Interventional and Counterfactual Causal Reasoning for LLM-based AI Agents: A Dataset and Evaluation in Portuguese

Resumen

Large Language Models (LLMs) are increasingly central to advancements in generative AI across various domains. While some view these models as a potential step toward artificial general intelligence, their capacity to perform complex causal reasoning remains unverified. Causal reasoning, particularly at Pearl’s interventional and counterfactual levels, is critical for achieving true general intelligence. In this study, we propose a causal reasoning framework that includes a three-axis taxonomy for causality, designed to capture the intent, action requirements, and the three rungs of causality as defined by Pearl: associational, interventional, and counterfactual; and a human-in-the-loop approach to generate golden collections of natural causal questions, annotated according to the proposal taxonomy. We evaluated the seed questions of a golden collection in Portuguese using the LLM GPT-4o and Llama3.1 with two prompt strategies. Our findings reveal that both LLMs face significant challenges in addressing interventional and counterfactual causal queries. These results suggest limitations in the indiscriminate use of these LLMs for extending annotation to additional natural questions or for developing LLM-based causal AI agents.