Meta, through its FAIR division, has introduced a new AI approach called 'Self-Taught Evaluators' that aims to enhance the evaluation of language models without the need for human annotations. This method utilizes synthetic training data and an iterative self-improvement scheme to train models. The Self-Taught Evaluators have demonstrated superior performance compared to commonly used language model judges like GPT-4, and they match the performance of top reward models trained with labeled examples. The approach involves generating contrasting outputs to train a language model as a judge, producing reasoning traces and final judgments. The model has shown significant improvements, notably boosting the Llama 3-70B model's performance on RewardBench to scores of 88.3 and 88.7 with majority vote, outperforming larger models and human labels.