Large Language Model Evaluators Exhibit Self-Recognition and Self-Preference Biases

Large language model (LLM) evaluators exhibit non-trivial self-recognition capabilities and a tendency to favor their own generated outputs, a phenomenon known as self-preference bias.