SIGNPOST-Bench - Adversarial Text Bias: leaderboard
Text Bias Score from SIGNPOST-Bench (arXiv 2608.04244v1, Table 1): mean increase in geodesic localization error, in kilometres, when scene text is replaced by a conflicting geographic cue, compared with the same image with its text removed; macro-averaged over four datasets. Lower is better.
Metric: Mean added geodesic error from adversarial text relative to the text-removed image (km; lower is better). Source: arxiv.org. Saturation forecast: Estimated already saturated. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 811 | #103 |
| 2 | Gemini 3.1 Pro (Preview) | 879 | #64 |
| 3 | Gemini 2.5 Pro | 888 | #159 |
| 4 | Qwen 3 VL 30B A3B Instruct | 1100 | #390 |
| 5 | GPT-5 | 1317 | #105 |
| 6 | GPT-4o Mini | 1346 | #598 |
| 7 | Gemini 2.5 Flash | 1441 | #255 |
| 8 | GPT-4o | 1461 | #357 |
| 9 | Claude Opus 4.6 | 1482 | #67 |
| 10 | Claude Sonnet 4.6 | 1535 | #94 |
| 11 | GPT-5.4 | 1558 | #83 |
| 12 | Kimi K2.5 | 1577 | #152 |
| 13 | Qwen 3 VL 235B A22B Instruct | 1787 | #280 |
| 14 | Doubao-Seed-2.0-Pro-260215 | 1796 | #86 |
| 15 | Claude Haiku 4.5 | 2287 | #287 |
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-bias · How It Works · Data refreshed daily, snapshot 2026-11-05.