Title: Re-examining Parameter-Based Knowledge Editing in Large Language Models: Theoretical Boundaries and Empirical Findings

Abstract:

The technique of parameter-based knowledge editing, which modifies the internal knowledge of large language models (LLMs) through targeted weight adjustments, has garnered considerable interest. Nevertheless, the majority of current approaches neglect essential theoretical constraints and are seldom assessed within practical, real-world scenarios. This study introduces a theoretical framework grounded in the Dimensional Collapse Hypothesis to elucidate how localized parameter modifications can spread along unstable pathways within the representation space. This propagation triggers widespread interference, potentially leading to a collapse in reasoning abilities. Leveraging these insights, we perform a rigorous empirical assessment that systematically manipulates variables such as knowledge complexity, the volume of edits, evaluation metrics, and baseline methodologies. The data indicates that parameter-based editing techniques invariably impair the fundamental capabilities of LLMs. Conversely, a straightforward retrieval-based baseline consistently outperforms all parameter-editing approaches across every tested condition. These outcomes underscore the necessity of prioritizing the preservation of core LLM functionalities following knowledge updates in future research endeavors.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC