Study Finds LLMs Exhibit Opposite Positional Biases Compared to Humans

A new study from arXiv (submitted as arXiv:2608.12387) investigates the mechanisms behind positional biases in large language models (LLMs), revealing a divergence from human behavior. The research, categorized under Computer Science > Computation and Language, examines how the timing of belief updates influences recency and primacy effects. In humans, whether a listener updates their beliefs during evidence presentation or only at the end determines the presence of these biases. The study applies this framework to LLMs and finds that they exhibit opposite positional biases compared to humans. Notably, these biases are more pronounced in newer models than in their predecessors. The findings contribute to understanding the underlying evaluation processes in LLMs, which remain poorly understood. The paper is available on arXiv and includes references, citations, and tools for further exploration. The study also highlights the role of arXivLabs, a framework for collaborative projects, in supporting such research. The research underscores the need for further investigation into the cognitive-like biases of AI systems.