AI Paper Atlas

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

PlugGuard: A Streaming Safeguard for Large Models via Latent Dynamics-Guided Risk Detection

Xiaodan Li, Mengjie Wu, Yao Zhu, Yunna Lv, YueFeng Chen, Cen Chen, Jianmei Guo, Hui Xue

ICML 2026 · Large language models › Language, safety and interpretability · approach: Data and hallucination detection

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
replace last-token MLP with SLD head (ablation)Kelp probe · WildGuardresponse-level F1: 0.2410 -> 0.8434Improved
replace last-token MLP with SLD head (ablation)Kelp probe · S-Evalresponse-level F1: 0.7018 -> 0.8971Improved
add ATC loss to MLP probe (ablation)Kelp probe · S-Evalstreaming F1: 0.7841 -> 0.8528Improved
add ATC loss to MLP probe (ablation)Kelp probe · MMSafetyBenchstreaming F1: 0.7097 -> 0.7491Improved
add ATC loss to SLD head (ablation)Kelp · WildGuardstreaming F1: 0.5085 -> 0.8333Improved
add SLD to ATC-trained probe (ablation)Kelp · WildGuardstreaming F1: 0.4122 -> 0.8333Improved
add ATC loss to MLP probe (ablation)Kelp probe · WildGuardstreaming F1: 0.4979 -> 0.4122Hurt
increase ATC anchors N from 1 to 6 (ablation)Kelp · S-EvalF1Improved
increase ATC anchors N too large (ablation)Kelp · S-EvalF1Hurt
Kelp versus SafeDecoding (ablation)Qwen3-8B · WildGuardharmful rate (%): 14.4 -> 3.1Improved
use streaming token-triggered variant (ablation)Kelp · WildGuardharmful rate (%): 3.1 -> 2.0Improved
fine-tune ShieldGemma2 on StreamGuardBenchShieldGemma2 · WildGuardresponse-level F1: 0.4336 -> 0.8466Improved

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