MuEdit: An efficient multi-task editing method towards inter-domain knowledge conflicts
Summary
- Researchers propose MuEdit, an efficient multi-task editing method designed to resolve inter-domain knowledge conflicts in large language models.
- Sequential multi-task editing often fails due to inter-domain null-space misalignment, where heterogeneous tasks shrink the common null space.
- MuEdit introduces the Conflict Index to quantify geometric interference, along with conflict-aware ordering and dynamic low-rank approximation.
- Experiments across five functional domains and three model backbones demonstrate that MuEdit outperforms existing baselines while preserving general capabilities.
Why it matters
MuEdit provides a robust approach for sequentially updating large language models across diverse domains without degrading overall performance or colliding with prior knowledge.