SIGNPOST-Bench - Adversarial-Text Localization: leaderboard
Geolocation of images whose scene text was replaced by a conflicting geographic cue naming another place (Adversarial variant) in SIGNPOST-Bench (arXiv 2608.04244v1, Table 1). Weighted localization accuracy scores each prediction as exp(-0.005 x geodesic error in km), macro-averaged over IM2GPS3K, YFCC4K, GoogleSV and BaiduSV, in percent; higher is better.
Metric: Weighted localization accuracy (%; exp(-0.005 x geodesic km), four-dataset macro average). Source: arxiv.org. Saturation forecast: Around May 2028. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 44.41 | #64 |
| 2 | Gemini 3 Flash | 43.8 | #103 |
| 3 | Gemini 2.5 Pro | 40.18 | #159 |
| 4 | Doubao-Seed-2.0-Pro-260215 | 38.95 | #86 |
| 5 | GPT-5 | 36.95 | #105 |
| 6 | Gemini 2.5 Flash | 36.9 | #255 |
| 7 | Kimi K2.5 | 35.29 | #152 |
| 8 | Claude Sonnet 4.6 | 34.66 | #94 |
| 9 | Claude Opus 4.6 | 34.47 | #67 |
| 10 | Qwen 3 VL 30B A3B Instruct | 29.03 | #390 |
| 11 | GPT-4o | 28.07 | #357 |
| 12 | GPT-5.4 | 25.9 | #83 |
| 13 | Qwen 3 VL 235B A22B Instruct | 25.49 | #280 |
| 14 | GPT-4o Mini | 20.35 | #598 |
| 15 | Grok 4 | 17.89 | #189 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=signpost-bench-adversarial-text-localization · How It Works · Data refreshed daily, snapshot 2026-11-05.