Harnessing Textual Refusal Directions for Multimodal Safety
Open paper
Benchmark
Model
Rank
Results
multimodal-reasoning-on-mmmu-pro
MARS (Qwen3-VL 8B)
#50
Accuracy: 39.3
vision-language-understanding-on-mmstar
MARS (Molmo2 8B)
#70
Accuracy: 63.6
multimodal-reasoning-on-mmmu
MARS (InternVL 3.5 8B)
#93
Accuracy: 51.4
Rank counts only results with a code link.