TRAP - Multimodal Task Accuracy: leaderboard

Metric: Task accuracy (%) on the 200 multimodal-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: Around 2029. 22 models tracked.

Top models

#ModelScore
1GPT-5 Mini73
2GPT-572
3Qwen 3 VL 32B71
4Gemini 2.5 Flash69
5GPT-5.4 Mini67.5
6Gemini 2.5 Flash Lite66.5
7InternVL3.5-8B64
8GPT-4o Mini60
9Claude Sonnet 4.555
10Claude Haiku 4.555
11Gemini 2.5 Pro52

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