🔍Introduction

引言

This research explores the impact of using polite language in interactions with large language models (LLMs). The analysis reveals that politeness has a dual effect on LLM responses, potentially enhancing response accuracy and relevance, while sometimes leading to unnecessary complexity and possible hallucination generation. The magnitude and direction of these effects largely depend on prompt structure, clarity of user intent, and the specific training methodology of the LLM. Using appropriate polite language while ensuring the core question remains clear may be the optimal strategy for balancing these effects.

本研究探討在與大型語言模型（LLM）互動時，使用禮貌性用語對對話結果的影響。分析結果表明，禮貌用語對LLM回應的影響是雙面的，既可能提高回應準確性和相關性，也可能在某些情況下導致不必要的複雜性和潛在的幻覺生成。影響的大小和方向很大程度上取決於提示的結構、用戶意圖的清晰度以及特定LLM的訓練方式。使用適當的禮貌用語並確保核心問題明確可能是平衡這些影響的最佳策略。

In everyday human communication, polite language is an essential component of social etiquette. With the widespread application of LLMs across various domains, people naturally extend this communication habit to human-machine dialogues.

在日常人際交流中，禮貌用語是社交禮節的重要組成部分。隨著LLM在各領域的廣泛應用，人們自然地將這種交流習慣延伸到了人機對話中。

📊Forms of Politeness in LLM Prompts

禮貌用語在LLM提示中的表現形式

Politeness in LLM interactions can take various forms, primarily including:

禮貌用語在與LLM互動中可以采取多種形式，主要包括：

Form of Politeness 禮貌形式	Request words (such as "please", "kindly", "if you could"); Gratitude expressions (such as "thank you", "appreciate it"); Respectful address (using "您" instead of "你" in Chinese); Humble modifiers (such as "if convenient", "may I ask"); Greetings (opening and closing remarks) 請求詞（如「請」、「麻煩」、「勞駕」）；感謝詞（如「謝謝」、「感激」）；尊敬詞（如「您」而非「你」）；謙虛修飾語（如「如果方便的話」、「冒昧請教」）；寒暄語（如開場白和結束語）

These polite expressions are a natural part of human communication, but they may have unexpected effects on LLM processing and responses. LLMs are trained on vast amounts of text data, learning various language patterns and usage contexts, which enables them to recognize and appropriately respond to different degrees of polite expression.

這些禮貌表達方式是人類交流的自然部分，但它們可能對LLM的處理和回應產生意想不到的影響。LLM通過大量文本數據訓練，學習了語言的各種模式和使用情境，這使它們能夠識別並適當回應不同程度的禮貌表達。

🧬Positive Effects of Politeness

禮貌用語對LLM回應的正面影響

Using polite language may positively influence LLM responses in various situations.

使用禮貌用語在多種情況下可能對LLM回應產生積極影響。

🧑‍🦲Improving Response Quality and Completeness

提升回應質量與完整性

Research Finding 研究發現	Example 例子	Explanation 解釋
Politeness influences more comprehensive responses 禮貌影響更全面的回應	"Please explain quantum computing in detail, thank you" vs. "Explain quantum computing" 「請詳細解釋量子計算，謝謝」vs.「解釋量子計算」	Polite requests often get more structured, comprehensive answers 禮貌請求通常獲得更結構化、全面的答案
Training patterns associate politeness with detail 訓練模式將禮貌與詳細關聯	LLMs learn from human communication patterns LLM從人類溝通模式中學習	Models learn connections between polite requests and expectations for detailed responses 模型學習禮貌請求與期望詳細回應之間的關聯

💥Enhancing Intent Understanding and Context Comprehension

增強意圖理解與上下文把握

🦗Reducing Harmful Content Generation

減少潛在的有害內容生成

Polite language may guide LLMs to adopt a more cautious and professional response style, reducing the likelihood of generating inappropriate or controversial content. Polite prompts often set a positive tone for the conversation, potentially activating patterns in LLMs associated with professional, helpful, and responsible responses. （來源）.

禮貌用語可能引導LLM採取更謹慎和專業的回應風格，減少生成不適當或有爭議內容的可能性。禮貌的提示通常設定了對話的正面基調，可能激活LLM中與專業、有幫助且負責任回應相關的模式。

🔬Potential Negative Effects

禮貌用語的潛在負面影響

Despite the many advantages of polite language, it may also have negative effects on LLM responses in certain situations.

儘管禮貌用語有諸多優點，但在某些情況下也可能對LLM的回應產生負面影響。

Increased Prompt Complexity 增加提示複雜性	Excessive politeness may distract the LLM from the core question. For example, "I'm terribly sorry to bother you, if convenient, could you kindly tell me about, if possible, basic information on global warming, really appreciate your help and patience, thank you!" makes it hard to identify the main query. （來源）
Misleading Certainty Threshold 誤導模型的確定性閾值	Extremely polite and humble expressions may unintentionally signal to the LLM that the user is uncertain or has limited knowledge, potentially leading to oversimplified answers.
Increased AI Hallucination Risk 增加AI幻覺的可能性	Research found that highly polite prompts like "Could you kindly share some insights about this obscure historical event?" compared to direct "What is this historical event?" are more likely to cause LLMs to generate plausible-sounding but potentially incorrect content when lacking specific information. （來源）

🧪Contextual Factors

脈絡決定因素

Different types of LLMs may have different sensitivities to polite language, depending on their training data and alignment methods.

不同類型的LLM可能對禮貌用語有不同的敏感度，這取決於其訓練數據和對齊方法。 （來源）.

Language Differences 語言差異	- English: Serves as a baseline language in most research, typically corresponding to a moderate degree of politeness. - Chinese: Displays unique politeness conventions that have specific effects on LLM behavior. - Japanese: Often requires a higher baseline level of politeness than English. - 英語：作為多數研究的基線語言，通常對應中等程度的禮貌要求。 - 中文：展現獨特的禮貌慣例，對LLM行為產生特定影響。 - 日語：通常需要比英語更高的基線禮貌程度。 （來源）
LLM Model Type LLM模型類型	- GPT-4: Relatively insensitive, performance is more stable. - GPT-3.5: Sensitivity between GPT-4 and LLaMA, performance varies. - LLaMA: Highly sensitive, performance often proportional to politeness level. - Smaller models (e.g., LLaMA 8B): More sensitive to prompt quality (including politeness), inappropriate politeness may reduce quality. - GPT-4：相對不敏感，表現較穩定。 - GPT-3.5：敏感度介於GPT-4和LLaMA之間，表現不一。 - LLaMA：高度敏感，表現常與禮貌程度成正比。 - 較小型模型（如LLaMA 8B）：對提示語品質（含禮貌）更敏感，不當禮貌可能降低品質。
RLHF-Tuned Models RLHF微調模型	RLHF fine-tuned models (e.g., GPT-4o, Claude 3) may have "semantic compliance drift" risk with polite/emotional prompts, potentially bypassing safety mechanisms. 經過RLHF微調的模型（如GPT-4o, Claude 3）對禮貌/情感性提示語存在「語義順從漂移」風險，可能繞過安全機制。 （來源）

🧠Best Practices

最佳實踐

Based on understanding the effects of polite language, here are some recommendations for optimizing LLM interactions. （來源）.

Maintain Moderate Politeness 保持適度禮貌	Use basic polite terms (such as "please" and "thank you"), but avoid excessive modifiers and lengthy expressions. 使用基本禮貌用語（如「請」和「謝謝」），但避免過度修飾和冗長的表達。
Structure Prompts Effectively 結構化提示	Separate polite language from the core question, for example, using it at the beginning or end of the prompt, keeping the main question clearly visible. 將禮貌用語與核心問題分開，例如在提示的開始或結束處使用，保持主要問題的清晰可見。
Express Expectations Clearly 明確表達期望	Politely but clearly express expectations for the answer, including the desired level of detail and format. 禮貌地但明確地表達對回答的期望，包括所需的細節級別和格式。

✨Conclusion · 結論

In interactions with LLMs, the influence of polite language is complex and multidimensional. Based on existing research and understanding of how LLMs work, there are both positive effects (such as improving response quality and relevance) and potential negative effects (such as increasing complexity and risk of hallucinations). （來源）.

Overall, moderate politeness combined with clear question formulation seems to be the best strategy. This balance can maintain important elements of politeness in human communication while minimizing interference with LLM processing.

在與LLM互動時，禮貌用語的影響是複雜且多維的。從現有研究和LLM工作原理的理解來看，既有正面影響（如提高回應質量和相關性），也有潛在的負面影響（如增加複雜性和幻覺風險）。總體而言，適度的禮貌用語結合清晰的問題表述似乎是最佳策略。這種平衡可以維持人類交流中重要的禮貌元素，同時最大限度地減少對LLM處理的干擾。 （來源）.