WildRoadBench: leaderboard
We introduce WildRoadBench, a wild aerial road-damage grounding benchmark that couples direct visual grounding by vision-language models with autonomous research-and-engineering by LLM-driven agents on a single professionally annotated UAV corpus. The same image set and the same per-class AP_50 metric are evaluated under two protocols. The VLM Track measures whether a fixed VLM can localise domain
Metric: AP50 (self-reported). Source: benchmarklist.com. Saturation forecast: Around July 2027. 25 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 42.1 | #80 |
| 2 | Qwen 3.6 Plus | 36.8 | #120 |
| 3 | Gemini 3.1 Flash Lite | 25.8 | #191 |
| 4 | Claude Sonnet 4.6 | 25 | #90 |
| 5 | Qwen 2.5 VL 7B Instruct | 23.6 | #654 |
| 6 | Qwen 3 VL 8B Instruct | 21.3 | #420 |
| 7 | GPT-5.4 | 20.9 | #79 |
| 8 | Gemini 2.5 Flash | 16.3 | #245 |
| 9 | Kimi K2.5 | 16.1 | #144 |
| 10 | Qwen 3 VL 8B (Thinking) | 14.8 | |
| 11 | Claude Opus 4.6 | 12.7 | #63 |
| 12 | InternVL3-8B | 10.9 | #620 |
| 13 | GPT-4o | 9.5 | #343 |
| 14 | InternVL3-14B | 7.9 | #505 |
| 15 | InternVL2-8B | 2.2 | #817 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=wildroadbench · How It Works · Data refreshed daily, snapshot 2026-10-20.