Here is the translated data: This article introduces recent research findings that prefix language models (prefixLM), based on Transformer architecture, demonstrate superior performance in contextual learning compared to causal language models (causalLM). The paper, through theoretical analysis and experimental validation, reveals the advantages of prefix language models in convergence behavior and optimization solutions, and provides corresponding open-source code. This research holds significant importance for understanding the working principles and optimization behaviors of language models in contextual learning.