Reported on 36 benchmarks across 6 tasks
Note: results are matched by exact model name. Different papers may use the same name for different model variants.