Exploring Synergies between Causal Models and LargeLanguage Models for Enhanced Understanding and Inference
Yaru Sun, Ying Yang, Wenhao Fu · 2024
Large language models (LLMs) have sparked a new wave of excitement in the field of artificial intelligence, thanks to their robust generative capabilities. However, they fall short when it comes to comprehending factual knowledge and logical reasoning. In contrast, causal models demonstrate superior interpretability, resilience against disturbances, and decision-support capabilities. By integrating event generation mechanisms and external knowledge, causal models can enhance the reasoning and interpretability of LLMs. Nevertheless, the complex construction and iterative nature of causal models pose challenges that push the boundaries of current frameworks. Thus, leveraging the strengths of both LLMs and causal models can effectively address the limitations of LLMs in logical reasoning, complex inference, and causal deduction, as well as tackle the complexities and difficulties encountered in the establishment and analysis of causal models. These challenges include distinguishing between relevance and causality, handling reverse relationships, and managing interactions. This paper proposes a technical roadmap for a collaborative approach between LLMs and causal models, exploring four different methods of collaboration: causal relationship modeling, causal knowledge injection, causal perception, and causal relationship constraints. We review and summarize existing work while identifying future research directions in harnessing the synergy between LLMs and causal models.