A testing and verification framework for AI models based on neural networks, built on a three-tier architecture (front-end, back-end, and pipeline), with native PyTorch support and an ACT intermediate representation (IR) that enables refinement-based precision and supports diverse model architectures, input formats, and specification types.
Install Miniconda and create a running environment.
conda env create -f environment.yml # Install required lib packages to run ACT
conda activate act-py312 # Activate an environment (python-3.12) # Activate the environment
git clone --recursive https://github.com/SVF-tools/ACT.git
cd ACT
2. Apply and download the Gurobi license (Optional for MILP optimization)
cp /path/to/your/gurobi.lic ./modules/gurobi/gurobi.lic # put gurobi.lic file in ./modules/gurobi/ directory
python -m act.pipeline --help
- Kaijie Liu and Yulei Sui. Detecting Unsoundness in Neural Network Verifiers via Concrete–Abstract Consistency. ACM/IEEE International Conference on AI-Powered Software Engineering (AIware 2026)
ACT is licensed under the GNU Affero General Public License v3.0 (AGPL-3.0).
This project was developed with the assistance of GitHub Copilot to enhance code readability and efficiency. AI-generated suggestions were reviewed and tested by the contributors before inclusion.