On the Superimposed Noise Accumulation Problem in Sequential Knowledge Editing of Large Language Models

TL;DR

The paper identifies superimposed noise as a key limitation in sequential knowledge editing of large language models and analyzes it through a decomposition of updates into influence and activation components. It introduces DeltaEdit, a method with dynamic orthogonal constraints that project new edits into a null space relative to history updates, coupled with a dynamic threshold to curb interference. Empirical results show DeltaEdit substantially improves editing performance (e.g., 16.8% over AlphaEdit on CounterFact with Llama3-8B) while preserving hidden representations and general capabilities. This work advances robust, long-horizon sequential knowledge editing by reducing noise accumulation and maintaining model reliability.

Abstract

Sequential knowledge editing techniques aim to continuously update knowledge in large language models at low cost, preventing models from generating outdated or incorrect information. However, existing sequential editing methods suffer from a significant decline in editing success rates after long-term editing. Through theoretical analysis and experiments, our findings reveal that as the number of edits increases, the model's output increasingly deviates from the desired target, leading to a drop in editing success rates. We refer to this issue as the superimposed noise accumulation problem. Our further analysis demonstrates that the problem is related to the erroneous activation of irrelevant knowledge and conflicts between activated knowledge. Based on this analysis, a method named DeltaEdit is proposed that reduces conflicts between knowledge through dynamic orthogonal constraint strategies. Experiments show that DeltaEdit significantly reduces superimposed noise, achieving a 16.8% improvement in editing performance over the strongest baseline.

Figures (11)

• Figure 1: Output deviation problem of edited LLM. The correctly activated output (green part) is interfered with by the incorrectly activated output (red parts).
• Figure 2: The changes in Efficacytop and $noise_E$ with the number of edits. The left figure displays results for GPT2-XL using AlphaEdit and MEMIT, the right figure shows results for LLaMA3-8B using only AlphaEdit. MEMIT is excluded from the figure of LLaMA3-8B due to its excessively high $noise_E$, making visualization difficult.
• Figure 3: The variation curve of $k^\top \beta$ for AlphaEdit and MEMIT as the number of edits increases, with GPT2-XL serving as the edited model. As the results of Llama3-8B are not suitable for display in a line chart, they are presented in Appendix \ref{['sec:sup-llama']}
• Figure 4: The changes in Efficacytop and $noise_E$ with the number of edits. The top figure shows results for GPT2-XL, the bottom figure shows results for LLaMA3-8B.
• Figure 5: The distribution of hidden representations of pre-edited and post-edited Llama3-8B after dimensionality reduction. The top and right curve graphs display the marginal distributions for two reduced dimensions. The dashed lines represent the 0.95 confidence intervals.