Reported on 110 benchmarks across 11 tasks
Note: results are matched by exact model name. Different papers may use the same name for different model variants.