Neurosymbolic Auditing of Natural-Language Software Requirements
A new neurosymbolic pipeline called VERIMED uses large language models combined with an SMT solver to audit natural-language software requirements for ambiguity, inconsistency, vacuousness, and safety violations. The approach translates requirements into formal logic, detects ambiguity through stochastic variation in formalizations, and exposes defects via solver queries. The research focuses on medical-device software requirements and finds that stochastic variation across independent formalizations signals ambiguity, with bidirectional SMT equivalence checking identifying inconsistent interpretations. The paper is available on arXiv under identifier 2605.13817.
Key facts
- VERIMED is a neurosymbolic pipeline for auditing natural-language software requirements.
- It uses large language models and an SMT solver.
- The pipeline translates requirements into formal logic.
- Ambiguity is detected through stochastic variation in formalizations.
- Inconsistency, vacuousness, and safety violations are exposed via solver queries.
- The research focuses on medical-device software requirements.
- Stochastic variation across independent formalizations signals ambiguity.
- Bidirectional SMT equivalence checking identifies inconsistent interpretations.
Entities
Institutions
- arXiv