PACER: New Framework for Scalable Causal Discovery from Interventional Data

Researchers have introduced PACER, which stands for Perturbation-driven Acyclic Causal Edge Recovery. This new framework is designed for causal discovery while ensuring that the results are acyclic. Unlike existing methods that rely on soft acyclicity constraints—which can create optimization problems, lead to numerical issues, and limit scalability—PACER addresses these challenges. It works by modeling a distribution over directed acyclic graphs (DAGs) using a combination of variable permutations and edge probabilities. This allows for straightforward optimization of valid causal structures without the hassle of surrogate penalties. PACER also supports a unified approach for both observational and interventional data and is particularly useful in high-dimensional settings. The findings are published in arXiv:2605.15353.