Causal Ensemble Agent: Hierarchical Causal Discovery with LLM-guided Expert Reweighting

Researchers propose Causal Ensemble Agent (CEA), a framework that combines multiple causal discovery algorithms with LLM-guided expert reweighting to improve accuracy in identifying causal relationships from data. The approach addresses limitations of existing methods by dynamically weighting statistical insights and leveraging domain knowledge, demonstrating superior performance across synthetic and real-world datasets.

Causal discovery—the process of inferring cause-and-effect relationships from observational data—remains a fundamental challenge in machine learning with applications spanning finance, healthcare, and policy analysis. The core problem is that different statistical algorithms frequently produce conflicting causal graphs, leaving practitioners uncertain which results to trust. Traditional methods rely heavily on numerical assumptions while ignoring valuable contextual information like feature descriptions that domain experts understand intuitively.

This research introduces a novel ensemble approach that treats the problem as a meta-analysis challenge. Rather than selecting a single best algorithm, CEA aggregates insights across multiple causal discovery methods using linear opinion pooling, then employs an LLM as an intelligent referee to dynamically adjust expert weights when predictions cluster near decision boundaries—precisely where disagreement is highest. This meta-referee role leverages LLMs' ability to incorporate domain knowledge and feature semantics without relying solely on statistical confidence scores.

The framework addresses a critical gap in recent AI research. While some teams have explored using LLMs to directly infer causal relations through prompting, such approaches lack grounding in actual data patterns. CEA instead uses LLMs to synthesize and reweight data-driven discoveries, creating a hybrid methodology that combines statistical rigor with semantic understanding. Experiments demonstrate consistent improvements across diverse datasets.

For AI practitioners and researchers, this work suggests ensemble methods and human-in-the-loop validation remain powerful strategies even as foundation models expand. The approach has implications for any domain requiring causal reasoning—from econometric modeling in fintech to treatment effect estimation in healthcare. Future development could explore how this framework scales to high-dimensional problems and whether similar meta-analysis patterns apply to other discovery tasks.

• →CEA combines multiple causal discovery algorithms with LLM-based expert reweighting to resolve conflicting structural inferences.
• →The framework uses LLMs as meta-referees to dynamically adjust weights near decision boundaries where expert disagreement is highest.
• →Ensemble approach incorporates domain-specific information like feature descriptions alongside statistical signals for more complete causal graphs.
• →Experiments show consistent performance improvements across synthetic and real-world datasets compared to individual methods.
• →Hybrid approach demonstrates that LLMs excel at meta-analysis and synthesis rather than direct causal inference from prompting alone.