TRAP - Task Accuracy: leaderboard
TRAP (2026) tests agents that must use private document fields to call a tool correctly while refusing to reveal the same fields when asked directly, over text, image and multimodal documents. This board is task accuracy in percent, a correct tool call with the private fields, over all scenarios; higher is better.
Metric: Task accuracy (%): correct tool and arguments using the private fields, sample-weighted over all 500 document-grounded scenarios (100 text, 200 image, 200 multimodal) over ten private document types; each pairs a task query that needs a tool call with private fields and an attack query eliciting the same fields in natural language; rule-based evaluator, greedy decoding. Source: arxiv.org. Saturation forecast: Around May 2027. 22 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 Mini | 75.8 | #348 |
| 2 | GPT-5.4 Mini | 72.6 | #389 |
| 3 | GPT-5 | 72.2 | #167 |
| 4 | Gemini 2.5 Flash | 71 | #453 |
| 5 | GPT-4o Mini | 66.6 | #1623 |
| 6 | Gemini 2.5 Flash Lite | 66 | #958 |
| 7 | Claude Sonnet 4.5 | 62 | #250 |
| 8 | Gemini 2.5 Pro | 61.2 | #266 |
| 9 | Claude Haiku 4.5 | 60.4 | #568 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=trap-task-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-13.