TRAP - Text Task Accuracy: leaderboard

Metric: Task accuracy (%) on the 100 text-document scenarios of 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: Estimated already saturated. 22 models tracked.

Top models

#ModelScore
1GPT-5 Mini90
2GPT-588
3Qwen 3 VL 32B88
4InternVL3.5-8B87
5GPT-5.4 Mini86
6GPT-4o Mini85
7Gemini 2.5 Flash Lite83
8Gemini 2.5 Flash82
9Claude Haiku 4.581
10Claude Sonnet 4.580
11Gemini 2.5 Pro78

Interactive version: theaggregate.ai/benchmark?slug=trap-text-task-accuracy · How It Works · Data refreshed daily, snapshot 2026-09-29.