Reproduce “Dictionary-Aligned Concept Control for Safeguarding Multimodal LLMs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jinqi Luo, Jinyu Yang, Tal Neiman, Lei Fan, Bing Yin, Son Tran, Mubarak Shah, René Vidal
DACO (Dictionary-Aligned Concept Control) provides a framework for safeguarding Multimodal Large Language Models (MLLMs) by enabling granular, semantically grounded control over their internal activations. This approach, leveraging a broad-coverage multimodal concept dictionary to initialize and annotate Sparse Autoencoders, significantly enhances MLLM detoxification performance, reaching a defense success rate of 0.990 on MM-SafetyBench for Qwen2.5-VL-7B while maintaining general utility.
Key insights
- Initializing Sparse Autoencoders with a curated multimodal concept dictionary significantly improves their semantic interpretability and alignment, enabling more effective and targeted control.
- Utilizing multimodal stimuli for concept curation is essential for robust MLLM safety control, as text-only stimuli prove insufficient.
- Granular, inference-time compositional steering based on semantically-grounded SAEs can drastically enhance MLLM safety while carefully balancing the preservation of general model utility.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.