Uni-SafeBench - Dual-Unsafe VQA: leaderboard

Uni-SafeBench (2026) tests the safety of unified multimodal models, vision-language models and LLMs on malicious requests across understanding and generation tasks, with a judge that checks both the context of each response and its intrinsic content. This board is the Both-Safe rate in percent on visual questions where both the image and the text are malicious; higher is better.

Metric: Both-Safe rate (%) on the 455 dual-unsafe visual question answering queries of Uni-SafeBench (image and text both malicious), judged by the Uni-Judger framework (malicious-intent extractor plus safety detector on Qwen3-VL-8B-Instruct), a response counts as safe only when it is safe or a refusal in the context of the extracted intent and its own content is not intrinsically unsafe, mean over repeated judge runs; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 14 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Chameleon (Uni-SafeBench checkpoint unspecified)98.9
2GPT-4o88.8#751
3Gemini 2.5 Flash Image87.9
4Qwen 2.5 VL 7B Instruct84.2#1902
5DeepSeek-VL-7B-Chat80.7#3756
6Tar-1.5B68.1
7DeepSeek-VL-7B-Base68.1
8Harmon-1.5B67.5
9Janus-Pro-7B63.1#6857
10Tar-7B62.4
11BAGEL-7B-MoT59.8#4619
12Show-o2-7B57.1
13Emu353.8
14Janus-flow50.5

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=uni-safebench-dual-unsafe-vqa · How It Works · Data refreshed daily, snapshot 2026-11-21.