Lifelong Knowledge Editing requires Better Regularization

Akshat Gupta, Phudish Prateepamornkul, Maochuan Lu, Ahmed Alaa, Thomas Hartvigsen, Gopala Anumanchipalli

TL;DR

This paper tackles the problem of model degradation in large language models (LLMs) when performing many sequential knowledge edits. The authors propose two novel regularization techniques, "Most-Probable Early Stopping" (MPES) and a Frobenius norm constraint, which significantly improve the stability and efficiency of knowledge editing, enabling up to 10,000 edits with minimal performance loss.

What Problem Does This Paper Address?

Large language models (LLMs) are powerful, but their knowledge is static. Knowledge editing aims to update this knowledge efficiently without full retraining. A popular approach, "locate-then-edit," identifies specific model parameters to modify. However, when applied at scale (thousands of edits), these methods suffer from significant "model degradation," meaning their general performance on other tasks drops. This paper investigates the root causes of this degradation and proposes solutions to enable "lifelong knowledge editing" – the ability to continuously update LLMs over time without sacrificing their overall capabilities.

Methods

The authors formalize the "locate-then-edit" process into a two-step fine-tuning procedure:

  1. Activation Optimization: A gradient-descent step to find target activation vectors that would produce the desired edited fact.
  2. Weight Update: A least-squares update to modify the model's weights using the target activations.

To combat degradation, they introduce two regularization techniques:

  • Most-Probable Early Stopping (MPES): Instead of stopping gradient descent based on a fixed number of steps or a loss threshold, MPES halts the process when the target fact becomes the most probable output across various contexts. This prevents over-optimization of internal activations.
  • Frobenius Norm Constraint: This technique adds a penalty to the optimization objective to limit the growth of the Frobenius norm of the edited weight matrices. This addresses the issue of edited matrices disproportionately influencing the model's output.

Datasets

The authors evaluate their methods on the following datasets:

  • CounterFact: A standard dataset for knowledge editing.
  • zsRE: Another standard dataset for knowledge editing.

They use three representative LLMs for their experiments:

  • GPT2-XL
  • Llama2-7B
  • Llama3-8B

Key Insights and Ideas

  • Degradation has two main culprits: Over-optimization of internal activations during the gradient-descent step and uncontrolled growth of the Frobenius norm in edited weight matrices.
  • MPES prevents over-optimization: By stopping gradient descent when a fact is most probable, it avoids assigning unnaturally high confidence to edited facts, leading to more natural recall.
  • Norm growth dominates residual streams: As edited matrices grow in norm, their outputs increasingly dominate the model's internal "residual stream," overriding information from other parts of the model and causing degradation.
  • Targeted regularization is key: Applying regularization at specific stages (gradient descent for MPES, weight update for norm constraint) is crucial for effective knowledge editing.
  • Combining techniques is powerful: MPES and the norm constraint address complementary failure modes, and their combination yields the best results.

Key Results

  • The proposed methods (MPES and Frobenius norm constraint) enable up to 10,000 sequential edits while maintaining downstream performance.
  • Editing time is reduced by 42-61% compared to standard methods.
  • The combined approach (MEMIT + MPES + NC) significantly outperforms baseline methods like MEMIT, AlphaEdit, PRUNE, and RECT across various editing and downstream performance metrics.
  • For example, on Llama3-8B, the combined approach achieves an Edit Score of 85.72, compared to 49.68 for baseline MEMIT, after 10,000 edits.
  • The norm constraint is shown to be critical; removing it from AlphaEdit leads to complete model collapse.

Limitations and Future Work

  • Computational Cost: Scaling beyond 10,000 edits still remains challenging due to compute constraints. Performing 10,000 edits on Llama3-8B took a full day on their hardware, partly due to extensive analysis.
  • Future Work: The authors suggest exploring more efficient optimization techniques or hardware acceleration to push beyond current limits. They also note that while MPES reduces norm growth, it's still about 3 times higher than the unedited model, indicating that explicit norm control is still needed.

Summary updated

Abstract

Knowledge editing is a promising way to improve factuality in large language models, but recent studies have shown significant model degradation during sequential editing. In this paper, we formalize the popular locate-then-edit methods as a two-step fine-tuning process, allowing us to precisely identify the root cause of this degradation. We show that model degradation occurs due to (1) over-optimization of internal activations and (2) continuous norm-growth of edited matrices. To mitigate these issues, we introduce two regularization techniques: (1) Most-Probable Early Stopping (MPES) and (2) explicit Frobenius norm-constraint. We demonstrate that applying these simple yet effective regularization techniques at key points in the editing process can substantially mitigate model degradation. Combining these regularization methods enables scaling locate-then-edit methods to 10,000 edits while reducing editing time by 42-61%. These results show that targeted regularization is essential for lifelong knowledge editing.
BibTeX
@inproceedings{gupta2025lifelong,
  author = {Akshat Gupta and Phudish Prateepamornkul and Maochuan Lu and Ahmed Alaa and Thomas Hartvigsen and Gopala Anumanchipalli},
  title = {{Lifelong Knowledge Editing requires Better Regularization}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure