On Zero-Initialized Attention: Optimal Prompt and Gating Factor Estimation

TL;DR

This paper provides a theoretical foundation for zero-initialized attention in LLMs, linking it to mixture-of-expert models and demonstrating its effectiveness with linear and non-linear prompts through empirical validation.

Contribution

It offers the first rigorous theoretical analysis of zero-initialized attention, connecting it to mixture-of-expert models and exploring optimal prompt and gating factor estimation.

Findings

Non-linear prompts outperform linear prompts in experiments.

Zero-initialized attention surpasses vanilla attention even with limited data.

Theoretical connection established between zero-initialized attention and mixture-of-expert models.

Abstract

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.