Comment by yorwba

Winning against their own baseline, not against the current best-performing model. Which apparently is S5 currently https://paperswithcode.com/sota/long-range-modeling-on-lra with 87.46 overall vs. 58.31 here.