A Novel Method of Multi-hop Reasoning in Large Language Models by Pruning Toxic Model Knowledge

Rapid advancements in artificial intelligence have led to the widespread deployment of language models across various domains, from customer service to content generation, where the generation of accurate and ethically sound responses is paramount. Addressing the dual challenges of toxic content generation and the need for sophisticated reasoning, this research introduces a novel approach that integrates targeted pruning techniques to enhance multi-hop reasoning while simultaneously reducing the presence of toxic knowledge. By fine-tuning the pre-trained Llama model, the study explores how selective pruning of toxic pathways can improve the model's accuracy, response diversity, and robustness against adversarial prompts. The findings reveal that pruning not only mitigates harmful content but also significantly enhances the model's reasoning capabilities, demonstrating a meaningful reduction in Average Toxicity Score and an increase in Exact Match and F1 scores. Through a combination of empirical evaluations and robust analysis, the study provides evidence that strategic pruning serves as an effective tool for aligning the ethical and functional performance of language models. This research highlights the potential for pruning techniques to contribute to the development of more reliable, safe, and ethically aligned artificial intelligence systems, ensuring that they meet both societal standards and practical requirements in diverse applications.