Large Causal Models for Temporal Causal Discovery
arXiv:2602.18662v1 Announce Type: new Abstract: Causal discovery for both cross-sectional and temporal data has traditionally followed a dataset-specific paradigm, where a new model is fitted for each individual dataset. Such an approach limits the potential of multi-dataset pretraining. The concept of large causal models (LCMs) envisions a class of pre-trained neural architectures specifically designed for temporal causal discovery. Prior approaches are constrained to small variable counts, degrade with larger inputs, and rely heavily on synthetic data, limiting generalization. We propose a principled framework for LCMs, combining diverse synthetic generators with realistic time-series datasets, allowing learning at scale. Extensive experiments on synthetic, semi-synthetic and realistic benchmarks show that LCMs scale effectively to higher variable counts and deeper architectures while maintaining strong performance. Trained models achieve competitive or superior accuracy compared to
arXiv:2602.18662v1 Announce Type: new Abstract: Causal discovery for both cross-sectional and temporal data has traditionally followed a dataset-specific paradigm, where a new model is fitted for each individual dataset. Such an approach limits the potential of multi-dataset pretraining. The concept of large causal models (LCMs) envisions a class of pre-trained neural architectures specifically designed for temporal causal discovery. Prior approaches are constrained to small variable counts, degrade with larger inputs, and rely heavily on synthetic data, limiting generalization. We propose a principled framework for LCMs, combining diverse synthetic generators with realistic time-series datasets, allowing learning at scale. Extensive experiments on synthetic, semi-synthetic and realistic benchmarks show that LCMs scale effectively to higher variable counts and deeper architectures while maintaining strong performance. Trained models achieve competitive or superior accuracy compared to classical and neural baselines, particularly in out-of-distribution settings, while enabling fast, single-pass inference. Results demonstrate LCMs as a promising foundation-model paradigm for temporal causal discovery. Experiments and model weights are available at https://github.com/kougioulis/LCM-paper/.
Executive Summary
This article proposes a novel framework for large causal models (LCMs) that enables temporal causal discovery at scale. The authors combine diverse synthetic generators with realistic time-series datasets to learn complex causal relationships. LCMs can effectively handle higher variable counts and deeper architectures while maintaining strong performance. The results demonstrate LCMs as a promising foundation-model paradigm for temporal causal discovery, particularly in out-of-distribution settings. This approach has the potential to revolutionize the field of causal discovery by enabling fast and accurate inference. The framework's scalability and robustness make it an attractive solution for real-world applications.
Key Points
- ▸ LCMs enable temporal causal discovery at scale
- ▸ Combination of synthetic generators and time-series datasets for learning
- ▸ Effective handling of higher variable counts and deeper architectures
Merits
Strength in Scalability
LCMs can handle large datasets and complex causal relationships, making them a promising solution for real-world applications.
Robustness in Out-of-Distribution Settings
LCMs demonstrate strong performance in out-of-distribution settings, making them a reliable choice for applications where data distributions may vary.
Demerits
Dependency on Synthetic Data
LCMs rely on synthetic data to learn complex causal relationships, which may limit their generalizability to real-world settings.
Computational Requirements
Training LCMs requires significant computational resources, which may be a limitation for some users.
Expert Commentary
The work presented in this article is a significant contribution to the field of causal discovery. The authors' use of synthetic generators and time-series datasets to learn complex causal relationships is a novel approach that shows great promise. However, the reliance on synthetic data and the computational requirements of training LCMs are limitations that need to be addressed. Nevertheless, the results demonstrate the potential of LCMs to handle large datasets and complex causal relationships, making them a promising solution for real-world applications. The scalability and robustness of LCMs make them an attractive choice for applications where data distributions may vary. As the field of causal discovery continues to evolve, it is likely that LCMs will play a significant role in enabling fast and accurate inference in real-world settings.
Recommendations
- ✓ Further research is needed to address the limitations of LCMs, including the reliance on synthetic data and the computational requirements of training.
- ✓ LCMs have the potential to revolutionize the field of causal discovery, and further development and testing are necessary to fully realize their potential.
