BenchCompass - Task-Input Attack (Open Book): leaderboard
BenchCompass Pro: 115 answer-preserving attacked inputs (prior inversion, semantic split and other expert-initialised attack modes) with evidence supplied, measuring brittleness. Higher is better. Source: arxiv.org/abs/2609.18270.
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Around February 2027. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 81.7 | #64 |
| 2 | GPT-5.5 | 80 | #33 |
| 3 | DeepSeek V4 Pro | 80 | #112 |
| 4 | Claude Opus 4.7 | 74.8 | #51 |
| 5 | Gemma 4 31B (IT) | 74.8 | #178 |
| 6 | Qwen 3.5 397B A17B | 72.2 | #154 |
| 7 | Qwen 3.5 122B A10B | 70.4 | #184 |
| 8 | GLM-5.1 | 70.4 | #109 |
| 9 | Gemma 4 26B A4B (IT) | 60.9 | #293 |
| 10 | MiniMax-M2.5 | 60.9 | #316 |
| 11 | Mistral Large 3 | 51.3 | #410 |
| 12 | Llama 4 Maverick | 46.1 | #486 |
| 13 | Qwen 3 32B | 42.6 | #449 |
| 14 | Ministral 3 14B | 37.4 | #664 |
| 15 | Qwen 3 8B | 36.5 | #678 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=benchcompass-task-input-attack-open-book · How It Works · Data refreshed daily, snapshot 2026-11-05.