Improving constraint-based discovery with robust propagation and reliable LLM priors

2025-09-27

AI Summary

Causal structure learning faces two key challenges: constraint-based methods (e.g., PC) rely on idealized conditional independence tests, making them prone to cascading errors due to statistical uncertainty; meanwhile, recent LLM-assisted edge orientation approaches overlook the risk of LLM hallucination. To address these issues, we propose MosaCD—a novel framework that synergistically integrates statistical testing with LLM-derived prior knowledge. First, it employs query shuffling to detect and filter out LLM hallucinations, thereby constructing a high-confidence seed set of oriented edges. Second, it introduces a confidence-decay-aware robust propagation mechanism that prioritizes orientation of the most reliable edges while actively suppressing error accumulation. Evaluated on multiple real-world graph benchmarks, MosaCD consistently outperforms state-of-the-art constraint-based methods, achieving substantial improvements in causal graph reconstruction accuracy. Our results empirically validate the effectiveness of the “high-fidelity seed + robust propagation” paradigm for reliable causal discovery.

Technology Category

• Machine Learning: Causal Learning
• Reasoning under Uncertainty: Causality
• Constraint Satisfaction and Optimization: Constraint Learning and Acquisition

Application Category

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

Abstract

Learning causal structure from observational data is central to scientific modeling and decision-making. Constraint-based methods aim to recover conditional independence (CI) relations in a causal directed acyclic graph (DAG). Classical approaches such as PC and subsequent methods orient v-structures first and then propagate edge directions from these seeds, assuming perfect CI tests and exhaustive search of separating subsets -- assumptions often violated in practice, leading to cascading errors in the final graph. Recent work has explored using large language models (LLMs) as experts, prompting sets of nodes for edge directions, and could augment edge orientation when assumptions are not met. However, such methods implicitly assume perfect experts, which is unrealistic for hallucination-prone LLMs. We propose MosaCD, a causal discovery method that propagates edges from a high-confidence set of seeds derived from both CI tests and LLM annotations. To filter hallucinations, we introduce shuffled queries that exploit LLMs' positional bias, retaining only high-confidence seeds. We then apply a novel confidence-down propagation strategy that orients the most reliable edges first, and can be integrated with any skeleton-based discovery method. Across multiple real-world graphs, MosaCD achieves higher accuracy in final graph construction than existing constraint-based methods, largely due to the improved reliability of initial seeds and robust propagation strategies.

Problem

• Improving causal graph accuracy with robust propagation and LLM priors
• Addressing cascading errors in constraint-based causal discovery methods
• Filtering LLM hallucinations using shuffled queries and confidence propagation

Innovation

• Propagates edges from high-confidence seeds
• Filters hallucinations using shuffled queries
• Uses confidence-down propagation strategy