The paper reports that REVA uses a contextual probe for function-level vulnerability screening in about one second.
REVA screens functions for vulnerabilities in one model pass
2026-10-27Frontier watchWATCH
REVA combines a module threat model, static-analysis findings and call-graph context, then reads an LLM’s internal representation with a trained probe. On PrimeVul’s held-out test set, the authors report 60.0% accuracy in about one second per function on GLM-5.3-Flash, similar to their VulTriage comparison while over 250 times faster. No public code or independent validation was identified.
What this means for you
For teams reviewing agent-generated code, REVA suggests a cheap screening layer before slower agentic investigation. The absolute accuracy remains too low for autonomous blocking, and the unpublished implementation prevents operational validation.
FRONTIER WATCH
Public artifacts and validated results are different.
The recorded fact is the public work and its artifacts. Author-reported outcomes do not establish independent reproduction. Next checks include usable code, task conditions and third-party validation.
Evidence & related material
Technical details & scores
| Method | Threat model + static analysis + call graph as context; trained probe reads LLM hidden representations |
|---|---|
| Evaluation | PrimeVul held-out test set across three open LLMs |
| Reported Result | 60.0% accuracy on GLM-5.3-Flash; about one second per function; >250x faster than the compared agentic detector |
| Evidence Limit | Author-reported preprint; no code or independent validation found |
| Time Basis | arXiv v1 submitted 2026-10-26T17:50:09Z |
Technical impact 3
Engineering value 4
Adoption signal 1
Maturity 1
Verification cost 4
Risk 4