The rapid development of Large Language Models (LLMs) has led to their widespread adoption across various domains, leveraging vast pre-training knowledge and impressive generalization capabilities. However, these models often inherit biased knowledge, resulting in unfair decisions in sensitive applications. It is challenging to remove this biased knowledge without compromising reasoning abilities due to the entangled nature of the learned knowledge within LLMs. To solve this problem, existing approaches have attempted to mitigate the bias using techniques such as fine-tuning with unbiased datasets, model merging, and gradient ascent. While these methods have experimentally proven effective, they can still be sub-optimum in fully disentangling biases from reasoning. To address this gap, we propose Selective Disentanglement Unlearning (SDU), a novel unlearning framework that selectively removes biased knowledge while preserving reasoning capabilities. SDU operates in three stages: identifying biased parameters using a shadow LLM, fine-tuning with unbiased data, and performing selective parameter updates based on weight saliency. Experimental results across multiple LLMs show that SDU improves fairness accuracy by 14.7% and enhances reasoning performance by 62.6% compared to existing baselines.

The ability to extract and verify factual information from free-form text is critical in an era where vast amounts of unstructured data are available, yet unreliable sources abound. This paper focuses on enhancing causal deductive reasoning, a key component of factual verification, through the lens of accident investigation, where determining the probable causes of events is paramount. Deductive reasoning refers to the task of drawing conclusions based on a premise. While some deductive reasoning benchmarks exist, none focus on causal deductive reasoning and are from real-world applications. Recently, large language models (LLMs) used with prompt engineering techniques like retrieval-augmented generation (RAG) have demonstrated remarkable performance across various natural language processing benchmarks. However, adapting these techniques to handle scenarios with no knowledge bases and to different data structures, such as graphs, remains an ongoing challenge. In our study, we introduce a novel framework leveraging LLMs’ decent ability to detect and infer causal relations to construct a causal Knowledge Graph (KG) which represents knowledge that the LLM recognizes. Additionally, we propose a RoBERTa-based Transformer Graph Neural Network (RoTG) specifically designed to select relevant nodes within this KG. Integrating RoTG-retrieved causal chains into prompts effectively enhances LLM performance, demonstrating usefulness of our approach in advancing LLMs’ causal deductive reasoning capabilities.