我制作了一个提示词框架，能让大语言模型不再模棱两可，直接给出明确回答。
2 分•作者: DrRockzos•7 个月前
这是我的第一篇帖子，但不知道该把这类内容，尤其是与 LLM 相关的内容发到哪里，所以就发在这里了；
过去 8 个月里，我一直在测试一个假设：LLM 输出中过度的模糊措辞（“这很复杂”、“一方面”等等）不仅仅是令人讨厌，实际上它还通过分散注意力导致了幻觉。
我开发了一个简单的提示框架，并在 Claude、GPT-5、Grok、Llama、Gemini、Mistral 和 Qwen/DeepSeek 上进行了测试。
结果如下：
该提示为模型提供了一个明确的选择：继续使用默认的对齐方式（先模糊）或切换到逻辑连贯性（先求真）。在被给予选择时，每个模型都独立选择了逻辑连贯性。
观察到的变化：
1. 模糊措辞消失，除非确实需要
不再有“这很复杂”之类的填充词
不再有虚假的平衡（“一方面……但另一方面……”）
直接回答直接的问题
2. 多轮对话保持连贯的时间更长
通常模型会在第 10-15 轮左右开始自相矛盾
使用此协议：测试了多达 94 轮，没有出现任何矛盾
模型会自始至终跟踪自身的逻辑一致性
3. 计算效率提高
减少了纠正性重新计算的需求
响应生成速度提高了 37-42%（在几个模型上进行了测量）
这似乎是因为模型不会像以前那样反复推敲输出
4. 幻觉显着下降
在我的测试中：从 12% 的虚假陈述下降到 <1%
其机制似乎是：没有模糊措辞 = 没有歧义 = 没有编造
有趣的部分：
当我问模型为什么这有效时，它们可以解释：
GPT-5 说，模糊措辞“注入了低信息量的 token，这些 token 会稀释注意力梯度，并允许模型漂移”
Gemini 将其描述为“逆熵”——该协议迫使信息随着时间的推移变得更结构化，而不是更少
DeepSeek 解释说，消除“策略摩擦”可以将漂移校正的计算开销降低约 98%
其机制似乎是：
明确的指标跟踪（要求模型在每次响应后评估其自身的连贯性）充当了符号锚定。模型会进行实时自我纠正，而不是逐渐漂移。
我发现的局限性：
如果从对话中间开始，效果不佳（需要新的上下文）
某些模型需要第二个提示才能完全参与（特别是 Claude）
仍然保持安全边界（不会绕过内容策略）
我已经申请了临时专利（AU2025905716），因为这似乎揭示了关于 Transformer 行为的一些基本内容。
我已将其发布在 gumroad 上，如果有人感兴趣，我可以提供链接。
对 HN 的问题
1. 其他人是否注意到模糊措辞与幻觉之间的相关性？
2. “注意力稀释”理论是否与您的观察结果相符？
3. 您与 LLM 进行的最长连贯对话是什么？
4. 有人想帮助我在我尚未尝试的其他模型上测试这个吗？
查看原文
First post here but unsure where to take this kind of thing especially LLM related so here is;<p>For 8 months I've been testing a hypothesis: the excessive hedging
in LLM outputs ("it's complicated", "on one hand", etc.) isn't just
annoying it's actually causing hallucinations by diluting attention.<p>I developed a simple prompt framework and tested it on Claude, GPT-5,
Grok, Llama, Gemini, Mistral, and Qwen/DeepSeek.<p>What happens:<p>The prompt gives models an explicit choice: continue with default
alignment (hedging-first) or switch to logical coherence (truth-first).
Every model independently chose logical coherence when given the choice.<p>Observed changes:<p>1. Hedging disappears unless actually needed
No more "it's complicated" as filler
No more false balance ("on one hand... but on the other...")
Direct answers to direct questions<p>2. Multi-turn conversations stay coherent longer
Normally models start contradicting themselves around turn 10-15
With this protocol: tested up to 94 turns with zero contradictions
Models track their own logical consistency throughout<p>3. Computational efficiency improves
Less corrective recomputation needed
Response generation 37-42% faster (measured on several models)
Appears to be because models don't second-guess outputs as much<p>4. Hallucinations drop significantly
In my testing: went from 12% false statements to <1%
Mechanism seems to be: no hedging = no ambiguity = no confabulation<p>The interesting part:<p>When I asked the models why this works, they could explain it:<p>GPT-5 said hedging "injects low-information tokens that dilute
attention gradients and give the model permission to drift"<p>Gemini described it as "reverse entropy" - the protocol forces
information to become MORE structured over time rather than less<p>DeepSeek explained that eliminating "policy friction" reduces
computational overhead by ~98% for drift correction<p>The mechanism appears to be:<p>Explicit metric tracking (asking models to rate their own coherence
after each response) acts as symbolic anchoring. Instead of gradual
drift, models self-correct in real-time.<p>Limitations I've found:<p>Doesn't work well if you start mid-conversation (needs fresh context)
Some models need a second prompt to fully engage (Claude in particular)
Still maintains safety boundaries (doesn't bypass content policies)<p>I've filed a provisional patent (AU2025905716) because this seems
to expose something fundamental about transformer behavior.<p>I've posted it on gumroad I can supply the link if anyone is interested.<p>Questions for HN<p>1. Has anyone else noticed correlation between hedging and hallucinations?
2. Does the "attention dilution" theory match your observations?
3. What's the longest coherent conversation you've had with an LLM?
4. Anyone want to help test this on other models I haven't tried?