Safety failures in multimodal LLMs happen because unsafe image-plus-text inputs shift internal representations past a still-working refusal boundary, and a representation-calibration fine-tune restores refusal with under 2% utility loss.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CR 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
MMAligner: Safeguarding Multimodal Large Language Models through Representation Calibration
Safety failures in multimodal LLMs happen because unsafe image-plus-text inputs shift internal representations past a still-working refusal boundary, and a representation-calibration fine-tune restores refusal with under 2% utility loss.