CDFM: Towards a General-Purpose Causal Discovery Foundation Model

2026-07-14 · Source: stat.ML updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

CDFM, the Causal Discovery Foundation Model, introduces a unified, general-purpose framework for zero-shot structural inference from observational data. It addresses the fragmentation of traditional causal discovery algorithms, which are tailored to specific causal mechanisms. CDFM formulates a principled variational framework, treating unknown causal mechanisms as latent variables and decomposing the marginal likelihood into tractable learning modules. Pretrained on a massive, diverse space of synthetic structural causal models, CDFM internalizes complex statistical asymmetries. Experiments show CDFM consistently outperforms traditional algorithms across 15 mechanism families, varying graph sizes (D=10 to D=100), and sample sizes (N=500 to N=4000) in AUROC and F1 scores, and performs strongly on real-world benchmarks like Causal Chamber and Tübingen cause-effect pairs. It also empirically validates identifiability boundaries.

Key takeaway

For Machine Learning Engineers and Research Scientists working with heterogeneous real-world datasets, CDFM offers a significant shift from manually selecting assumption-specific causal discovery algorithms. You should consider integrating foundation models like CDFM to automatically infer causal structures, reducing the exhaustive loop of statistical tests. This approach promises more robust and scalable causal insights, especially for complex systems with unknown underlying mechanisms.

Key insights

CDFM unifies causal discovery by learning generalizable representations from diverse mechanisms for zero-shot structural inference.

Principles

• Causal prior mechanisms are indispensable for identifiability.
• Variational decomposition guides architecture design for mechanism inference.
• Strict identifiability of mechanism space enables unique graph identification.

Method

CDFM uses a variational framework to treat unknown causal mechanisms as latent variables, decomposing marginal likelihood into mechanism inference, data reconstruction, and graph inference modules.

In practice

• Pretrain on massive, diverse synthetic structural causal models.
• Employ soft causal masking for self-supervised imputation.
• Use graph-adaptive threshold calibrator for binary graph conversion.

Topics

• Causal Discovery
• Foundation Models
• Structural Causal Models
• Variational Inference
• Zero-shot Learning
• Missing Data Imputation

Code references

• DMIRLAB-Group/CDFM

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• CDFM: Towards a General-Purpose Causal Discovery Foundation Model — CDFM offers a unified, zero-shot approach to causal discovery by pretraining on diverse synthetic causal models, outperforming traditional methods.
• Relational Structural Causal Models — Relational Structural Causal Models enable AI to perform causal and combinatorial reasoning in dynamic, multi-object environments.
• Unsupervised Causal Abstractions Discovery — This work proposes an unsupervised method to discover high-level causal abstractions directly from low-level observational data.
• Bridge: Basis-Driven Causal Inference Marries VFMs for Domain Generalization — The "Bridge" framework improves object detection domain generalization by using causal inference to block confounders.
• Identifiability of Causal Graphs under Non-Additive Conditionally Parametric Causal Models — CPCMs offer flexible causal discovery by modeling conditional distributions with diverse parametric families.
• Partial Causal Structure Learning for Valid Selective Conformal Inference under Interventions — Selective conformal prediction under interventions can be robustly maintained by learning partial causal structures.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by stat.ML updates on arXiv.org.