研究提示词如何影响模型内部表示，发现相关提示不必然带来更好表征。
Do Prompts Reshape Representations? An Empirical Study of Prompting Effects on Embeddings
- 通过探测实验分析不同提示模板对嵌入的影响。
- 提示相关性与表征质量无一致关联，高相关提示未必提升性能。
- 挑战了'提示越相关越好'的直觉，适合关注提示机制的研究者。
提示词是零样本场景下利用语言模型的常用方法。然而，预训练模型在无任务特定监督下完成多样化任务的内在机制仍不明确。研究提示词与内部表征质量之间的关系，有助于理解预训练嵌入如何支持上下文中的任务求解。本实证研究对多种零样本分类任务的提示嵌入进行了系统探测实验，分析了不同提示模板组合的效果。结果表明，虽然提示会影响表征质量，但这种变化与提示对目标任务的相关性并无一致关联。这一发现质疑了‘更相关提示必然带来更好表征’的假设。我们进一步探讨了可能导致该反常行为的潜在因素。
原文摘要 · Abstract (English)
Prompting is a common approach for leveraging LMs in zero-shot settings. However, the underlying mechanisms that enable LMs to perform diverse tasks without task-specific supervision remain poorly understood. Studying the relationship between prompting and the quality of internal representations can shed light on how pre-trained embeddings may support in-context task solving. In this empirical study, we conduct a series of probing experiments on prompt embeddings, analyzing various combinations of prompt templates for zero-shot classification. Our findings show that while prompting affects the quality of representations, these changes do not consistently correlate with the relevance of the prompts to the target task. This result challenges the assumption that more relevant prompts necessarily lead to better representations. We further analyze potential factors that may contribute to this unexpected behavior.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。