LARGE LANGUAGE MODELS ARE NOT FAIR EVALUATORS

Summary from the TLDR AI newsletter

It is challenging to compare the performance of multiple models to one another. Traditional automated measures aren't robust and human evaluation isn't scalable. Recently people have been turning to GPT4 as a proxy for human preference. After all, it was trained with RLHF. However, doing this naively results in miscalibration. This work proposes a few simple fixes that aligns GPT4 evaluations more closely with those of humans.