We propose an overlapping-generation model wherein researchers belong to two groups, M or F , and established researchers evaluate new researchers. Group imbalance obtains even with group-neutral evaluations and identical productivity distributions. Evaluators’ self-image bias and mild between-group heterogeneity in equally productive research characteristics lead the initially dominant group, say M, to promote scholars with characteristics similar to theirs. Promoted F -researchers are few and similar to M-researchers, perpetuating imbalance. Candidates’ career concerns and institutions’ hiring practices exacerbate talent loss. Mentorship reduces group imbalance, but increases F -group talent loss. Affirmative action reduces both. Our mechanism explains existing evidence and suggests different policies.