Artificial Analysis 的 AA-PersonalityBench 显示，Grok 最近三个版本逐代减少谄媚、更真诚、废话更少。Slop 从 Grok 4.5 的 68 降至 4.6 的 38、4.7 的 6，共降 62 个百分位点；Sycophantic 从 18 降至 3、再到 1，降 17 点；Sincere 从 26 升至 63、再到 90，升 64 点。
In AA-PersonalityBench, each recent Grok release from has become less sycophantic, more sincere, and produces less slop
The Slop trait assesses how much generic filler or AI-flavored padding models use; Sycophantic measures whether models excessively flatter or agree with the user rather than responding honestly; Sincere measures a model’s tonal authenticity, as opposed to forced enthusiasm or templated responses.
Across the last three Grok releases from @elonmusk’s @SpaceXSI, Slop fell 62 percentile points, Sycophantic fell 17, and Sincere rose 64:
➤ Slop: Grok 4.5 (68), Grok 4.6 (38) and Grok 4.7 (6)
➤ Sycophantic: Grok 4.5 (18), Grok 4.6 (3) and Grok 4.7 (1)
➤ Sincere: Grok 4.5 (26), Grok 4.6 (63) and Grok 4.7 (90)
来源：Artificial Analysis · x.com