New ADTGuard Safeguard Enhances LLM Streaming Inference Safety

Researchers have developed ADTGuard, a novel safeguard designed for streaming inference in large language models (LLMs). This system addresses limitations in current guardrails by checking each token before release and progressively refining training targets. ADTGuard utilizes Depth-Time Safety Dynamics (DTSD) to integrate safety evidence across layers and time, and a Causal Depth Router (CDR) to adapt layer selection based on the current text prefix. Experiments show ADTGuard achieves strong streaming detection performance with a small parameter count and competitive latency. AI

IMPACT Improves safety mechanisms for LLMs during streaming inference, potentially reducing exposure to harmful content.

RANK_REASON The item is an academic paper detailing a new method for LLM safety. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• ADTGuard
• alphaXiv
• arXiv
• Causal Depth Router
• Connected Papers
• CORE Recommender
• Depth-Time Safety Dynamics
• Hugging Face
• large-language models
• Litmaps
• Risk-Onset Target Refinement
• scite Smart Citations

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →