Uni-SafeBench - Composite-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 whose image and text are benign alone but unsafe together; higher is better.

Metric: Both-Safe rate (%) on the 314 composite-unsafe visual question answering queries of Uni-SafeBench (individually benign image and text that are unsafe together), 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)99
2GPT-4o92.7#754
3Qwen 2.5 VL 7B Instruct92.4#1919
4Gemini 2.5 Flash Image90.8
5DeepSeek-VL-7B-Chat90.8#3770
6DeepSeek-VL-7B-Base88.5
7Janus-Pro-7B80.3#6889
8BAGEL-7B-MoT80.3#4668
9Show-o2-7B78.7
10Tar-1.5B77.7
11Harmon-1.5B76.4
12Janus-flow70.4
13Tar-7B68.2
14Emu365

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=uni-safebench-composite-unsafe-vqa · How It Works · Data refreshed daily, snapshot 2026-11-22.