THE PERFECT BLEND: REDEFINING RLHF WITH MIXTURE OF JUDGES

Summary from the TLDR AI newsletter

Meta has a great new paper about using a mixture of judges models to properly perform multi task RLHF in post training. It dramatically improves final performance of the models on a number of benchmarks.