On Zero-Initialized Attention: Optimal Prompt and Gating Factor Estimation
This work provides foundational insights for researchers and practitioners using efficient fine-tuning techniques in large language models, though it is incremental as it builds on existing empirical methods.
The paper tackled the lack of theoretical understanding of zero-initialized attention in LLaMA-Adapter by connecting it to mixture-of-expert models and proving optimal estimation of prompts and gating factors, with non-linear prompts outperforming linear ones on benchmarks and both surpassing vanilla attention even with limited data.
The LLaMA-Adapter has recently emerged as an efficient fine-tuning technique for LLaMA models, leveraging zero-initialized attention to stabilize training and enhance performance. However, despite its empirical success, the theoretical foundations of zero-initialized attention remain largely unexplored. In this paper, we provide a rigorous theoretical analysis, establishing a connection between zero-initialized attention and mixture-of-expert models. We prove that both linear and non-linear prompts, along with gating functions, can be optimally estimated, with non-linear prompts offering greater flexibility for future applications. Empirically, we validate our findings on the open LLM benchmarks, demonstrating that non-linear prompts outperform linear ones. Notably, even with limited training data, both prompt types consistently surpass vanilla attention, highlighting the robustness and adaptability of zero-initialized attention.