AI Agents Powered by Open-Source Language Models for Causal Understanding in Portuguese

• Uriel Lasheras UNIFOR
• Elioenai Alves UNIFOR
• Caio Ponte UNIFOR
• Carlos Caminha UFC
• Diego Silva UNIFOR
• Vládia Pinheiro UNIFOR

Resumo

Causal reasoning is a key capability for developing more robust, fair, and explainable AI agents. However, the ability of open-source Large Language Models (LLMs) to perform causal reasoning, especially in low-resource languages like Portuguese, remains a significant challenge. This paper introduces two AI agents powered by open-source LLMs: the Causal Classifier Agent and the Causal Judge Agent. For the fine-tuning of the Causal Classifier Agent, we leveraged an expanded version of the Golden Collection with a corpus of 2,500 natural language questions in Portuguese designed to support multilevel causal classification, combined with an enhanced Few-Shot Learning prompting strategy. The results demonstrate that, with targeted training and carefully designed prompts, smaller open-source LLMs can match or even outperform larger models in causal reasoning tasks. In the Causal Judge Agent project, we adopted an enhanced Few-Shot Learning approach and employed an open-source LLM to identify gaps and errors in the fine-tuning process of the first agent. This study highlights the potential of corpus-based fine-tuning and LLM-as-a-judge techniques as cost-effective strategies for improving causal reasoning capabilities in open LLMs.