Refusal in Language Models Is Mediated by a Single Direction

Cite

Arditi, Andy, et al. Refusal in Language Models Is Mediated by a Single Direction. arXiv:2406.11717, arXiv, 30 Oct. 2024. arXiv.org, https://doi.org/10.48550/arXiv.2406.11717.

Metadata

Title: Refusal in Language Models Is Mediated by a Single Direction Authors: Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Panickssery, Wes Gurnee, Neel Nanda Cite key: arditi2024a

Links

• Online Link
• Zotero PDF Link

Abstract

Conversational large language models are fine-tuned for both instruction-following and safety, resulting in models that obey benign requests but refuse harmful ones. While this refusal behavior is widespread across chat models, its underlying mechanisms remain poorly understood. In this work, we show that refusal is mediated by a one-dimensional subspace, across 13 popular open-source chat models up to 72B parameters in size. Specifically, for each model, we find a single direction such that erasing this direction from the model’s residual stream activations prevents it from refusing harmful instructions, while adding this direction elicits refusal on even harmless instructions. Leveraging this insight, we propose a novel white-box jailbreak method that surgically disables refusal with minimal effect on other capabilities. Finally, we mechanistically analyze how adversarial suffixes suppress propagation of the refusal-mediating direction. Our findings underscore the brittleness of current safety fine-tuning methods. More broadly, our work showcases how an understanding of model internals can be leveraged to develop practical methods for controlling model behavior.

Notes

From Obsidian

(As notes and annotations from Zotero are one-way synced, this section include a link to another note within Obsidian to host further notes)

Refusal-in-Language-Models-Is-Mediated-by-a-Single-Direction

From Zotero

(one-way sync from Zotero)

Annotations

Highlighting colour codes

• Note: highlights for quicker reading or comments stemmed from reading the paper but might not be too related to the paper
• External Insight: Insights from other works but was mentioned in the paper
• Question/Critic: questions or comments on the content of paper
• Claim: what the paper claims to have found/achieved
• Finding: new knowledge presented by the paper
• Important: anything interesting enough (findings, insights, ideas, etc.) that’s worth remembering

Link to original

From Zotero

(one-way sync from Zotero)