VFM verifier scores many natural-language governance criteria in parallel
Researchers introduce VFM, a tree-attention verifier that checks AI agent output against hundreds of written criteria at once, returning a separate satisfied, violated, or inapplicable verdict for each rule.
In plain words
AI agents are supposed to follow many written rules, and someone must check whether each output complies. VFM is a model that reads the context, the output, and many criteria together, then gives a three-way verdict for every rule in one pass. Its authors report it is fast enough to run while an agent is operating.
What was announced
The paper, an arXiv abstract by the VFM authors, describes a verifier that encodes shared context and target text with tree attention and scores N natural language criteria in parallel, returning independent per-criterion ternary verdicts rather than a single pass-fail result. It was pretrained on CriteriaBank, a corpus of 394K (context, target, criterion, verdict) tuples covering more than 321K distinct criterion strings.
According to the abstract, a finetuned VFM-4B outperforms GPT-5.4 on a synthetic governance benchmark and beats DynaGuard-4B on trace-level failure detection in the DynaBench multi-criterion benchmark. Its per-criterion verdicts can localize a violated rule and support escalation to a larger chain-of-thought judge. The authors also report that CriteriaBank pretraining improves cross-domain transfer in low-data settings, and that on a single H100 GPU the model evaluates 1,000 criteria against a 4K-token context in under 2 seconds using under 11 GB of VRAM. These are the authors' own reported results, not independent evaluations.
What it does not show
- The evidence is the paper's abstract only; full methods, datasets, and evaluation details were not available for review.
- All performance comparisons and latency figures come from the authors' own benchmarks and have not been independently confirmed.
For security and IT teams
- Scope
- The reported scope is governance of AI agent outputs against large sets of natural language criteria, with per-criterion verdicts intended for runtime verification and escalation to a larger judge.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.