EvoEdit: Lifelong Free-Text Knowledge Editing through Latent Perturbation Augmentation and Knowledge-driven Parameter Fusion

TL;DR

LF-Edit addresses the challenge of updating outdated knowledge in deployed LLMs by enabling lifelong, free-text edits. The authors introduce MRLF-Bench to evaluate such edits and EvoEdit to inject knowledge via Latent Perturbation Augmentation and preserve history via Knowledge-driven Parameter Fusion. The results show EvoEdit outperforms existing editing methods on new knowledge acquisition while maintaining previous knowledge, across a large, multi-rank benchmark. The work advances practical continual learning for LLMs and provides a foundation for future free-text editing research.

Abstract

Adjusting the outdated knowledge of large language models (LLMs) after deployment remains a major challenge. This difficulty has spurred the development of knowledge editing, which seeks to accurately and efficiently modify a model's internal (parametric) knowledge without retraining it from scratch. However, existing methods suffer from two limitations. First, they depend on structured triplets that are misaligned with the free-text nature of LLM pretraining and fail to capture the nuanced relationships among facts. Second, they typically support one-time knowledge updates, with relatively limited research on the problem of sequential or lifelong editing. To address these gaps, we propose a new task, Lifelong Free-text Knowledge Editing (LF-Edit), which enables models to incorporate updates expressed in natural language and supports continual editing over time. Despite its promise, LF-Edit faces the dual challenge of integrating new knowledge while mitigating the forgetting of prior information. To foster research on this new task, we construct a large-scale benchmark, Multi-Rank Lifelong Free-text Editing Benchmark (MRLF-Bench), containing 16,835 free-text edit requests. We further design a cognitively inspired multi-rank evaluation framework encompassing four levels: memorization, understanding, constrained comprehension, and reasoning. To tackle the challenges inherent in LF-Edit, we introduce a novel approach named EvoEdit that enhances knowledge injection through Latent Perturbation Augmentation and preserves prior information via Knowledge-driven Parameter Fusion. Experimental results demonstrate that EvoEdit substantially outperforms existing knowledge editing methods on the proposed LF-Edit task.

Figures (6)

• Figure 1: Overview of the lifelong free-text knowledge editing (LF-Edit) task. The target of the task is to enable LLMs to continuously acquire knowledge from free text and efficiently update the internal knowledge of the model.
• Figure 2: An example of a free-text knowledge editing instance. The edit request serves to update the model, while multi-rank queries are employed to assess the performance of the edited model.
• Figure 3: The architecture of the proposed EvoEdit for the lifelong free-text editing task, which consists of two key modules: 1) Latent Perturbation Augmentation ($\S$\ref{['PA']}), which injects the noise into the embedding layer to enhance the diversity of input representations, thereby improving the model’s capacity to comprehend and assimilate new knowledge. 2) Knowledge-driven Parameter Fusion ($\S$\ref{['KPF']}), which identifies and selects the most critical parameters across multiple models for integration, enabling the preservation of previously acquired knowledge.
• Figure 4: BLEU scores on historical edits after after editing 500 instances for the knowledge editing methods. T=X represents the performance of the edited model on multi-rank queries at step X/50 (editing 50 instances at each step).
• Figure 5: The performance comparison between different parameter fusion strategies on the LLaMA-2 and LLaMA-3, respectively. "DPF" represents the direct fusion of the original model, the edited model from the last step, and all parameters of the current model, without considering the importance of the parameters.