DACO curates a 15,000-concept dictionary from 400K image-caption pairs and uses it to initialize an SAE that enables granular, concept-specific steering of MLLM activations, raising safety scores on MM-SafetyBench and JailBreakV while preserving general capabilities.
Rap- guard: Safeguarding multimodal large language models via rationale-aware defensive prompting
2 Pith papers cite this work. Polarity classification is still indexing.
2
Pith papers citing it
citation-role summary
background 1
citation-polarity summary
years
2026 2verdicts
UNVERDICTED 2roles
background 1polarities
background 1representative citing papers
Empirical comparison finds no single inference-time defense dominates for MLLMs, combinations cause 97-100% over-refusal on benign queries, and adaptive selection based on model and attack type is recommended.
citing papers explorer
-
Dictionary-Aligned Concept Control for Safeguarding Multimodal LLMs
DACO curates a 15,000-concept dictionary from 400K image-caption pairs and uses it to initialize an SAE that enables granular, concept-specific steering of MLLM activations, raising safety scores on MM-SafetyBench and JailBreakV while preserving general capabilities.
-
Comparative Analysis of Inference-Time Defense Methods for Multimodal Large Language Models
Empirical comparison finds no single inference-time defense dominates for MLLMs, combinations cause 97-100% over-refusal on benign queries, and adaptive selection based on model and attack type is recommended.