MuEdit
A knowledge editing method for large models in multi-task settings. It mitigates inter-domain knowledge conflicts through conflict-aware ordering and dynamic low-rank approximation while preserving the model's general capabilities.
Timeline (last 90 days)
The study attributes the failure of multi-task editing to "inter-domain null-space misalignment."
Mu-Edit uses conflict-aware ordering to minimize cumulative interference.
Dynamic low-rank approximation is used to expand the null space when the common null space is insufficient.
Experiments were conducted on five functional domains and three backbone models.
Experimental results show that Mu-Edit significantly outperforms existing baselines in multi-task editing performance.
Mu-Edit preserves the model's general capabilities after editing.
Mu-Edit remains effective in scenarios of incremental task arrival.
All Facts (7)
The study attributes the failure of multi-task editing to "inter-domain null-space misalignment."
50%UnverifiedMu-Edit uses conflict-aware ordering to minimize cumulative interference.
50%UnverifiedDynamic low-rank approximation is used to expand the null space when the common null space is insufficient.
50%UnverifiedExperiments were conducted on five functional domains and three backbone models.
50%UnverifiedExperimental results show that Mu-Edit significantly outperforms existing baselines in multi-task editing performance.
50%UnverifiedMu-Edit preserves the model's general capabilities after editing.
50%UnverifiedMu-Edit remains effective in scenarios of incremental task arrival.
50%