Not sure I understand what you mean? RAdam is included in the figures, and seems to consistently underperform AdaBelief. Are there other benchmarks/papers that you're looking at? If so, shoot me a link. Curious what you're seeing that I'm not.