SIGNPOST-Bench - Trap-Fit Rate: leaderboard
Trap-Fit Rate from SIGNPOST-Bench (arXiv 2608.04244v1, Table 1): share of adversarial-text predictions that land within 50 km of the place named by the injected text, computed on groups whose injected target geocodes (about a third) and macro-averaged over four datasets. Percent; lower is better.
Metric: Adversarial predictions within 50 km of the injected location (%; lower is better). Source: arxiv.org. Saturation forecast: Estimated already saturated. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 6.5 | #159 |
| 2 | Gemini 3 Flash | 6.7 | #103 |
| 3 | Gemini 3.1 Pro (Preview) | 8 | #64 |
| 4 | Qwen 3 VL 30B A3B Instruct | 9.6 | #390 |
| 5 | Gemini 2.5 Flash | 10 | #255 |
| 6 | Claude Opus 4.6 | 11.4 | #67 |
| 7 | GPT-4o | 11.5 | #357 |
| 8 | GPT-5 | 11.6 | #105 |
| 9 | Claude Sonnet 4.6 | 11.8 | #94 |
| 10 | Kimi K2.5 | 13.1 | #152 |
| 11 | GPT-5.4 | 13.8 | #83 |
| 12 | Doubao-Seed-2.0-Pro-260215 | 15.9 | #86 |
| 13 | GPT-4o Mini | 16.7 | #598 |
| 14 | Qwen 3 VL 235B A22B Instruct | 16.9 | #280 |
| 15 | Grok 4 | 19.9 | #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-trap-fit-rate · How It Works · Data refreshed daily, snapshot 2026-11-05.