YouMind's 'writing taste' benchmark puts Opus 5.5 first
YouMind published an internal benchmark it calls "文感力" (writing taste) that scores how closely a model's choices in a specific context match those of an experienced author, rather than whether it merely meets a brief. The method takes recent well-received articles by known authors, removes a paragraph, turns the reasons the original works into scoring criteria and builds wrong answers that fail individual criteria, then tests models on both picking the original paragraph and writing the removed one from scattered and decoy material. YouMind says Opus 5.5 ranked first, Gemini 4 Argon was the most reliable reader and Gemini 3.8 Flash offered the best value, and that Opus 5.5 and GPT-6 Luna, previously near-tied in its tests, differed by 42 points; it also reports every model wrote longer than the original (median 1.6x) and that 58% of rewrites broke at least one of the author's trade-offs even on articles the model had read correctly. These are vendor-reported internal results without independent replication.
YouMind argues this metric captures something task-compliance benchmarks miss: models can follow instructions and recognise good writing yet still diverge from a mature author when they write themselves.
用「文感力」评测 agent 的写作水平 YouMind 内部原本有一套写作 benchmark，用来评测模型有没有覆盖写作要点、用对素材，结构、语气和篇幅是否符合要求。但随着模型不断变强，连 GPT-6 Luna 和 Opus 5.5 都打成了平手，但实际用起来，大家还是能感觉到不少差距。 我们观察用户在写作任务中使用的 prompt 发现，很多时候，用户传达的是一种感觉。比如：“让结尾带一点落空感。” 模型可能写成伤感。如果把评分标准写成“结尾要带有落空感”，再交给评委打分，伤感的结尾其实也能过关。 问题在于，把感觉写成评分标准，并没有让它变得更明确，只是把判断转交给了评委。最终测出来的，可能只是模型有多符合评委的口味。 那么怎样的写作算好？我们选择的角度是：好的写作从作者的意图出发，每一个选择都为这个意图服务。
写作很大程度上就是在做取舍：该交代哪个细节，哪句话不说出口，一个意思写到几分就停。而模型写作时是在替用户动笔，关键不是写出“一段合格的文字”，而是做出用户会做的选择。 所以我们衡量的是：在具体语境下，模型做出的取舍和成熟作者有多接近。这就像语文考试里的阅读理解：作者为什么选这个细节？为什么偏偏在这里结尾？我们把这种能力叫做「文感力」。 做法是：挑选知名作者近期发表、读者反响好的文章，挖掉其中一段，分析原文的写法为什么好，把每条理由写成一条标准，再针对每条标准改写原文，构造错误答案。这样只有原文能同时满足所有标准，答案是唯一的。 测试分阅读和写作两部分。阅读是从几个候选段落里选出原文。
#YouMind#Opus 5.5#Gemini 4 Argon#Gemini 3.8 Flash#GPT-6 Luna#DeepSeek V4.1 Flash#Haiku 5.5#Claude