Runtime Verification of Multiple Natural Language Criteria for Agent Governance
VFM-4B checks 1,000 natural-language rules against a 4,000-token context in under two seconds on a single H100, the authors report.
The verifier evaluates rules in parallel, marking each as satisfied, violated or inapplicable. The finetuned model outperforms GPT-5.4 on a synthetic governance benchmark and DynaGuard-4B on trace-level failure detection in DynaBench. Its individual verdicts identify violated rules and support focused escalation to a larger reasoning judge. Pretraining on CriteriaBank’s 394,000 examples improves transfer across domains, particularly when finetuning data is scarce.
ArXiv · AI/CL/LG's note
The verifier evaluates rules in parallel, marking each as satisfied, violated or inapplicable. The finetuned model outperforms GPT-5.4 on a synthetic governance benchmark and DynaGuard-4B on trace-level failure detection in DynaBench. Its individual verdicts identify violated rules and support focused escalation to a larger reasoning judge. Pretraining on CriteriaBank’s 394,000 examples improves transfer across domains, particularly when finetuning data is scarce.
ArXiv · AI/CL/LG's note
score 5