Mitigating Hallucinations in Large Language Models via Causal Reasoning

Abstract

Large language models often hallucinate because they lack explicit causal reasoning during generation. This work proposes CDCR-SFT, a supervised fine-tuning framework that trains LLMs to explicitly construct causal DAGs and reason over variable relationships before producing answers. By strengthening the causal reasoning ability of LLMs, CDCR-SFT improves causal reasoning performance and reduces hallucination across standard benchmarks.