This paper investigates the challenges of knowledge editing in language models, focusing on the propagation of updated knowledge to interconnected facts. The authors introduce ReCoE, a novel reasoning-based benchmark that evaluates knowledge editing across six common reasoning schemes: superlative, comparative, sorting, counting, aggregation, and subtraction. The dataset includes both factual knowledge and counterfactual knowledge, enabling the assessment of how well models can reason with edited information. The study analyzes existing knowledge editing techniques, including input-augmentation, finetuning, and locate-and-edit. It finds that all methods perform poorly on the ReCoE benchmark, especially in certain reasoning schemes. The analysis reveals key limitations in current knowledge editing methods, including poor fact recall, incoherent reasoning, and inadequate propagation of updated knowledge. The authors propose a framework to evaluate knowledge editing methods based on three key aspects: (1) effectiveness of editing individual facts, (2) accuracy in recalling relevant facts, and (3) logical coherence of the thought process. They also introduce a reasoning-based framework to assess the underlying challenges of knowledge propagation. The study highlights the limitations of current knowledge editing methods, particularly in their ability to effectively propagate new facts for coherent reasoning. It shows that models edited through locate-and-edit methods, such as MEMIT, exhibit a severe decline in generation coherence, leading to nonsensical outputs. The results indicate that these methods fail to maintain the model's fundamental language modeling abilities. The paper concludes that current knowledge editing methods are not effective in propagating updated knowledge for coherent reasoning. It provides a clear direction for future research in this field, aiming to enhance the efficacy and reliability of knowledge editing in computational models. The authors also highlight the importance of using reasoning-based benchmarks to evaluate knowledge editing methods and identify areas for improvement.This paper investigates the challenges of knowledge editing in language models, focusing on the propagation of updated knowledge to interconnected facts. The authors introduce ReCoE, a novel reasoning-based benchmark that evaluates knowledge editing across six common reasoning schemes: superlative, comparative, sorting, counting, aggregation, and subtraction. The dataset includes both factual knowledge and counterfactual knowledge, enabling the assessment of how well models can reason with edited information. The study analyzes existing knowledge editing techniques, including input-augmentation, finetuning, and locate-and-edit. It finds that all methods perform poorly on the ReCoE benchmark, especially in certain reasoning schemes. The analysis reveals key limitations in current knowledge editing methods, including poor fact recall, incoherent reasoning, and inadequate propagation of updated knowledge. The authors propose a framework to evaluate knowledge editing methods based on three key aspects: (1) effectiveness of editing individual facts, (2) accuracy in recalling relevant facts, and (3) logical coherence of the thought process. They also introduce a reasoning-based framework to assess the underlying challenges of knowledge propagation. The study highlights the limitations of current knowledge editing methods, particularly in their ability to effectively propagate new facts for coherent reasoning. It shows that models edited through locate-and-edit methods, such as MEMIT, exhibit a severe decline in generation coherence, leading to nonsensical outputs. The results indicate that these methods fail to maintain the model's fundamental language modeling abilities. The paper concludes that current knowledge editing methods are not effective in propagating updated knowledge for coherent reasoning. It provides a clear direction for future research in this field, aiming to enhance the efficacy and reliability of knowledge editing in computational models. The authors also highlight the importance of using reasoning-based benchmarks to evaluate knowledge editing methods and identify areas for improvement.