Reproduce “Optimizing Against Safety Representations: Activation-Guided Adversarial Suffixes and the Geometry of Refusal” 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
Ege Çakar, Hannah Guan, Kayden Kehe
Researchers at Harvard University introduced Activation-Guided GCG to target internal safety representations in large language models and Soft-GCG for efficient adversarial suffix generation. Their findings reveal that while current safety representations are distributed and vulnerable, larger models exhibit increased resistance to these attacks, and Soft-GCG achieves a 33x speedup over previous methods.
Key insights
- Internal safety representations in LLMs are distributed across the forward pass (layers and token positions) rather than being causally localized to a single site.
- Current RLHF-based alignment techniques result in refusal representations that are geometrically simple and accessible to gradient-based optimization.
- Continuous relaxation with temperature annealing, as implemented in Soft-GCG, can dramatically accelerate adversarial suffix generation while maintaining or improving attack efficacy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.