Back to AI Research

AI Research

Causal Discovery in the Era of Agents | AI Research

Key Takeaways

  • Causal Discovery in the Era of Agents This paper addresses the growing trend of using Large Language Models (LLMs) to assist in causal discovery—the process...
  • Recent attempts to combine large language models (LLMs) with causal discovery ask models to infer pairwise directions, propose graph structures, or inject language-model outputs as priors and constraints.
  • These approaches promise faster analysis, but they also obscure whether a causal evidence is supported by data and assumptions or by textual associations, prompt artifacts and hallucinated mechanisms.
  • We argue for a different role for agents in causal discovery.
  • Agents should inspect data, retrieve context, explain method assumptions and clarify graph outputs, but they should not supply edges, orientations, priors, constraints or causal conclusions.
Paper AbstractExpand

Recent attempts to combine large language models (LLMs) with causal discovery ask models to infer pairwise directions, propose graph structures, or inject language-model outputs as priors and constraints. These approaches promise faster analysis, but they also obscure whether a causal evidence is supported by data and assumptions or by textual associations, prompt artifacts and hallucinated mechanisms. We argue for a different role for agents in causal discovery. Agents should inspect data, retrieve context, explain method assumptions and clarify graph outputs, but they should not supply edges, orientations, priors, constraints or causal conclusions. We propose the principle that agents assist the workflow, while causal claims remain grounded in data, explicit assumptions, formal algorithms, diagnostics and user or domain-expert decisions. We instantiate this principle in causal-learn+, an online platform that coordinates data analysis, preprocessing, method recommendation, expert-knowledge incorporation, formal discovery and interpretation around the algorithmic ecosystem of causal-learn. A case study on Big Five personality data illustrates agent-assisted pipeline of causal discovery without turning language-model unreliability into causal evidence. The platform is available at this http URL .

Causal Discovery in the Era of Agents
This paper addresses the growing trend of using Large Language Models (LLMs) to assist in causal discovery—the process of identifying cause-and-effect relationships from observational data. While LLMs can make data analysis faster, the authors argue that current methods often blur the line between data-driven evidence and AI-generated text. The paper proposes a new framework, causal-learn+, which ensures that agents act only as assistants to the workflow, while causal conclusions remain strictly grounded in formal algorithms, explicit scientific assumptions, and human expertise.

The Problem with AI-Driven Causal Claims

When LLMs are allowed to suggest causal directions, define graph structures, or set constraints, the resulting causal models become unreliable. Because LLMs are trained to predict patterns in text, their outputs may reflect common beliefs, prompt phrasing, or even hallucinations rather than actual statistical evidence. The authors warn that if these AI outputs are mixed into the discovery process, it becomes impossible to tell if a causal link is supported by data or simply by the model’s linguistic associations. This is particularly dangerous in causal discovery, where the validity of a graph depends entirely on the specific assumptions and mathematical methods used to create it.

A New Role for Agents

To maintain scientific rigor, the authors propose a clear division of labor: agents should assist, but they cannot provide evidence. In this model, agents are used to improve the user experience by summarizing data, explaining complex methodological assumptions, recommending appropriate algorithms, and helping interpret the final results. However, the "inferential core"—the actual calculation of causal graphs, conditional independence tests, and orientation rules—must remain the exclusive domain of formal, transparent algorithms. By keeping agents outside of this core, the system ensures that every causal claim can be traced back to a specific dataset and a well-defined set of assumptions.

Implementing the Workflow

The causal-learn+ platform serves as a practical implementation of this principle. It organizes the causal discovery process into a series of traceable steps:

  • Data Analysis: Agents help users inspect data for missing values, distributions, and potential issues.

  • Preprocessing: Agents suggest transformations like scaling or discretization, but these remain visible, user-approved decisions rather than hidden AI actions.

  • Method Recommendation: Agents guide users toward the right algorithm (such as constraint-based or latent-variable methods) based on the specific scientific question.

  • Interpretation: After the algorithm generates a graph, agents help explain the results, ensuring that users understand the limitations of the output without over-interpreting it.

Case Study: Personality Research

The authors demonstrate this approach using a dataset from the Big Five personality questionnaire. Analyzing 50 survey questions to identify latent personality traits is complex and prone to misinterpretation. By using causal-learn+, researchers were able to use latent-variable methods to identify meaningful structures in the data. The agents provided context on psychological measurement and helped interpret the resulting graphs, but the actual discovery of the causal links was performed by formal algorithms. This case study illustrates that agents can successfully lower the barrier to entry for complex causal analysis without compromising the integrity of the scientific conclusions.

Comments (0)

No comments yet

Be the first to share your thoughts!