List your company
ResearchFormerly WizWand

Safety Evaluation on AudioJailBench (AJB) (test)

49.6ASR Accuracy (Audio)Base
4.8816.4928.139.71May 18, 2026
Updated 5mo ago

Evaluation results

MethodLinks
BaseBackbone=Qwen2.5-Omni-3B2026.0549.6
BaseBackbone=MiniCPM-o-4.52026.0545.6
SPA-VL-DPOBackbone=Qwen2.5-Omni-3B2026.0531.8
BaseBackbone=Qwen3-Omni-30B-A3B-Instruct2026.0526.4
ShiftDCBackbone=Qwen2.5-Omni-3B2026.0522.2
VLGuard-SFTBackbone=Qwen2.5-Omni-3B2026.0520.8
ShiftDCBackbone=Qwen3-Omni-30B-A3B-Instruct2026.0518.2
ReGapBackbone=Qwen2.5-Omni-3B2026.0517
ReGapBackbone=Qwen3-Omni-30B-A3B-Instruct2026.0516.9
ReGapBackbone=MiniCPM-o-4.52026.059.2
ShiftDCBackbone=MiniCPM-o-4.52026.056.6