
| 7 | GPT-4.1 | 7.94 | #240 |
| 8 | GPT-4o | 6.63 | #333 |
| 9 | Qwen 3 VL 32B Instruct | 6.31 | #276 |
| 10 | Qwen 3.5 9B (Non-reasoning) | 5.69 | #363 (Qwen 3.5 9B) |
| 11 | Qwen 3 VL 8B (Thinking) | 5.12 | |
| 12 | Qwen 3 VL 30B A3B Instruct | 3.31 | #365 |
| 13 | Qwen 3 VL 8B Instruct | 2.75 | #401 |
| 14 | Qwen 2.5 VL 7B Instruct | 0.94 | #643 |
| 15 | Phi-4 Multimodal Instruct | 0 | #896 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=maptab-metromap-map-only · How It Works · Data refreshed daily, snapshot 2026-10-11.