Empirical Study on Updating Key-Value Memories in Transformer Feed-forward Layers
This work provides incremental insights into understanding and optimizing transformer architectures for specific tasks like knowledge editing and fine-tuning.
The study empirically compares updating keys versus values in transformer feed-forward layers during knowledge editing and fine-tuning tasks for large language models, finding that tuning keys is more effective for knowledge editing while tuning values is better for fine-tuning.
The feed-forward networks (FFNs) in transformers are recognized as a group of key-value neural memories to restore abstract high-level knowledge. In this work, we conduct an empirical ablation study on updating keys (the 1st layer in the FFNs layer) or values (the 2nd layer in the FFNs layer). We compare those two methods in various knowledge editing and fine-tuning tasks of large language models to draw insights to understand FFNs further. Code is available at $\href{https://github.com/qiuzh20/Tuning-keys-v.s.-values}{this\,repo}$.