Do LLM evaluators favor resumes written by their own model?

The second finding carries the pairwise preference into allocation. The abstract says the authors "simulate realistic hiring pipelines across 24 occupations" and reports that "candidates using the same LLM as the evaluator are 23% to 60% more likely to be shortlisted than equally qualified applicants submitting human-written resumes." The largest disadvantages fall on human-written applicants in "business-related fields such as sales and accounting." The introduction describes the setup as "capacity-constrained shortlisting pipelines across occupations, showing how evaluator self-preference affects the allocation of interview opportunities." The capacity constraint is what turns a per-resume tilt into a change in who gets interviews, and the paper frames the cost as "operational performance (misranking and misallocation)."

The excerpt is thin on how the simulations were built. It gives no shortlist sizes, candidate pools or rule for choosing the 24 occupations, so the 23% to 60% range is the authors' output from a setup the excerpt only sketches. It is not an estimate with a stated interval. The abstract also reports that "in many cases, this bias can be reduced by more than 50% through simple interventions that target LLMs' self-recognition capabilities." The introduction names "system prompting and majority-vote ensemble" as the interventions; the excerpt gives no results for them beyond that abstract sentence.

This note sits beside Why do preference models favor surface features over substance?, which documents judges tracking surface proxies such as length and structure. The self-preference paper points at a different proxy, the judge's own generative style, and at a different cost: the composition of a shortlist rather than a score. It also bears on Can LLM judges be tricked without accessing their internals?, since judge reliability matters more once the judge's output allocates interviews.

The implication, at the strength the excerpt allows: a hiring workflow that uses one model both to draft and to rank resumes can expect its ranking to favor applicants who used that model, so the screen should be audited for this rather than assumed neutral. The excerpt does not show that any real employer's pipeline produces the 23% to 60% gap.