Structured Causal CoT Prompting for Enhanced Causal Reasoning in LLMs

Abstract

Large language models (LLMs) have achieved remarkable success on a range of natural language tasks, yet their ability to reason about causal relationships remains limited. We investigate whether a Structured Causal Chain-of-Thought (CoT) prompting approach can improve LLM causal reasoning. In this approach, prompts explicitly guide the model to enumerate causal variables, mechanisms, and inference steps, as opposed to a plain CoT prompt that simply asks for step-by-step thinking. We evaluate four state-of-the-art LLMs on a suite of 12 causal reasoning problems derived from three classic scenarios. Each scenario has four variants to test different causal inference forms. Responses are scored by a qualitative rubric. Our results show that structured causal CoT prompting substantially outperforms plain CoT prompting across all models and scenarios. In particular, GPT-5.1 under structured prompting attains the highest correct rate, while smaller models see dramatic gains under structured prompts, especially on complex intervention questions. We analyze these trends and discuss why explicit structural guidance aids causal inference. Our findings suggest that even without model retraining, thoughtful prompt engineering can significantly enhance LLM reasoning in higher-order tasks. This work provides practical strategies for improving LLM causal reasoning and insights into their current limitations.