Reproduce “Edit Less, Achieve More: Dynamic Sparse Neuron Masking for Lifelong Knowledge Editing in LLMs” 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
Jinzhe Liu, Junshu Sun, Shufan Shen, Chenxue Yang, Shuhui Wang
Neuron-Specific Masked Knowledge Editing (NMKE) introduces a fine-grained, neuron-level method for updating Large Language Models, which dynamically identifies and selectively modifies relevant neurons to embed new knowledge. This approach effectively maintains high editing accuracy and preserves the model's general capabilities across thousands of sequential edits, outperforming previous methods in lifelong learning scenarios.
Key insights
- LLM FFN neurons can be functionally distinguished into 'knowledge-general' (fundamental information) and 'knowledge-specific' (contextual details), a distinction crucial for targeted editing.
- Dynamically selecting neurons for masking based on entropy-guided ratios for both knowledge-general and knowledge-specific types optimally balances editing success and general capability preservation.
- Precise, neuron-level sparse modifications are essential for minimizing internal parameter disruption and preventing catastrophic forgetting, enabling robust lifelong knowledge editing.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.