[2603.02235] Talking with Verifiers: Automatic Specification Generation for Neural Network Verification

Computer Science > Machine Learning arXiv:2603.02235 (cs) [Submitted on 13 Feb 2026] Title:Talking with Verifiers: Automatic Specification Generation for Neural Network Verification Authors:Yizhak Y. Elboher, Reuven Peleg, Zhouxing Shi, Guy Katz, Jan Křetínský View a PDF of the paper titled Talking with Verifiers: Automatic Specification Generation for Neural Network Verification, by Yizhak Y. Elboher and 4 other authors View PDF HTML (experimental) Abstract:Neural network verification tools currently support only a narrow class of specifications, typically expressed as low-level constraints over raw inputs and outputs. This limitation significantly hinders their adoption and practical applicability across diverse application domains where correctness requirements are naturally expressed at a higher semantic level. This challenge is rooted in the inherent nature of deep neural networks, which learn internal representations that lack an explicit mapping to human-understandable features. To address this, we bridge this gap by introducing a novel component to the verification pipeline, making existing verification tools applicable to a broader range of domains and specification styles. Our framework enables users to formulate specifications in natural language, which are then automatically analyzed and translated into formal verification queries compatible with state-of-the-art neural network verifiers. We evaluate our approach on both structured and unstructured datasets, de...