STARE uses step-wise RL to attack multimodal models, achieving 68% higher attack success rate while revealing that adversarial optimization concentrates conceptual toxicity early and detail toxicity late in the generation trajectory.
emnlp-main.595/
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CR 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
STARE: Step-wise Temporal Alignment and Red-teaming Engine for Multi-modal Toxicity Attack
STARE uses step-wise RL to attack multimodal models, achieving 68% higher attack success rate while revealing that adversarial optimization concentrates conceptual toxicity early and detail toxicity late in the generation trajectory.