TLDR: This research introduces CDCR-SFT, a supervised fine-tuning framework that trains Large Language Models (LLMs) to explicitly construct and reason over causal Directed Acyclic Graphs (DAGs). By teaching LLMs to understand cause-and-effect relationships, the framework significantly improves their causal reasoning capabilities, achieving state-of-the-art accuracy on benchmarks like CLADDER (surpassing human performance). Crucially, this enhanced causal reasoning also leads to a notable reduction in logically inconsistent hallucinations across various NLP tasks, demonstrating a direct link between causal understanding and LLM reliability.
Large Language Models, or LLMs, have shown incredible capabilities in understanding and generating human-like text. However, they often suffer from a significant drawback: hallucinations. These are instances where the LLM generates information that sounds plausible and coherent but is factually incorrect or logically inconsistent. This issue is particularly problematic in complex reasoning tasks, where the models might rely on surface-level patterns rather than true understanding.
Recent research has highlighted a connection between an LLM’s ability to perform causal reasoning and its tendency to hallucinate. Essentially, models with stronger causal reasoning skills tend to produce fewer illogical hallucinations. This observation led researchers to explore whether improving an LLM’s causal reasoning could directly help in reducing these inconsistencies.
Traditional reasoning methods used in LLMs, such as Chain-of-Thought (CoT), Tree-of-Thought (ToT), and Graph-of-Thought (GoT), primarily operate at a linguistic or semantic level. This means they process words and their meanings but don’t explicitly model the underlying cause-and-effect relationships between different variables. They might generate a sequence of steps that looks like reasoning, but they lack the ability to truly understand conditional dependencies or how one event directly influences another. This limitation means they can’t effectively block misleading connections or ensure that their reasoning holds up in hypothetical ‘what-if’ scenarios.
To address this fundamental gap, a new framework called Causal-DAG Construction and Reasoning (CDCR-SFT) has been introduced. This is a supervised fine-tuning approach that trains LLMs to explicitly build a ‘causal Directed Acyclic Graph’ (DAG) at the variable level and then perform reasoning based on this graph. A causal DAG is a powerful tool that visually represents variables and the direct causal links between them, showing how information flows and influences outcomes.
A crucial part of this new framework is the creation of a specialized dataset called CausalDR. This dataset, comprising 25,368 samples, is specifically designed to teach LLMs how to construct these causal DAGs and then reason over them. Each sample in CausalDR includes an input question, a clear causal DAG, a step-by-step reasoning trace that follows the graph’s structure, and the correct answer. This dataset was built using an automated process with a powerful reasoning model, DeepSeek-R1, and further enhanced through a technique called Causal DAG Augmentation to increase its diversity.
During the fine-tuning process, CDCR-SFT guides the LLM to generate this structured causal DAG inference sequence. If the model makes a mistake in constructing the DAG – for example, by reversing a causal link or missing an important variable – immediate feedback (gradient signals) helps correct these errors. This explicit supervision ensures that the model learns the correct causal directions and dependencies, leading to more accurate causal reasoning.
Experiments were conducted on four different LLMs, including Llama-3.1-8B-Instruct and Mistral-7B-Instruct-v0.2, across various benchmarks. On causal reasoning tasks like CLADDER and WIQA, CDCR-SFT consistently showed significant improvements. For instance, with the Llama-3.1-8B-Instruct model, it achieved an impressive 95.33% accuracy on CLADDER, surpassing human performance for the first time. This demonstrates that explicitly modeling causal structures helps LLMs generalize better across different causal reasoning scenarios.
Beyond improving reasoning, CDCR-SFT also proved effective in reducing hallucinations. On the HaluEval benchmark, which assesses hallucinations in dialogue, question answering, and summarization tasks, the CDCR-SFT method consistently outperformed baseline methods. For the Llama-3.1-8B model, it achieved an overall accuracy of 54.93%, showing a clear reduction in logical inconsistencies. These reductions were achieved without specific hallucination-focused training, suggesting that they are a natural outcome of enhanced causal reasoning capabilities.
An ablation study further confirmed that these performance gains are indeed due to the explicit causal DAG construction and reasoning strategy, rather than just exposure to more causal data. The quality of the generated causal DAGs also significantly improved with CDCR-SFT, indicating that the models learned to correctly represent and reason with causal structures.
Also Read:
- Boosting LLM Logical Reasoning with Structured Chain-of-Thought
- Making Knowledge Graph Rules Understandable with AI
In conclusion, the research demonstrates that by equipping LLMs with structured reasoning capabilities that align with the underlying causal nature of real-world problems, we can achieve more trustworthy and reliable models. This approach, detailed in the paper Mitigating Hallucinations in Large Language Models via Causal Reasoning, offers a promising path forward for developing LLMs that are not only coherent but also logically consistent and less prone to generating false information.
