Expert-guided Causal Structure Learning (CSL) incorporates prior knowledge to improve the accuracy of causal discovery, yet the acquisition of such knowledge is often restricted by the availability of human experts. While Large Language Models (LLMs) provide an alternative source of causal priors, LLM-derived knowledge can be inconsistent with the true causal structure due to hallucinations or contextual misinterpretations. This paper introduces a structural constraint measurement framework, which defines constraint strength and constraint quality to describe reliability and effectiveness, enabling a systematic evaluation of LLM-derived constraints. Using this framework, we evaluate five categories of structural constraints: Edge Existence (EEC), Edge Forbidden (EFC), Path Existence (PEC), Path Forbidden (PFC), and Order Constraints (OC). Our theoretical and empirical analyses demonstrate that while EEC offers high constraint strength, it exhibits low quality when derived from LLMs; conversely, PFC and OC provide a balanced trade-off between search-space pruning and reliability. Building on these insights, we propose a two-level CSL optimization framework that partitions the search space by node order and refines the structure using global path constraints. The results show that this framework provides an effective way to incorporate noisy LLM-derived priors into CSL, particularly in settings where expert knowledge is limited.

Differentiable structure learning of causal directed acyclic graphs (DAGs) is an emerging field in causal discovery, leveraging powerful neural learners. However, the incorporation of ancestral constraints, essential for representing abstract prior causal knowledge, remains an open research challenge. This paper addresses this gap by introducing a generalized framework for integrating ancestral constraints. Specifically, we identify two key issues: the non-equivalence of relaxed characterizations for representing path existence and order violations among paths during optimization. In response, we propose a binary-masked characterization method and an order-guided optimization strategy, tailored to address these challenges. We provide theoretical justification for the correctness of our approach, complemented by experimental evaluations on both synthetic and real-world datasets.

Traditional domain-specific causal discovery relies on expert knowledge to guide the data-based structure learning process, thereby improving the reliability of recovered causality. Recent studies have shown promise in using the Large Language Model (LLM) as causal experts to construct autonomous expert-guided causal discovery systems through causal reasoning between pairwise variables. However, their performance is hampered by inaccuracies in aligning LLM-derived causal knowledge with the actual causal structure. To address this issue, this paper proposes a novel LLM-driven causal discovery framework that limits LLM’s prior within a reliable range. Instead of pairwise causal reasoning that requires both precise and comprehensive output results, the LLM is directed to focus on each single aspect separately. By combining these distinct causal insights, a unified set of structural constraints is created, termed a harmonized prior, which draws on their respective strengths to ensure prior accuracy. On this basis, we introduce plug-and-play integrations of the harmonized prior into mainstream categories of structure learning methods, thereby enhancing their applicability in practical scenarios. Evaluations on real-world data demonstrate the effectiveness of our approach.

Causal discovery faces significant challenges as the number of hypotheses grows exponentially with the number of variables. This complexity becomes particularly daunting when dealing with large sets of variables. We introduce a novel divide-and-conquer method that uniquely handles this challenge. The existing division strategies often rely on conditional independency (CI) tests or data-driven clustering to split variables, which can suffer from the typical data scarcity in large-scale settings, thus leading to inaccurate division results. The proposed method overcomes this by implementing a data-independent division strategy, which constructs a prior structure, informed by potential causal relationships identified using a Large Language Model (LLM), to guide recursively dividing variables into sub-sets. This approach avoids the impact of data insufficiency and is robust against potential incompleteness in the prior structure. In the merging phase, we adopt a score-based refinement strategy to address fake causal links caused by hidden variables in sub-sets, which eliminates edges in the intersected parts of sub-sets to optimize the score of local structures. While maintaining both correctness and completeness under the faithfulness assumption, this novel merging approach demonstrates enhanced performance than the conventional CI-test based merging strategy in practical scenarios. Empirical evaluations on various large-scale datasets demonstrate the proposed approach's superior accuracy and efficiency compared to existing causal discovery methods.

This paper introduces an evaluation method for large language models (LLMs) based on multi-model factual cognition consistency. Traditional evaluation methods, especially in terms of factuality assessments, face challenges in constructing extensive domain-specific question sets and relying on specific model answers. These methods fall short in the face of dynamic and diverse model development. To overcome these limitations, the proposed approach does not depend on a fixed set of standard answers. Instead, it utilizes the responses of multiple models to construct a dynamic, relative evaluation benchmark. We first developed a framework to capture and compare the cognitive consistency of different models when addressing specific questions. Subsequently, a dynamic iterative algorithm was designed to evaluate models based on these sets of answers. Experiments across multiple domains demonstrated the effectiveness of this method. This innovative evaluation strategy not only provides a more comprehensive and flexible approach to understanding and assessing the performance of LLMs in various scenarios but also offers practical guidance for future model development and improvement.

Currently, numerous deep learning (DL) methods have been proposed for the recognition of ground penetrating radar (GPR) B-scan images. Due to the sensitivity of GPR imaging to local underground conditions, DL models trained in other underground environment areas are likely to fail in new areas. Consequently, organizing and labeling new GPR images to train models have become a widely adopted approach in practical applications. However, expert annotation is costly, making it often difficult to collect large-scale datasets with high-quality annotations. Some studies have attempted to improve the quality of image annotations by integrating the efforts of multiple annotators. This process can be constrained by the varying levels of expertise and attention spans of the annotators, which may lead to the presence of errors or contradictions in the provided labels. To address these challenges, this article proposes a method for underground target recognition aimed at multisource annotation tasks. A probability multisource label aggregation (PMLA) module is designed to estimate the reliability of multisource labels, and a label-sensitive regularization (LSR) module is introduced to mitigate the negative impact of potentially erroneous labels on model training. Extensive experiments are conducted on multiple GPR B-scan datasets. The experimental results demonstrate the advantages of the proposed method in handling annotation conflicts and improving the accuracy of underground target recognition.