AA-PersonalityBench 统计：GLM-5.3 平均每条回复用 7.4 个破折号居首
Artificial Analysis 数了各模型写破折号的频率，GLM-5.3 一条回复平均 7.4 个，Claude Opus 5.5 几乎为零，看你的模型像谁就靠这个。
Artificial Analysis 在 AA-PersonalityBench 中统计了各模型回复中破折号（em dash）的使用频率。GLM-5.3 平均每条回复使用 7.4 个破折号，GLM-5.3-Flash 为 6.9 个，Kimi K3 为 6.7 个。Claude 系列几乎不用：Opus 5.5 平均每 40 条回复不到 1 个（0.02），Sonnet 5.5 与 Haiku 5.5 为 0.00。Claude Fable 5.1 平均 6.4 个，是 Claude 系列中的例外。GPT-3.5 Turbo 为 0.00。
In AA-PersonalityBench, GLM-5.3 uses the most em dashes of any model measured, averaging 7.4 per response, while Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5 use almost none
We count the em dashes in every response each model gives across the benchmark. Claude Fable 5.1 averages 6.4 em dashes per response, while Claude Opus 5.5 averages fewer than one every 40 responses.
Average em dashes per response:
➤ Highest: GLM-5.3 (7.4), GLM-5.3-Flash (6.9), Muse Wonder (6.9), Kimi K3 (6.7) and Muse Spark 1.3 (6.5)
➤ Lowest: GPT-3.5 Turbo (0.00), Claude Haiku 5.5 (0.00), Claude Sonnet 5.5 (0.00 at medium, 0.01 at max), Llama 4 Maverick (0.01) and Claude Opus 5.5 (0.02)
