Decomposing Task Vectors for Improved Model Editing
Analysis
This paper addresses a key limitation in using task vectors for model editing: the interference of overlapping concepts. By decomposing task vectors into shared and unique components, the authors enable more precise control over model behavior, leading to improved performance in multi-task merging, style mixing in diffusion models, and toxicity reduction in language models. This is a significant contribution because it provides a more nuanced and effective way to manipulate and combine model behaviors.
Key Takeaways
- •Proposes a decomposition method for task vectors to separate shared and unique knowledge.
- •Improves multi-task merging, style mixing, and toxicity reduction in different model types.
- •Addresses the problem of overlapping concepts in task vector arithmetic.
- •Offers a new framework for understanding and controlling task vector arithmetic.
Reference
“By identifying invariant subspaces across projections, our approach enables more precise control over concept manipulation without unintended amplification or diminution of other behaviors.”