<p>Doctoral training relies heavily on supervisor–student collaboration as a channel of academic socialization, yet its longer-run implications for productivity and independence remain debated. Using integrated data from Academic Family Tree and OpenAlex, we construct a cross-disciplinary, cross-national dataset spanning 1960–2024 with more than 1.8 million supervisor–student dyads. We examine how the scale, quality, authorship structure, and continuity of doctoral collaboration are associated with post-PhD research productivity, research impact, and multiple measures of academic independence. Supervisor–student co-authorship intensity shows a robust inverted U-shaped association across outcomes, as productivity, impact, and independence rise at low to moderate collaboration levels but display diminishing returns at higher intensity. The model-implied turning points are not universal numeric targets because publication and authorship regimes vary substantially across disciplines. A higher share of student first authorship during the PhD is strongly associated with earlier and stronger transitions into solo-authored publishing, although this association is attenuated in larger teams. Collaboration quality and stability, captured by top-tier co-authorship and consecutive years of co-authorship, are positively associated with later outcomes, whereas heavier dependence on supervisor-involved publishing is consistently detrimental for independence. These patterns vary across disciplines, institutional tiers, and national systems. Based on observational evidence supported by extensive controls and robustness checks, the findings suggest that doctoral training should strike a balance between collaboration and independence, reinforce transparent authorship norms, and ensure effective management of team structures.</p>