PyRAT ranked third place at VNN-Comp 2026

We are delighted to announce that PyRAT ranked third at the International Competition of Verification of Neural Network (VNN-COMP).

This year competition provided new challenges that the team had to overcome:

  1. A new format, the VNN-LIB 2.0 vastly increased the range of possible properties, notably including multiple neural network properties. This proved to be a challenge for all verifiers, as this required rework on some core parts of the tools that may induce soundness bugs;
  2. Six new tools with unknown capabilities were competing this year; further increasing the competition stakes.

The PyRAT team was not afraid to sail against those difficult seas. They were able to cover most of the new benchmarks and to stay on the podium. PyRAT scored 2nd on the extended track again, further demonstrating its ability to cover a vast diversity of specifications.

On top of that, the PyRAT team continued its own technical and scientific roadmap. At the end of 2025, they published a tool paper at SAS20251 and participated in numerous scientific gathering (yearly PEPR AI days) and secured industrial contracts to bring formal verification of neural networks to critical systems.

  1. Lemesle, A., Lehmann, J., Gall, T.L., Chihani, Z. (2026). Verifying Neural Networks with PyRAT. In: Oh, H., Sui, Y. (eds) Static Analysis. SAS 2025. Lecture Notes in Computer Science, vol 16100. Springer, Cham. https://doi.org/10.1007/978-3-032-07106-4_2