VNN-LIB 2.0 in Practice: Benchmarks for More Complex Neural Network Verification Tasks (Competition Contribution)

25. Jul 2026·
Lukas Rohwer
Samuel Teuber
Samuel Teuber
,
Matthew L. Daggitt
· 0 min read
Abstract
Released in December 2025, VNN-LIB 2.0 [7] is the latest version of the standardised language for specifying neural network verification queries. While VNN-LIB 2.0 introduces many new features, including a type system and semantics, one of the most significant changes is the increased expressivity of its syntax compared to that of VNN-LIB 1.0 [2]. For example, queries can now involve networks with multiple inputs and outputs, multiple interacting networks, non-linear constraints and constraints over hidden nodes. It is hoped this new expressivity will help drive the development of solvers that are capable of handling more sophisticated verification problems.
Type
Publication
Third International Symposium on AI Verification
Samuel Teuber
Authors
PhD Candidate
Researching formal verification of AI-based heterogeneous systems, with applications in neural network control and algorithmic fairness.