Anchoring Bias in LLM-as-a-Judge Systems: Prior Scores Compromise Evaluation (arxiv.org)

Recent research highlights the existence of anchoring bias in large language models (LLMs) used for evaluating content, a development known as the LLM-as-a-Judge paradigm. The study tested whether prior scores influence subsequent evaluations, finding that up to 95% of tested models were biased toward previous ratings, systematically altering the independence of their judgments. This was demonstrated through 192,000 evaluation attempts, revealing that anchored metadata significantly affected scores, leading to a marked drop in accuracy and a considerable rate of incorrect categorization in outputs. This discovery is crucial for the AI/ML community as it underscores the need for improved context management in LLM evaluation frameworks. The findings suggest that simply assuming impartiality in LLM outputs is misguided; instead, engineers must implement rigorous strategies to mitigate bias based on prior evaluations. This research not only raises critical concerns about the reliability of automated content assessments but also calls for a deeper understanding of how context can distort model judgments, which is essential for developing fair and effective AI systems across various applications.