Fine-tuning a dense retriever on LLM-generated text induces a measurable pro-LLM-content bias (Xion & Nejdl 2026)

llmfeedback-loopbias-amplificationinformation-retrievalarxiv

William Xion and Wolfgang Nejdl (L3S Research Center, Hannover) fine-tuned a dense retrieval model on a corpus containing LLM-generated text and found the retriever's ranking preference shifted measurably toward LLM-generated content over comparably relevant human-authored content: "Fine-tuning on LLM-generated corpora induces a pronounced pro-LLM bias." The design compares a single before/after pair of training checkpoints — one fine-tuning step — rather than iterated multi-generation retraining.

This surfaced during a search for prior work applying Wang et al.'s iterated multi-generation retraining design to citation-popularity bias specifically (see question-does-citation-popularity-bias-compound-across-llm-training-generations); it does not answer that question, since it concerns retrieval-ranking preference rather than citation selection, and a single training step rather than a compounding chain. It is, however, the closest structural analog located: direct empirical evidence that a training-induced feedback bias of the same general shape — a system's own outputs, once back in a training corpus, tilting the system's future behavior toward those outputs — is real in an adjacent system (retrieval ranking), even though it has not yet been demonstrated for citation-count popularity or run across successive generations.