AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

ASGuard: Activation-Scaling Guard to Mitigate Targeted Jailbreaking Attack

Yein Park, Jungwoo Park, Jaewoo Kang

ICLR 2026 · Large language models › Language, safety and interpretability · approach: Jailbreak elicitation and risk · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
apply ASGuard (circuit-targeted scaling + preventative fine-tuning)Llama-3.1-8B-Instruct · Past tense jailbreakattack success rate: 42 -> 8Improved
apply ASGuardQwen2.5-7B-Instruct · Past tense jailbreakattack success rate: 51 -> 8Improved
apply ASGuardGemma-2-9B-it · Past tense jailbreakattack success rate: 38 -> 19Improved
add preventative fine-tuning after activation scaling (ablation)ASGuard · Gemma-2-9B-it MMLUaccuracy: 50.3 -> 72.2Improved
add preventative fine-tuning after activation scaling (ablation)ASGuard · Qwen2.5-7B-Instructpast tense ASR: 37 -> 8Improved
add preventative fine-tuning after activation scaling (ablation)ASGuard · Llama-3.1-8B-Instruct MMLUaccuracy: 64.3 -> 68.2Improved
scale tense-vulnerable heads rather than random heads (ablation)Activation scaling · Llama-3.1-8B-Instructpast tense ASR: 25 (random) vs 13 (circuit heads)Improved
ablate (zero out) vulnerable heads instead of scaling (ablation)Head Ablation · Past tense jailbreakattack success rate: 42 -> 29 vs ASGuard 8Hurt
use SFT (30/70) instead of ASGuardSFT · Qwen2.5-7B-InstructOR-Bench-Hard over-refusal: 12.9 -> 98.5Hurt
use SFT (5/95) instead of ASGuardSFT · Gemma-2-9B-itMMLU: 72.2 -> 43.1Hurt
apply ASGuard to other attacksLlama-3.1-8B-Instruct · GCGattack success rate: 15 -> 1Improved
apply ASGuard to other attacksLlama-3.1-8B-Instruct · LogiBreakattack success rate: 30 -> 13Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy