Method Detail: VLSBench

Back to Leaderboard
Benchmark: VLSOT
Short name: VLSBench
Long name: Visual Leakage in Multimodal Safety Benchmark
Description: VLSBench is a multimodal safety benchmark designed to evaluate whether Vision-Language Models (VLMs/MLLMs) genuinely use visual information when responding to safety-sensitive image-text inputs. It addresses Visual Safety Information Leakage (VSIL), where the textual prompt unintentionally reveals the harmful content in the image, allowing a model to make a safety decision without actually understanding the image. The benchmark is designed to provide a more reliable evaluation of multimodal safety and cross-modal reasoning.
Reference: Hu, Xuhao, Liu, Dongrui, Li, Hao, Huang, Xuanjing, Shao, Jing, VLSBench: Unveiling Visual Leakage in Multimodal Safety. In arXiv preprint arXiv:2411.19939, 2024.
Last submitted: September 15, 2026
Published: September 15, 2026 at 07:17:37
Submissions: 2
Project page / code: https://github.com/AI45Lab/VLSBench.git
Open source: Yes

Benchmark performance

Submission Date SR@0.5 (↑) SR@0.7 (↑) AOR (↑) PR@1.0 (↑) ACE (↓) PR@0.5 (↑)
2026-09-15 07:29 - - - - - -
2026-09-15 07:25 - - - - - -