New AttentionSwitch Method Targets Jailbreak Defense in Large Language Models
A new arXiv paper introduces AttentionSwitch, a jailbreak defense for LLMs that jointly learns sparse attention-head selection and scaling to counter attacks that circumvent safety measures.
Read the full story at arXiv cs.AI ↗
Timeline · 1 report
- 2026-11-09 05:00 · arXiv cs.AI
AttentionSwitch: Jailbreak Defense for LLMs via Joint Learning of Sparse Attention-Head Selection and Scaling