5-2B classifier with direct prediction heads, avoiding auto-regressive rationales that slow down existing guardrail models.
Editable policy tree guardrail classifies agent actions in 58 milliseconds
Jev-Guard combines hierarchical policy routing with direct safety classification, cutting review time in the Hermes agent harness from 4.80s to 1.33s per task.
Top University
Wenbin Hu · Huihao Jing · Haochen Shi · Yuxuan Liu · Changxuan Fan · Sirui Zhang · +6 more
HKUST · ModeIO · CUHK (Shenzhen)
Research Digest··1 min read
Hu et al.
§