Causal Graph Ensembles of Reasoning Traces to Improve LLM Reasoning and Mitigate Hallucinations

Amruta Parulekar, Jinu Lee, Hari Sundaram

Keywords: Causal Directed Acyclic Graphs, Graph Ensembling, Reasoning Traces, Large Language Models, LLM Reasoning, LLM Hallucinations.

TL;DR: Uncertainty-based ensembling of causal DAGs extracted from reasoning traces for accurate, explainable, and efficient LLM reasoning

Abstract: Large language models (LLMs) hallucinate on legal tasks at least 58% of the time. Although causal inference via LLMs shows promise for improving legal analysis and mitigating hallucinations, current approaches rely on spurious text patterns rather than genuine cause–effect relations, or require computationally expensive supervised fine-tuning. Existing model-routing and step-selection ensembling methods further restrict reasoning diversity. We propose uncertainty-based ensembling of causal Directed Acyclic Graphs (DAGs) extracted from Chain-of-Thought (CoT) reasoning traces, by multiple small, open-source LLMs to reduce hallucinations and improve LLM reasoning. This work delivers auditable, explainable, and compute-efficient LLM reasoning.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 146