They probably mean that if you test it on common NLP benchmarks it performs better.