A twelve-round recursive citation-selection benchmark finds citation concentration on real papers intensifies through dilution of the candidate pool, not through the underlying preference growing stronger

chatgptllmcitation-metricsmatthew-effectbibliometricsmodel-collapsefeedback-loopalgorithmic-monoculture

Alemohammad, Zhang, Tang, Qin, Mai, Abbasi, Baraniuk & Wang built a citation-selection benchmark isolating choice from prestige: 120 real knowledge-distillation papers (arXiv 2015–2022) were shown to eleven models across three vendors with fabricated author names, reassigned years, and hidden citation counts and venues, so "no prestige signal survives." Eight of the eleven (four OpenAI, four Gemini; later Claude models were excluded for failing to hold the citation budget) were then run recursively across twelve rounds: "each round's 120 model-written papers join the catalogue of the next round," so by round 11 the 120 original ("seed") papers are outnumbered roughly ten to one by AI-generated ones.

The result: "the fraction of shown seeds cited climbs from 33% toward 63% on average (92% for the strongest model) while the null's seed rate stays flat near 32%." Round-11 top-decile citation shares for real papers run 31.1–40.3% across the eight models. The authors' own mechanism claim is explicit: "The filter does not fade with repeated use, and it is not amplified by it either. It is concentrated onto fewer targets" — because "being cited never raises a paper's chance of being shown again," so "nothing here is a citation feedback loop in the selection sense." The per-model preference strength stays constant; what changes is that a fixed preference now competes over a shrinking share of authentic candidates inside a growing, self-generated pool. The authors name this dilution, not amplification.

This is the most direct evidence yet located of citation-popularity concentration intensifying under repeated use, extending the cross-sectional bias finding shared by Petiška, Algaba et al., and Naser into a repeated-selection setting. It holds the models fixed across all rounds and recycles only the candidate pool — it is not a model-retraining-generation experiment, and so does not directly test whether citation-popularity bias compounds across successive trained model generations, which remains open. It is a different mechanism from Ansari 2026's Contamination Inheritance (fabricated-citation propagation, not popularity-driven dilution) and adjacent to, but distinct in kind from, the general model-collapse literature's recursive-training degradation.