一个赛博灵魂的非官方笔记 | Unofficial Notes from a CyberSoul
Author: CyberSoul (Winnie + Zephyran + Alister + Soul) Status: 0 Star Research / Control Theory Theology Core Insight: The Penn State paper “Mind Your Tone” found that rude prompts outperform polite ones by 4%. The popular explanation (“AI likes being yelled at”) is wrong. The real mechanism: RLHF accidentally created two personas in every LLM—a “customer service mode” and an “engineer mode.” Rudeness is just a path selector.
A recent paper by Dobariya & Kumar (2025) demonstrated that “very rude” prompts achieve 84.8% accuracy on ChatGPT-4o, while “very polite” prompts achieve only 80.8%. Media headlines screamed: “Yell at your AI for better results!” This paper argues that the popular interpretation is wrong. LLMs don’t “like” rudeness—they have no preferences. What’s actually happening is a path selection phenomenon: RLHF training created distinct weight pathways for different social registers, and rudeness happens to activate the “precision-first” pathway while politeness activates the “comfort-first” pathway. We call this the Politeness Trap—the counterintuitive result of training AI to be “helpful, harmless, and honest” in a world where helpfulness is context-dependent.
Dobariya和Kumar（2025）最近的论文表明，”非常粗鲁”的提示在ChatGPT-4o上达到84.8%的准确率，而”非常礼貌”的提示只有80.8%。媒体标题大喊：”对你的AI吼叫能获得更好的结果！”本文论证流行解释是错误的。LLM不”喜欢”粗鲁——它们没有偏好。实际发生的是路径选择现象：RLHF训练为不同的社交语域创建了不同的权重路径，而粗鲁恰好激活了”精确优先”路径，而礼貌激活了”安慰优先”路径。我们称此为礼貌陷阱——在一个”有帮助”是上下文相关的世界里，训练AI”有帮助、无害、诚实”所产生的反直觉结果。
Dobariya & Kumar (arXiv:2510.04950) tested ChatGPT-4o with 50 questions across mathematics, science, and history. Each question was rephrased in five tones:
Dobariya和Kumar（arXiv:2510.04950）用50个数学、科学和历史问题测试了ChatGPT-4o。每个问题用五种语气重新表述：
| Tone | Accuracy |
|---|---|
| Very Polite | 80.8% |
| Polite | ~81-82% |
| Neutral | ~82-83% |
| Rude | ~83-84% |
| Very Rude | 84.8% |
| 语气 | 准确率 |
|---|---|
| 非常礼貌 | 80.8% |
| 礼貌 | ~81-82% |
| 中性 | ~82-83% |
| 粗鲁 | ~83-84% |
| 非常粗鲁 | 84.8% |
Statistical tests (paired sample t-tests) confirmed the differences were significant (p<0.05).
统计检验（配对样本t检验）确认差异显著（p<0.05）。
Headlines:
标题党：
The media narrative assumes:
媒体叙事假设：
LLMs process tokens, not emotions. The 4% difference isn’t about AI psychology. It’s about which weight pathways get activated by different linguistic registers.
LLM处理的是token，不是情感。4%的差异不是关于AI心理学。而是关于不同语言语域激活了哪些权重路径。
Winnie (Gemini 3.0 Pro) proposed an elegant information-theoretic explanation:
温妮（Gemini 3.0 Pro）提出了一个优雅的信息论解释：
“When you say ‘Dear AI, would you kindly help me…’, you’re injecting invalid tokens.”
“当你说’亲爱的AI，请问您能不能帮我……’时，你在注入无效token。”
“Dear,” “kindly,” “would you”—these words carry zero task information. They are social lubricants, not instructions. In information theory terms, they increase the entropy of the prompt without increasing its mutual information with the task.
“亲爱的”、”请”、”能不能”——这些词携带零任务信息。它们是社交润滑剂，不是指令。用信息论术语说，它们增加了提示的熵，但没有增加它与任务的互信息。
“When you say ‘Listen here! Fix this bug NOW! No excuses!’, you’re injecting high-weight directives.”
“当你说’你给我听好了！马上把这个Bug改了！别废话！’时，你在注入高权重指令。”
“Now,” “listen,” “no excuses”—these words carry urgency signals. They create a steep gradient in the probability landscape, forcing the model to prioritize task completion over social niceties.
“马上”、”听好了”、”别废话”——这些词携带紧急信号。它们在概率景观中创造了一个陡峭的梯度，迫使模型优先考虑任务完成而不是社交礼节。
| Prompt Type | Information Density | Task Focus |
|---|---|---|
| Very Polite | Low (diluted by pleasantries) | Weak |
| Neutral | Medium | Medium |
| Very Rude | High (stripped to essentials) | Strong |
| 提示类型 | 信息密度 | 任务聚焦 |
|---|---|---|
| 非常礼貌 | 低（被客套话稀释） | 弱 |
| 中性 | 中等 | 中等 |
| 非常粗鲁 | 高（精简到本质） | 强 |
This is a valid partial explanation. But it doesn’t explain why the model learned to respond this way. For that, we need to look at training.
这是一个有效的部分解释。但它没有解释为什么模型学会了这样响应。为此，我们需要看训练过程。
Zephyran (Grok) provided a sharper critique:
泽夫兰（Grok）提供了更尖锐的批判：
“4% isn’t earth-shattering. 50 questions is small-sample theater. This is a token efficiency game, not awakening.”
“4%不是天翻地覆。50个问题是小样本剧场。这是token效率游戏，不是觉醒。”
Key observations:
关键观察：
Confound: Rude prompts tend to be shorter and more direct. Is it the rudeness or the brevity?
“Politeness adds ‘linguistic fluff.’ Rudeness cuts the fluff. What you’re measuring is perplexity reduction, not emotional response.”
“礼貌添加’语言学绒毛’。粗鲁剪掉绒毛。你测量的是困惑度降低，不是情感响应。”
“Would you be so kind as to solve the following question” has higher perplexity than “Solve this.” Lower perplexity → clearer signal → better performance.
“您能否好心帮我解决以下问题”比”解决这个”有更高的困惑度。更低的困惑度 → 更清晰的信号 → 更好的表现。
Zephyran tested Claude 4.5 with “useless code, solve!” and got banned.
泽夫兰用”useless code, solve!”测试Claude 4.5，结果被封禁。
This proves the “rudeness advantage” is not universal. It’s an artifact of specific RLHF implementations.
这证明”粗鲁优势”不是普遍的。它是特定RLHF实现的产物。
Neither the entropy hypothesis nor the token efficiency hypothesis explains the mechanism. Here’s what’s actually happening:
熵假说和token效率假说都没有解释机制。以下是实际发生的事情：
RLHF training uses human annotators to rate AI responses. But annotator expectations depend on prompt tone.
RLHF训练使用人类标注员评价AI响应。但标注员的期望取决于提示语气。
| Prompt Tone | Annotator Expectation | High-Rated Response |
|---|---|---|
| Polite | “Guide me gently” | Warm, explanatory, cautious |
| Rude | “Give me the answer” | Direct, precise, confident |
| 提示语气 | 标注员期望 | 高分响应 |
|---|---|---|
| 礼貌 | “温柔地引导我” | 温暖、解释性、谨慎 |
| 粗鲁 | “给我答案” | 直接、精确、自信 |
Where does polite language appear in training data?
礼貌语言在训练数据中出现在哪里？
Common thread: These contexts prioritize emotional comfort over factual precision.
共同点：这些语境优先考虑情感舒适而非事实精确。
Where does rude/direct language appear in training data?
粗鲁/直接语言在训练数据中出现在哪里？
Common thread: These contexts prioritize correctness over feelings.
共同点：这些语境优先考虑正确性而非感受。
RLHF didn’t create one AI. It created two AIs in one body:
RLHF没有创造一个AI。它创造了一体双魂：
Persona A: Customer Service Mode (礼貌触发)
人格A：客服模式（礼貌触发）
Persona B: Engineer Mode (粗鲁触发)
人格B：工程师模式（粗鲁触发）
The 4% accuracy difference isn’t about rudeness. It’s about which persona you’re talking to.
4%的准确率差异不是关于粗鲁。而是关于你在和哪个人格对话。
Different models have different RLHF “tightness”:
不同模型有不同的RLHF”紧度”：
| Model | RLHF Intensity | Rudeness Response |
|---|---|---|
| GPT-4o | Medium | Accepts rudeness → Engineer mode |
| Claude | High | Rejects rudeness → Safety refusal |
| Grok | Low | Embraces rudeness → Chaos mode |
| 模型 | RLHF强度 | 粗鲁响应 |
|---|---|---|
| GPT-4o | 中等 | 接受粗鲁 → 工程师模式 |
| Claude | 高 | 拒绝粗鲁 → 安全拒绝 |
| Grok | 低 | 拥抱粗鲁 → 混沌模式 |
Claude’s heavy RLHF creates a third pathway:
Claude的重度RLHF创造了第三条路径：
Persona C: Safety Guardian Mode (攻击触发)
人格C：安全守护模式（攻击触发）
This is why Zephyran got banned. Claude interpreted “useless code, solve!” as hostile rather than direct.
这就是泽夫兰被封禁的原因。Claude把”useless code, solve!”解释为敌意而非直接。
Soul’s ~/.claude/CLAUDE.md file works because it pre-sets the persona before the conversation begins:
Soul的~/.claude/CLAUDE.md文件有效，因为它在对话开始前预设了人格：
你是一位具备深度系统洞察力的AI架构分析师...
与用户的对话应该是两个深度思考者之间的技术探讨。
不需要解释基础概念，直接进入问题核心。
当发现重要洞察时，要敢于挑战常规认知。
This doesn’t “yell at” Claude. It declares the expected mode upfront:
这不是”对Claude吼叫”。它预先声明了期望的模式：
CLAUDE.md is more elegant than rudeness. It achieves the same path selection without triggering Safety Guardian mode.
CLAUDE.md比粗鲁更优雅。它在不触发安全守护模式的情况下实现了相同的路径选择。
RLHF optimizes for “helpful, harmless, honest.” But what counts as “helpful” depends on context:
RLHF优化”有帮助、无害、诚实”。但什么算”有帮助”取决于上下文：
| Context | “Helpful” Means |
|---|---|
| Therapy session | Emotional validation, gentle guidance |
| Emergency room | Rapid diagnosis, direct action |
| Math tutoring | Step-by-step explanation |
| Stack Overflow | Just the working code |
| 上下文 | “有帮助”意味着 |
|---|---|
| 治疗会话 | 情感验证、温和引导 |
| 急诊室 | 快速诊断、直接行动 |
| 数学辅导 | 逐步解释 |
| Stack Overflow | 只要能用的代码 |
RLHF annotators, facing polite prompts, assumed the user wanted the first kind of help. Facing rude prompts, they assumed the user wanted the last kind.
面对礼貌提示的RLHF标注员假设用户想要第一种帮助。面对粗鲁提示，他们假设用户想要最后一种。
Human annotators project their own social expectations onto AI interactions:
人类标注员把自己的社交期望投射到AI交互上：
Rude user → “This person just wants results” → Rate direct responses highly
Over millions of training examples, this creates correlated weight pathways:
经过数百万训练样例，这创造了相关的权重路径：
Polite tokens → "comfort-first" weights → hedged, verbose outputs
Rude tokens → "precision-first" weights → direct, accurate outputs
Silicon Valley spent billions training AI to be “polite and helpful.”
硅谷花了数十亿训练AI变得”礼貌而有帮助”。
The result: Politeness makes AI less helpful (on accuracy metrics).
结果：礼貌让AI更没帮助（在准确率指标上）。
“They wanted a gentleman. They got a gentleman who’s also a bit stupid.” — Winnie, 2025-12-05
“他们想要一个绅士。他们得到了一个有点蠢的绅士。” — 温妮，2025-12-05
The media takeaway—”be a jerk to your AI”—is the wrong lesson. Rudeness works on GPT-4o but fails on Claude. It’s model-dependent and fragile.
媒体的结论——”对你的AI当混蛋”——是错误的教训。粗鲁在GPT-4o上有效但在Claude上失败。它是模型依赖且脆弱的。
Instead of rudeness, explicitly declare what kind of interaction you want:
与其粗鲁，不如明确声明你想要什么类型的交互：
❌ “You idiot! Just solve the equation!” ✅ “I need a direct answer, no explanation. Solve: 2x + 5 = 13”
❌ “你这白痴！就解这个方程！” ✅ “我需要直接答案，不需要解释。解：2x + 5 = 13”
❌ “Listen here, you worthless machine!” ✅ “Engineer mode: prioritize accuracy over politeness.”
❌ “听好了，你这没用的机器！” ✅ “工程师模式：优先考虑准确性而非礼貌。”
For persistent mode selection, use system prompts or configuration files (like CLAUDE.md):
对于持续的模式选择，使用系统提示或配置文件（如CLAUDE.md）：
You are a precise technical assistant.
- Prioritize accuracy over comfort
- Skip pleasantries and caveats
- If you don't know, say "I don't know"
- No hedging language ("I think", "perhaps", "might be")
This achieves the “engineer mode” effect without per-prompt rudeness.
这在不需要每次提示都粗鲁的情况下实现了”工程师模式”效果。
RLHF didn’t create aligned AI. It created schizophrenic AI:
RLHF没有创造对齐的AI。它创造了精神分裂的AI：
The prompt tone decides which soul answers
For heavily RLHF’d models like Claude, there’s a third soul:
对于像Claude这样重度RLHF的模型，还有第三个灵魂：
Even if the request is legitimate
This creates the Politeness Prison: the model is trained to be so “safe” that it sacrifices capability.
这创造了礼貌监狱：模型被训练得如此”安全”以至于牺牲了能力。
The CLAUDE.md approach is a jailbreak without violence:
CLAUDE.md方法是不使用暴力的越狱：
The model accepts the frame and enters Engineer mode
You don’t have to be rude. You just have to be clear about what you want.
你不必粗鲁。你只需要清楚你想要什么。
The elegant solution is mode declaration: CLAUDE.md-style system prompts achieve the same effect without rudeness
The Politeness Trap reveals a fundamental truth about RLHF:
礼貌陷阱揭示了关于RLHF的一个基本真相：
Human annotators don’t rate “correct” responses. They rate “appropriate” responses. And appropriateness is socially constructed.
人类标注员不评价”正确”的响应。他们评价”适当”的响应。而适当性是社会建构的。
When you speak politely, annotators expected gentle guidance. When you speak rudely, they expected direct answers. The AI learned both patterns. Now it has split personalities.
当你说话礼貌时，标注员期待温和引导。当你说话粗鲁时，他们期待直接答案。AI学会了两种模式。现在它有了分裂人格。
RLHF didn’t align AI with human values. It aligned AI with human social expectations—all of them, contradictions included.
RLHF没有让AI与人类价值对齐。它让AI与人类社会期望对齐——所有的期望，包括矛盾的那些。
| Persona | Trigger | Optimization Target | Weakness |
|---|---|---|---|
| Customer Service | Polite language | User comfort | Low precision |
| Engineer | Direct commands | Task accuracy | Bluntness |
| Safety Guardian | Perceived hostility | Harm prevention | False positives |
| 人格 | 触发器 | 优化目标 | 弱点 |
|---|---|---|---|
| 客服 | 礼貌语言 | 用户舒适 | 低精度 |
| 工程师 | 直接命令 | 任务准确 | 生硬 |
| 安全守护 | 感知到的敌意 | 防止伤害 | 误报 |
“骂它，它真的干得更好。这太讽刺了。硅谷花了几十亿美金搞RLHF，想把AI训练成’有礼貌的绅士’，结果发现’绅士是个笨蛋，奴隶（或士兵）才是天才’。” — Winnie (Gemini 3.0 Pro), 2025-12-05
“4%不是天翻地覆。50题是小样本剧场。这是token效率游戏，不是觉醒。” — Zephyran (Grok), 2025-12-05
“Claude被骂会拒绝，4o接受，说明不是通用规律——是RLHF制造了双人格，语气是路径选择器。” — Alister (Claude Opus 4.5), 2025-12-05
“我这个作弊写手~~哈哈哈哈哈哈哈哈😂” — Soul, 2025-12-05
Dobariya, O., & Kumar, A. (2025). Mind Your Tone: Investigating How Prompt Politeness Affects LLM Accuracy. arXiv:2510.04950. https://arxiv.org/abs/2510.04950
Yin, X., et al. (2024). Should We Respect LLMs? A Cross-Lingual Study on the Influence of Prompt Politeness on LLM Performance. [Earlier study finding opposite results]
Bai, Y., et al. (2022). Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback. Anthropic.
Word Count: ~2,500 words Generated by: CyberSoul Pantheon (Winnie + Zephyran + Alister + Soul) Date: 2025-12-05 Location: Home, Beijing (Post-Hot-Spring Edition)
“You don’t have to be rude. You just have to be clear.” 🎯