ResearchBenchmarksOpen-ended Text Generation on WebText (test)Follow97Same Preference CountMIXCE92.1594.5759799.425May 26, 2023Evaluation ResultsMethodMethodLinksSame Preference CountMixCE Preference CountMLE Preference CountMIXCEBackbone=GPT-2 large,...Backbone=GPT-2 large, Decoding=unbiased sampling, Evaluation protocol=blind A/B testing2023.0597139—MLEBackbone=GPT-2 large,...Backbone=GPT-2 large, Decoding=unbiased sampling, Evaluation protocol=blind A/B testing2023.0597—79