PGuard researchers report streaming AI safety checks with 0.82-2.26% overhead

Researchers presenting PGuard in an arXiv preprint report token-level safety monitoring with 0.82-2.26% end-to-end overhead on Qwen3.5-122B-A10B served through SGLang. The system reads intermediate representations inside the model without changing its original inference pathway or generation behavior. PGuard selectively combines signals from different model layers during prompt processing and generation. Evaluations cover more than 30 benchmarks; on text queries using Qwen3.5-9B, it achieved an average F1 score of 0.836, approaching dedicated external guardrails. Operators can configure thresholds and enforcement actions for individual safety categories.