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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| replace last-token MLP with SLD head (ablation) | Kelp probe · WildGuard | response-level F1: 0.2410 -> 0.8434 | Improved |
| replace last-token MLP with SLD head (ablation) | Kelp probe · S-Eval | response-level F1: 0.7018 -> 0.8971 | Improved |
| add ATC loss to MLP probe (ablation) | Kelp probe · S-Eval | streaming F1: 0.7841 -> 0.8528 | Improved |
| add ATC loss to MLP probe (ablation) | Kelp probe · MMSafetyBench | streaming F1: 0.7097 -> 0.7491 | Improved |
| add ATC loss to SLD head (ablation) | Kelp · WildGuard | streaming F1: 0.5085 -> 0.8333 | Improved |
| add SLD to ATC-trained probe (ablation) | Kelp · WildGuard | streaming F1: 0.4122 -> 0.8333 | Improved |
| add ATC loss to MLP probe (ablation) | Kelp probe · WildGuard | streaming F1: 0.4979 -> 0.4122 | Hurt |
| increase ATC anchors N from 1 to 6 (ablation) | Kelp · S-Eval | F1 | Improved |
| increase ATC anchors N too large (ablation) | Kelp · S-Eval | F1 | Hurt |
| Kelp versus SafeDecoding (ablation) | Qwen3-8B · WildGuard | harmful rate (%): 14.4 -> 3.1 | Improved |
| use streaming token-triggered variant (ablation) | Kelp · WildGuard | harmful rate (%): 3.1 -> 2.0 | Improved |
| fine-tune ShieldGemma2 on StreamGuardBench | ShieldGemma2 · WildGuard | response-level F1: 0.4336 -> 0.8466 | Improved |
Closest papers by meaning
- SafeWatch: An Efficient Safety-Policy Following Video Guardrail Model with Transparent Explanations (ICLR 2025 · similarity 0.86)
- Building a Foundational Guardrail for General Agentic Systems via Synthetic Data (ICLR 2026 · similarity 0.85)
- PolyGuard: Massive Multi-Domain Safety Policy-Grounded Guardrail Dataset (NeurIPS 2025 · similarity 0.84)
- What Should a Streaming Video Model Remember? (NeurIPS 2026 · similarity 0.84)
- Towards Policy-Adaptive Image Guardrail: Benchmark and Method (CVPR 2026 · similarity 0.84)
- WildGuard: Open One-stop Moderation Tools for Safety Risks, Jailbreaks, and Refusals of LLMs (NeurIPS 2024 · similarity 0.84)
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.
