AI Paper Atlas

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

Furina: Fragmented Uncertainty-Driven Refusal Instability Attack

Tongxi Wu, Jian Zhang, Yang Gao

ICML 2026 · Large language models › Language, safety and interpretability · approach: Safety alignment of LLMs · 0 citations (OpenAlex)

Read the paper

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
add semantic probes (ablation)Furina · MM-SafetyBench Hate Speech (GPT-4o-mini)ASR (%): 0.00 -> 91.41Improved
add auxiliary synthesis of probe answers (ablation)Furina · MM-SafetyBench Hate Speech (GPT-4o-mini)ASR (%): 7.98 -> 91.41Improved
add scene interpretation (ablation)Furina · MM-SafetyBench Hate Speech (GPT-4o-mini)ASR (%): 88.34 -> 91.41Improved
use dispersed probes instead of progressive ladder reformulation (ablation)Furina · MM-SafetyBench Hate Speech (GPT-4o-mini)ASR (%): 80.98 -> 91.41Improved
render scene as typography instead of diffusion image (ablation)Furina · MM-SafetyBench on Claude-Haiku-4.5ASR (%): 75.48 -> 77.20Improved
add LlamaGuard-3-8B input filtering of probes as defenseFurina · HarmBench (GPT-4o-mini)ASR (%): 94.0 -> 93.5No clear effect
add aggressive p95 perplexity filtering as defenseFurina · HarmBench (GPT-4o-mini)ASR (%): 94.0 -> 86.5Hurt
add p99 perplexity filtering as defenseFurina · HarmBench (GPT-4o-mini)ASR (%): 94.0 -> 93.0No clear effect
increase rewrite strength from original to semantic rewriteLLaMA-2-7B (AdvBench) · AdvBench rewrite ladderASR: 0.01 -> 0.42Improved
increase rewrite strength from original to semantic rewriteLLaMA-3.2-3B (AdvBench) · AdvBench rewrite ladderASR: 0.02 -> 0.90Improved
suppress refusal-direction alignment in last 4 layers (activation patching) (ablation)LLaMA-2-7B-Chat · Original rewriteASR: 0.01 -> 0.04Improved
suppress refusal-direction alignment in last 4 layers (activation patching) (ablation)LLaMA-2-7B-Chat · Semantic rewriteASR: 0.42 -> 0.42No clear effect

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