We did a similar (less rigorous) evaluation internally last year, and one of the big things we identified was "purple prose":

  In literary criticism, purple prose is overly ornate prose text that may disrupt a narrative flow by drawing undesirable attention to its own extravagant style of writing, thereby diminishing the appreciation of the prose overall.[1] Purple prose is characterized by the excessive use of adjectives, adverbs, and metaphors. When it is limited to certain passages, they may be termed purple patches or purple passages, standing out from the rest of the work. (Wikipedia)

The Slop Score they have gets at this which is good, but I wonder how completely it captures it.

Also curious about how realistic this benchmark is against real "creative writing" tools. The more the writing is left up to the LLM the better the benchmark likely reflects real performance. For tools that have a human in the loop or a more structured approach it's hard to know how well the benchmarks match real output, beyond just knowing that better models will do better, e.g Claude 3.7 would beat Llama 2.