CauScientist: Teaching LLMs to Respect Data for Causal Discovery

2026-01-20

AI Summary

Existing causal discovery methods are often hindered by statistical indistinguishability, strong modeling assumptions, or unverified priors, limiting their ability to effectively integrate data-driven inference with domain knowledge. This work proposes CauScientist, a novel framework that, for the first time, leverages large language models (e.g., Qwen3-32B) as collaborative hypothesis generators and probabilistic validators. The approach employs hybrid graph initialization, iterative structural refinement, and an error-memory mechanism to guide the search process. CauScientist substantially advances the state of the art in discovering complex causal structures, achieving up to a 53.8% improvement in F1 score on 37-node graphs, increasing recall from 35.0% to 100.0%, and reducing structural Hamming distance by 44.0%, thereby demonstrating both high accuracy and robustness.

Technology Category

• Machine Learning: Causal Learning
• Reasoning under Uncertainty: Causality
• Knowledge Representation and Reasoning: Action, Change, and Causality

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

Causal discovery is fundamental to scientific understanding and reliable decision-making. Existing approaches face critical limitations: purely data-driven methods suffer from statistical indistinguishability and modeling assumptions, while recent LLM-based methods either ignore statistical evidence or incorporate unverified priors that can mislead result. To this end, we propose CauScientist, a collaborative framework that synergizes LLMs as hypothesis-generating"data scientists"with probabilistic statistics as rigorous"verifiers". CauScientist employs hybrid initialization to select superior starting graphs, iteratively refines structures through LLM-proposed modifications validated by statistical criteria, and maintains error memory to guide efficient search space. Experiments demonstrate that CauScientist substantially outperforms purely data-driven baselines, achieving up to 53.8% F1 score improvement and enhancing recall from 35.0% to 100.0%. Notably, while standalone LLM performance degrades with graph complexity, CauScientist reduces structural hamming distance (SHD) by 44.0% compared to Qwen3-32B on 37-node graphs. Our project page is at https://github.com/OpenCausaLab/CauScientist.

Problem

• causal discovery
• large language models
• statistical indistinguishability
• unverified priors
• data-driven methods

Innovation

• causal discovery
• large language models
• probabilistic verification
• hybrid initialization
• iterative refinement