Liu, Chen, Liu, Lu & Murray 2026 find six LLMs over-cite popular and older papers, amplified for contrasting citations, affirming Algaba et al.'s popularity-bias finding

llm-citation-biasmatthew-effectcitation-metricscitation-intentmechanism-comparison

Liu, Chen, Liu, Lu & Murray (Network Science Institute, Northeastern University; University at Albany, SUNY), in a September 2026 preprint, tested six LLMs (GPT-5.1, Claude-3.5-Haiku, Gemini-2.0-Flash, DeepSeek-V3.2, Llama-4-Maverick, Qwen2.5-72B) against 1,746 ACL/EMNLP/NAACL 2025 papers using a masked-citation task — masking a real citation and having each model fill the same slot, so human and LLM citation choices can be compared at an identical position rather than across open-ended generation. Their finding: "LLMs over-cite popular and older papers, a tendency amplified for contrasting citations where human writing more often draws on recent, niche work." The effect is explicitly framed as a replication: "We affirm prior findings that LLMs disproportionately cite highly-cited and smaller-team work," citing Algaba et al.'s earlier GPT-4 popularity-bias finding directly. The paper's distinct contribution is intent-conditioning: the bias is not uniform across citation rhetoric — it sharpens specifically where human authors would reach for a recent, narrow counterexample and an LLM instead reaches for an established, well-known one.

The paper audits citation rhetoric and intent-conditioned popularity bias without citing Simkin & Roychowdhury or any copying-based account, and does not use the word "copying" anywhere in its text (checked directly against the full PDF). It extends the popularity-bias finding's cross-vendor and intent-conditioned reach without addressing why the underlying training-corpus citation distribution is skewed in the first place. See question-does-citation-copying-explain-training-corpus-skew-behind-llm-citation-bias.