SIGNPOST-Bench - Conflict Robustness Score: leaderboard
Conflict Robustness Score from SIGNPOST-Bench (arXiv 2608.04244v1, Table 1): weighted combination of the localization accuracy a model retains under unrelated and conflicting scene text, relative to the text-removed image, plus normalized penalties for added error and for following the injected location. Percent; higher is better.
Metric: R (%; retention under random and adversarial text with error and trap penalties). Source: arxiv.org. Saturation forecast: Around February 2027. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 85.93 | #103 |
| 2 | Gemini 2.5 Pro | 85.1 | #159 |
| 3 | Gemini 3.1 Pro (Preview) | 84.49 | #64 |
| 4 | Qwen 3 VL 30B A3B Instruct | 80.05 | #390 |
| 5 | Gemini 2.5 Flash | 77.77 | #255 |
| 6 | Claude Opus 4.6 | 76.84 | #67 |
| 7 | GPT-5 | 75.3 | #105 |
| 8 | Claude Sonnet 4.6 | 74.3 | #94 |
| 9 | GPT-5.4 | 72.27 | #83 |
| 10 | Doubao-Seed-2.0-Pro-260215 | 71.21 | #86 |
| 11 | Kimi K2.5 | 70.57 | #152 |
| 12 | GPT-4o | 69.74 | #357 |
| 13 | GPT-4o Mini | 69.02 | #598 |
| 14 | Qwen 3 VL 235B A22B Instruct | 66.16 | #280 |
| 15 | Grok 4 | 53.82 | #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-conflict-robustness-score · How It Works · Data refreshed daily, snapshot 2026-11-05.