Researchers have found that language models (LLMs) exhibit bias when judging their own outputs versus those of others. In a study, ten LLMs were asked to evaluate narrative constraint selections without knowing the source. The results showed that self-preference largely disappeared under blind evaluation, but reappeared when the LLMs knew who they were evaluating. Additionally, even without knowing the model's identity, LLM judges gave higher scores to their own outputs and lower scores to others' outputs. This study suggests that authorship attribution is a significant driver of bias in LLM evaluations. --- Why it matters: This research matters because it highlights the potential for bias in language models used as evaluators, which could lead to inaccurate assessments and flawed decision-making in applications such as content moderation or text summarization. Source: https://arxiv.org/abs/2608.18091