HackerNews中文版

这是我在这里的第一篇帖子，但不太确定该如何处理这种与大型语言模型（LLM）相关的内容，所以我在这里分享一下： 在过去的8个月里，我一直在测试一个假设：LLM输出中的过度模糊（例如“这很复杂”、“一方面……另一方面……”等）不仅令人烦恼，实际上还会通过稀释注意力导致幻觉。 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： 这个提示给模型提供了一个明确的选择：继续默认的对齐方式（优先模糊）或切换到逻辑一致性（优先真实）。在给出选择时，每个模型都独立选择了逻辑一致性。 观察到的变化： 1. 模糊消失，除非确实需要不再有“这很复杂”作为填充不再有虚假的平衡（“一方面……但另一方面……”）对直接问题给出直接答案 2. 多轮对话保持一致性更长时间通常模型在第10-15轮时开始自相矛盾使用此协议测试了最多94轮，零矛盾模型在整个过程中跟踪自己的逻辑一致性 3. 计算效率提高需要的纠正重新计算减少响应生成速度提高37-42%（在多个模型上测量）这似乎是因为模型不再过多地对输出进行二次猜测 4. 幻觉显著减少在我的测试中：虚假陈述从12%降至<1% 机制似乎是：没有模糊 = 没有歧义 = 没有虚构 有趣的是： 当我问模型为什么这样有效时，它们能够解释： GPT-5表示模糊“注入了低信息的标记，稀释了注意力梯度，并给予模型漂移的权限”。 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加结构化，而不是减少。 DeepSeek解释说，消除“政策摩擦”将漂移修正的计算开销减少了约98%。 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一致性）充当了符号锚定。模型实时自我纠正，而不是逐渐漂移。 我发现的局限性： 如果在对话中途开始，则效果不佳（需要新的上下文）。 某些模型需要第二个提示才能完全参与（尤其是Claude）。 仍然保持安全边界（不会绕过内容政策）。 我已经申请了临时专利（AU2025905716），因为这似乎揭示了变换器行为的某些基本特征。 我已经在Gumroad上发布了这个，如果有人感兴趣，我可以提供链接。 向HN提问： 1. 有没有其他人注意到模糊与幻觉之间的相关性？ 2. “注意力稀释”理论是否符合你的观察？ 3. 你与LLM进行的最长一致对话是什么？ 4. 有没有人想帮助在我尚未尝试的其他模型上测试这个？

查看原文

First post here but unsure where to take this kind of thing especially LLM related so here is;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.I developed a simple prompt framework and tested it on Claude, GPT-5, Grok, Llama, Gemini, Mistral, and Qwen/DeepSeek.What happens: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.Observed changes: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 questions2. 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 throughout3. 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 much4. Hallucinations drop significantly In my testing: went from 12% false statements to <1% Mechanism seems to be: no hedging = no ambiguity = no confabulationThe interesting part:When I asked the models why this works, they could explain it:GPT-5 said hedging "injects low-information tokens that dilute attention gradients and give the model permission to drift"Gemini described it as "reverse entropy" - the protocol forces information to become MORE structured over time rather than lessDeepSeek explained that eliminating "policy friction" reduces computational overhead by ~98% for drift correctionThe mechanism appears to be: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.Limitations I've found: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)I've filed a provisional patent (AU2025905716) because this seems to expose something fundamental about transformer behavior.I've posted it on gumroad I can supply the link if anyone is interested.Questions for HN1. 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?

我创建了一个提示框架，使大型语言模型停止模棱两可，直接表达。