Mitigating Hallucinations in Large Language Models via Causal Reasoning

TL;DR

The paper tackles hallucinations in large language models by introducing explicit causal reasoning through causal-DAG construction. It presents CDCR-SFT, a supervised fine-tuning framework that trains models to build a variable-level DAG and reason over it, backed by the CausalDR dataset containing 25,368 high-quality samples that pair questions with causal graphs, reasoning traces, and answers. Across four LLMs and multiple benchmarks (CLADDER, WIQA, HaluEval), CDCR-SFT yields state-of-the-art causal reasoning (e.g., 95.33% on CLADDER, surpassing human performance) and reduces hallucinations on HaluEval by up to about 11%, demonstrating that explicit causal structure modeling improves output consistency. The work also provides a thorough experimental analysis, DAG-quality assessments, ablation studies, and releases code and data to enable replication and further research in structured causal reasoning for trustworthy AI.

Abstract

Large language models (LLMs) exhibit logically inconsistent hallucinations that appear coherent yet violate reasoning principles, with recent research suggesting an inverse relationship between causal reasoning capabilities and such hallucinations. However, existing reasoning approaches in LLMs, such as Chain-of-Thought (CoT) and its graph-based variants, operate at the linguistic token level rather than modeling the underlying causal relationships between variables, lacking the ability to represent conditional independencies or satisfy causal identification assumptions. To bridge this gap, we introduce causal-DAG construction and reasoning (CDCR-SFT), a supervised fine-tuning framework that trains LLMs to explicitly construct variable-level directed acyclic graph (DAG) and then perform reasoning over it. Moreover, we present a dataset comprising 25,368 samples (CausalDR), where each sample includes an input question, explicit causal DAG, graph-based reasoning trace, and validated answer. Experiments on four LLMs across eight tasks show that CDCR-SFT improves the causal reasoning capability with the state-of-the-art 95.33% accuracy on CLADDER (surpassing human performance of 94.8% for the first time) and reduces the hallucination on HaluEval with 10% improvements. It demonstrates that explicit causal structure modeling in LLMs can effectively mitigate logical inconsistencies in LLM outputs. Code is available at https://github.com/MrLYG/CDCR-SFT.

Figures (6)

• Figure 1: Average overall performance of our CDCR‑SFT applied to four LLMs on the causal reasoning benchmarks (CLADDER and WIQA) and the hallucination benchmark (HaluEval). Orange symbols denote the LLMs enhanced by CDCR‑SFT, demonstrating that CDCR‑SFT significantly improves causal reasoning capabilities and reduces hallucinations.
• Figure 2: Comparison of reasoning approaches: Existing methods (CoT, ToT, GoT) operate at linguistic/semantic levels without explicit causal structure; Our CDCR‑SFT constructs a variable-level causal DAG with directed edges representing causal relationships, enabling principled causal inference through graph-based reasoning.
• Figure 3: Textual representation of the causal DAG in Fig. \ref{['fig:cdcr_review']}.
• Figure 4: Proportional Distribution of 12 Canonical Causal DAG Structures in the CausalDR Dataset.
• Figure 5: Comparison of causal DAG quality scores (Node, Edge, and Structural Scores) generated by pretrained LLMs versus those enhanced with CDCR‑SFT, evaluated on the CLADDER dataset.