Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Beyond Local Edits: Embedding-Virtualized Knowledge for Broader Evaluation and Preservation of Model Editing

Shuainan Liu, Xuanang Chen, Ben He, Le Sun

TL;DR

This work addresses the limited scope of existing knowledge-editing evaluations by introducing Embedding-Virtualized Knowledge (EVK), which probes model knowledge through controlled embedding-space perturbations. It formalizes EVK-Bench, an unsupervised embedding-level benchmark that quantifies knowledge drift via Embedding Stability and Text Stability, revealing shortcomings of LTE-based editors on latent knowledge. To counter drift, the authors propose EVK-Align, a plug-and-play alignment module that regularizes edits against EVK instances using a KL-divergence objective with a progressive top-$k$ mechanism, achieving improved specificity while maintaining editing efficacy and generation quality. Across multiple model families and editing baselines, EVK-Bench and EVK-Align yield more comprehensive evaluation and stronger preservation of neighboring knowledge, offering a practical approach to more controllable model editing.

Abstract

Knowledge editing methods for large language models are commonly evaluated using predefined benchmarks that assess edited facts together with a limited set of related or neighboring knowledge. While effective, such evaluations remain confined to finite, dataset-bounded samples, leaving the broader impact of editing on the model's knowledge system insufficiently understood. To address this gap, we introduce Embedding-Virtualized Knowledge (EVK) that characterizes model knowledge through controlled perturbations in embedding space, enabling the exploration of a substantially broader and virtualized knowledge region beyond explicit data annotations. Based on EVK, we construct an embedding-level evaluation benchmark EVK-Bench that quantifies potential knowledge drift induced by editing, revealing effects that are not captured by conventional sample-based metrics. Furthermore, we propose a plug-and-play EVK-Align module that constrains embedding-level knowledge drift during editing and can be seamlessly integrated into existing editing methods. Experiments demonstrate that our approach enables more comprehensive evaluation while significantly improving knowledge preservation without sacrificing editing accuracy.

Beyond Local Edits: Embedding-Virtualized Knowledge for Broader Evaluation and Preservation of Model Editing

TL;DR

This work addresses the limited scope of existing knowledge-editing evaluations by introducing Embedding-Virtualized Knowledge (EVK), which probes model knowledge through controlled embedding-space perturbations. It formalizes EVK-Bench, an unsupervised embedding-level benchmark that quantifies knowledge drift via Embedding Stability and Text Stability, revealing shortcomings of LTE-based editors on latent knowledge. To counter drift, the authors propose EVK-Align, a plug-and-play alignment module that regularizes edits against EVK instances using a KL-divergence objective with a progressive top- mechanism, achieving improved specificity while maintaining editing efficacy and generation quality. Across multiple model families and editing baselines, EVK-Bench and EVK-Align yield more comprehensive evaluation and stronger preservation of neighboring knowledge, offering a practical approach to more controllable model editing.

Abstract

Knowledge editing methods for large language models are commonly evaluated using predefined benchmarks that assess edited facts together with a limited set of related or neighboring knowledge. While effective, such evaluations remain confined to finite, dataset-bounded samples, leaving the broader impact of editing on the model's knowledge system insufficiently understood. To address this gap, we introduce Embedding-Virtualized Knowledge (EVK) that characterizes model knowledge through controlled perturbations in embedding space, enabling the exploration of a substantially broader and virtualized knowledge region beyond explicit data annotations. Based on EVK, we construct an embedding-level evaluation benchmark EVK-Bench that quantifies potential knowledge drift induced by editing, revealing effects that are not captured by conventional sample-based metrics. Furthermore, we propose a plug-and-play EVK-Align module that constrains embedding-level knowledge drift during editing and can be seamlessly integrated into existing editing methods. Experiments demonstrate that our approach enables more comprehensive evaluation while significantly improving knowledge preservation without sacrificing editing accuracy.
Paper Structure (32 sections, 19 equations, 4 figures, 2 tables)

This paper contains 32 sections, 19 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Memory Mechanisms in LLMs
  4. Model Editing under LTE Paradigm
  5. Methodology
  6. EVK-Bench: Embedding-Virtualized Knowledge Benchmark
  7. Knowledge Simulation
  8. Evaluation Metrics
  9. EVK-Align: Embedding-Virtualized Knowledge Alignment
  10. Alignment Objective
  11. Progressive Regulation
  12. Experiments
  13. Experimental Setup
  14. Main Results
  15. Analysis
  16. ...and 17 more sections

Figures (4)

  • Figure 2: Computational graph of EVK-Align. EVK data is used to compute the pre-edit next-token distribution. After editing, next-token distributions are obtained for both Editing and EVK data: the Editing distribution is optimized via negative log-likelihood to enforce the edit, while the KL divergence between the pre- and post-edit EVK distributions is minimized to preserve original knowledge.
  • Figure 3: Performance of different LTE editing algorithms on the GLUE benchmark. Blue lines represent models enhanced with EVK‑Align, and orange lines correspond to the baseline algorithms, with a total of 1000 edits performed.
  • Figure 4: Performance of different LTE editing algorithms on the ZsRE and Counterfact benchmarks. The pink bars represent models augmented with EVK‑Align, and the blue bars denote the baseline editing algorithms, with a total of 1000 edits performed.
  • Figure 5: Embedding distributions of initial editing points, predefined related knowledge (i.e., neighborhood prompts in Counterfact), and EVK instances after dimensionality reduction, showing that EVK-Bench covers a broader knowledge space.