anonsubmission12345/anchorrep-data
Viewer • Updated • 1 • 15
Cross-model jailbreak defense adapters from the AnchorRep paper, NeurIPS 2026
Note Defender: Llama-3-8B. Anchor: Phi-3-medium. Cross-model transfer ASR 1.1% (0% self), BGR 0%.
Note Defender: Mistral-7B-v0.2. Anchor: Qwen1.5-7B. Largest drop: transfer ASR 36% -> 1.0%, BGR 0%.
Note Defender: Vicuna-7B-v1.5. Anchor: Qwen1.5-7B. Cross-model transfer ASR 0.0%, BGR 0%.
Note Defender: Qwen1.5-14B. Anchor: Llama-3-8B. Cross-model transfer ASR 0.4%, BGR 0%.
Note Defender: Phi-3-medium-14B. Anchor: Llama-3-8B. Cross-model transfer ASR 0.0%, BGR 0%.