The paper introduces a novel framework for efficient causal graph discovery using large language models (LLMs). Unlike previous LLM-based methods that rely on pairwise queries, which are impractical for larger causal graphs due to their quadratic complexity, the proposed framework employs a breadth-first search (BFS) approach, reducing the query complexity to linear. This method can incorporate observational data to enhance performance and achieves state-of-the-art results on real-world causal graphs of varying sizes. The framework consists of three stages: initialization, expansion, and insertion, where the LLM is used to identify independent variables, find variables caused by current nodes, and add edges while ensuring the resulting graph is a Directed Acyclic Graph (DAG). The effectiveness of the method is demonstrated through experiments on three causal graphs, showing superior performance compared to existing methods, including numerical and pairwise approaches. The paper also discusses limitations and suggests future directions for combining LLMs with observational data and advanced prompting strategies.The paper introduces a novel framework for efficient causal graph discovery using large language models (LLMs). Unlike previous LLM-based methods that rely on pairwise queries, which are impractical for larger causal graphs due to their quadratic complexity, the proposed framework employs a breadth-first search (BFS) approach, reducing the query complexity to linear. This method can incorporate observational data to enhance performance and achieves state-of-the-art results on real-world causal graphs of varying sizes. The framework consists of three stages: initialization, expansion, and insertion, where the LLM is used to identify independent variables, find variables caused by current nodes, and add edges while ensuring the resulting graph is a Directed Acyclic Graph (DAG). The effectiveness of the method is demonstrated through experiments on three causal graphs, showing superior performance compared to existing methods, including numerical and pairwise approaches. The paper also discusses limitations and suggests future directions for combining LLMs with observational data and advanced prompting strategies.