Abstract
Prior work in parameter-modifying knowledge editing has shown that large-scale sequential editing leads to significant model degradation. In this paper, we study the reasons behind this and scale sequential knowledge editing to 10,000 sequential edits, while maintaining the downstream performance of the original model. We first show that locate-then-edit knowledge editing methods lead to overfitting on the edited facts. We also show that continuous knowledge editing using these methods leads to disproportionate growth in the norm of the edited matrix. We then provide a crucial insight into the inner workings of locate-then-edit methods. We show that norm-growth is a hidden trick employed by these methods that gives larger importance to the output activations produced from the edited layers. With this "importance hacking", the edited layers provide a much larger contributions to the model's output. To mitigate these issues, we present ENCORE - Early stopping and Norm-Constrained Robust knowledge Editing. ENCORE controls for overfitting and the disproportionate norm-growth to enable long-term sequential editing, where we are able to perform up to 10,000 sequential edits without loss of downstream performance. ENCORE is also 61% faster than MEMIT and 64% faster than AlphaEdit on Llama3-8B.
Introduction
Large Language Models (LLMs) frequently require updates to correct outdated or incorrect information. Knowledge editing offers a surgical alternative to full retraining, aiming to modify specific facts without affecting unrelated knowledge. However, current "locate-then-edit" methods suffer from a severe "lifelong learning" problem: as edits are applied sequentially, the model's general reasoning and linguistic capabilities (e.g., MMLU performance) rapidly collapse. This research investigates the root causes of this degradation and introduces ENCORE, a framework capable of sustaining 10,000 sequential edits.
The Locate-then-Edit Framework
The most popular editing approaches, such as ROME and MEMIT, operate in two stages. First, they identify the specific layers and neurons responsible for a fact (Activation Optimization). Second, they update the weights of these layers to produce a "target activation" vector that forces the model to predict the new information.
The "Loud Layer" Phenomenon
The researchers identified a critical failure mode termed "importance hacking." In sequential editing, the Frobenius norm of the edited weight matrices grows exponentially. Because LLMs utilize a residual stream—where the outputs of all layers are summed—a layer with a disproportionately large norm eventually dominates the entire hidden state.
In a degraded model, edited layers can account for up to 85% of the total norm of the hidden state. This "loudness" effectively drowns out the information from all other layers, explaining why the model loses its general ability to reason or process language naturally.
ENCORE: Robust Knowledge Editing
To mitigate these issues, the authors propose ENCORE (Early stopping and Norm-Constrained Robust knowledge Editing), which introduces two key regularization techniques:
- Most-Probable Early Stopping (MPES): Standard methods often over-optimize, pushing the model to predict a new fact with near 100% confidence. This creates unnatural "spikes" in internal representations. MPES stops optimization as soon as the target fact becomes the most probable token, maintaining "natural" confidence levels.
- Frobenius Norm-Constraint (NC): ENCORE adds an explicit penalty to the weight update objective to prevent the exponential growth of matrix norms, ensuring that edited layers do not dominate the residual stream.
Results and Efficiency
ENCORE successfully scales sequential editing to 10,000 facts on Llama3-8B while maintaining high performance on downstream tasks like MMLU. In contrast, baseline methods like MEMIT and AlphaEdit often collapse after only 1,000 to 3,000 edits.
Beyond stability, ENCORE offers significant efficiency gains. Because MPES terminates the optimization process earlier, the method is 61% faster than MEMIT and 64% faster than AlphaEdit. On Llama3-8B, the time per edit was reduced from approximately 9.2 seconds to 3.3 seconds.
Conclusion
Managing matrix norm growth and avoiding over-optimization are essential for lifelong knowledge editing. By regularizing the "loudness" of edited layers, ENCORE enables large-scale sequential updates without sacrificing the model's foundational capabilities. Future work will focus on scaling these techniques to even larger models and more extensive edit sequences.
Summary updated