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.

This paper investigates the theory behind LLaMA-Adapter, a fine-tuning method for LLaMA models that uses zero-initialized attention to improve training stability and performance. Although this technique has shown strong practical results, its underlying principles were not well understood. The authors provide a theoretical explanation, linking zero-initialized attention to mixture-of-expert models and proving that both linear and non-linear prompts can be optimally estimated, with non-linear prompts offering greater flexibility for future applications. Experiments on open LLM benchmarks confirm that non-linear prompts outperform linear ones, and importantly, both consistently perform better than standard attention even with limited training data, highlighting the method’s robustness and adaptability.