When a new model climbs a creativity leaderboard (e.g., LM Arena’s creative-writing ranking), what has actually been measured - and what has improved?
Our new paper examines this and points to one shared target: long-tail performance.
15 benchmarks. 26 open-weight models. 🧵
