Study finds humans prefer fewer, more diverse citations in LLM outputs while models show inconsistent preferences

A pre-print study published on arXiv examines how citations influence human and LLM preferences in the context of scientific question answering. The research, titled *On the Role of Citations in Preference Data*, analyzes pairwise judgments from human evaluators and four open-source LLMs to determine how citations affect preferences.

The authors report that humans tend to prefer outputs with fewer but more diverse citations, suggesting a nuanced approach to attribution that balances comprehensiveness with conciseness. In contrast, LLMs exhibit citation-related preferences that vary by model and dataset, despite lacking access to the cited sources. This inconsistency highlights a gap between human judgment and model behavior in evaluating attributed outputs.

The study leverages mixed effects models to investigate the influence of citations on pairwise judgments, providing a quantitative framework for assessing citation impact. The authors discuss the implications of these findings for preference data collection, particularly in the context of reward modeling and modern LLM post-training methodologies.

The paper is authored by Yu Hou, Hal Daumé III, Rachel Rudinger, and William Walden, and was submitted to arXiv on July 6, 2026. It contributes to ongoing discussions about attribution mechanisms in AI systems, especially as they relate to user trust and the mitigation of hallucinations in outputs.