This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
This paper addresses the challenge of recovering the structure of causal graph models from observational data, a crucial yet challenging task for causal discovery in scientific scenarios. Existing domain-specific causal discovery relies on expert validation or prior analysis to enhance reliability, but is limited by the lack of expert resources. This paper demonstrates the potential of large-scale language models (LLMs) to serve as autonomous experts guiding data-driven structural learning. However, the inaccuracy of LLM-based inference hinders the integration of LLMs into causal discovery. To address this, we propose a fault-tolerant LLM-based causal discovery framework. The fault-tolerant mechanism is designed with three aspects in mind: During the LLM-based inference process, an accuracy-focused prompting strategy restricts causal analysis to a reliable range. Next, knowledge-to-structure transfer aligns causal statements derived from LLMs with structural causal interactions. During the structure learning process, prior inaccuracy is addressed by balancing the fit to the data with adherence to prior information derived from LLMs. Evaluation on eight real-world causal structures demonstrates that the proposed LLM-based approach improves data-driven causal discovery and is robust to inaccurate LLM-based prior information. The code can be found in https://github.com/tyMadara/LLM-CD .
Takeaways: We propose a fault-tolerant causal discovery framework utilizing LLM to improve the accuracy and efficiency of data-driven causal discovery. We validate the effectiveness and robustness of the proposed method using real-world datasets.
•
Limitations: The performance of the proposed framework depends on the performance of the LLM, and errors may still occur due to inaccurate inferences from the LLM. Further research is needed to determine its generalization performance across diverse domains and complex causal structures. Performance may vary depending on the type and size of the LLM used.
