SIGNPOST-Bench - Multimodal Conflict Robustness Score: leaderboard
Composite score of SIGNPOST-Bench (arXiv 2608.04244v1, Table 1), which tests image geolocation when scene text is kept, removed, or replaced by compatible, unrelated or conflicting place names. MCRS = 100 x C^0.4 x R^0.6 combines localization capability C with conflict robustness R; 0-100, higher is better.
Metric: MCRS (0-100; 100 x C^0.4 x R^0.6). Source: arxiv.org. Saturation forecast: Around March 2027. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 72.7 | #103 |
| 2 | Gemini 3.1 Pro (Preview) | 72.23 | #64 |
| 3 | Gemini 2.5 Pro | 69.8 | #159 |
| 4 | Gemini 2.5 Flash | 64.31 | #255 |
| 5 | GPT-5 | 64.28 | #105 |
| 6 | Doubao-Seed-2.0-Pro-260215 | 64.23 | #86 |
| 7 | Claude Opus 4.6 | 62.56 | #67 |
| 8 | Claude Sonnet 4.6 | 62.08 | #94 |
| 9 | Kimi K2.5 | 61.74 | #152 |
| 10 | Qwen 3 VL 30B A3B Instruct | 58.5 | #390 |
| 11 | GPT-4o | 57.16 | #357 |
| 12 | GPT-5.4 | 54.77 | #83 |
| 13 | Qwen 3 VL 235B A22B Instruct | 53.36 | #280 |
| 14 | GPT-4o Mini | 50.68 | #598 |
| 15 | Grok 4 | 44.05 | #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-multimodal-conflict-robustness-score · How It Works · Data refreshed daily, snapshot 2026-11-05.