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