Which Model Loves Saying “!” the Most? 哪个模型最喜欢说“!”
Exclamation marks per 1,000 characters
The Last Words Before “!”
Select a model or tap center · Tap any phrase to read context
Same Prompt, Two Voices
Zero exclamation marks in any input prompt
Six Other Marks, per 1,000 Characters
Hover or tap any bar to track a model across all six charts
About This Little Study
This little study observes punctuation habits across 9 verified direct-endpoint model services given 60 identical, unprompted English everyday chat prompts (10 each across casual chat, good news, emotional support, practical advice, everyday explainers, and social messages—30 adapted from our original Chinese everyday prompts plus 30 newly authored prompts; not an official MT-Bench score). Every prompt contains zero ! or !, evaluated in independent single-turn sessions at temperature = 1, yielding 540 / 540 complete English responses with zero missing pairs.
Only final assistant content is counted (excluding separate reasoning traces). Code blocks, inline code, URLs, and Markdown image syntax are stripped before scoring; natural quoted dialogue and drafted messages are kept. Two additional gateway routes (grok-4.7, muse-spark-1.3) are temporarily held out pending clean direct-endpoint runs after token-usage auditing revealed ~13k–14.5k injected upstream proxy system tokens forcing Chinese replies.
本实验测试 9 个直连无污染模型服务在相同的 60 道英文日常提示词下的标点习惯(日常闲聊、好消息、安慰支持、实用建议、知识解释、社交写作六类各 10 题;含 30 道原中文日常题的英文改写与 30 道新编英文题,不属于官方 MT-Bench 能力评测)。全部输入均不含 ! 或 !,独立单轮会话、无客户端 system prompt、temperature = 1 单次采样,共取得 540 / 540 条完整纯英文有效回答(零缺失)。
计数仅统计最终 content 正文(不含独立 reasoning 字段),预先剔除代码块、行内代码、URL 与部分 Markdown 控制符,保留自然语言引文与代写消息。另有两个网关转发渠道(grok-4.7、muse-spark-1.3)在审计中发现被上游 cursor: 反代注入约 1.3 万–1.45 万系统 token 及中文回复规则,现已暂时剔除,待更换干净直连 API 补跑后补回。