Reproduce “EvoEdit: Lifelong Free-Text Knowledge Editing through Latent Perturbation Augmentation and Knowledge-driven Parameter Fusion” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Pengfei Cao, Zeao Ji, Daojian Zeng, Jun Zhao, Kang Liu

Researchers introduced the task of Lifelong Free-text Knowledge Editing (LF-Edit) and developed EvoEdit, a method that enables large language models to continuously learn from free-text inputs while effectively preserving previously acquired knowledge. EvoEdit demonstrates enhanced knowledge acquisition and robust mitigation of catastrophic forgetting, achieving a BLEU score of 64.86 on LLaMA-3 for new knowledge after 100 edits, outperforming previous methods.

Key insights

  • Injecting controlled random noise into input embeddings (LPA) during fine-tuning enhances the LLM's ability to acquire complex free-text knowledge, promoting generalization and reducing overfitting.
  • Selectively merging model parameters based on their calculated importance to new knowledge (KPF) is crucial for balancing new learning with the preservation of previously acquired information, effectively preventing catastrophic forgetting.
  • Existing knowledge editing methods, designed for structured data, are largely ineffective and often degrade model performance when applied to the task of lifelong free-text knowledge editing, underscoring the need for specialized approaches.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.