Mitigating Hallucinations in Large Language Models via Causal Reasoning

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.

These methods generate reasoning structures only at inference time through prompting, without any training signal to correct mis-specified causal relationships. Consequently, when an LLM incorrectly identifies A as causing B (when B actually causes A), or fails to recognize a confounding variable C that influences both, no gradient flows back to fix these fundamental errors (Wang et al. 2023; Yao et al. 2023; Besta et al. 2024). As a result, they cannot block spurious backdoor paths or guarantee counterfactual consistency, leaving LLMs still vulnerable to logically inconsistent hallucinations (Wang et al. 2023; Yao et al. 2023; Besta et al. 2024). The mathematical constraints further compound this problem. Causal relationships inherently form a DAG that encodes multiple interconnected variables with conditional dependencies and multiple pathways of influence. A linear chain or even a tree structure cannot adequately represent scenarios where a variable influences multiple outcomes simultaneously or where effects depend on the interaction of multiple causes, both fundamental characteristics of causal DAG. This structural mismatch means that prompt-only variants such as CoT, ToT, GoT, and DoT cannot, by design, supervise LLMs to learn causal edge semantics, limiting their ability to enforce conditional independencies required for true causal inference.

Each sample in CausalDR comprises (1) an input question, (2) a causal DAG that explicitly describes variables and their relationships, (3) a graph-based reasoning trace that navigates the causal structure, and (4) the final answer.

structured reasoning methods such as Chain-of- Thought (CoT) prompting, which generates intermediate steps alongside final answers (Wei et al. 2022); Self- Consistency (CoT-SC), which samples multiple reasoning chains for robustness; Tree-of-Thoughts (ToT), which branches into alternative solution paths (Yao et al. 2023); and Graph-of-Thoughts (GoT), which links subproblems as nodes in a simple graph (Besta et al. 2024). However, these methods treat inference as linear sequences or trees and cannot represent directed acyclic graph (DAG) needed for causal analysis, where edges denote cause–effect relations and support interventions and counterfactual reasoning.