The identification of predictors for immunotherapy is often hampered by the absence of control groups in many studies, making it difficult to distinguish between prognostic and predictive biomarkers. This study presents biomarker analyses from the phase 3 CONTINUUM trial (NCT03700476), the first to show that adding anti-PD-1 (aPD1) to chemoradiotherapy (CRT) improves event-free survival (EFS) in patients with locoregionally advanced nasopharyngeal carcinoma. A dynamic single-cell atlas was profiled using mass cytometry on peripheral blood mononuclear cell samples from 12 pairs of matched relapsing and non-relapsing patients in the aPD1-CRT arm. Using a supervised representation learning algorithm, we identified a Ki67+ proliferating regulatory T cells (Tregs) population expressing high levels of activated and immunosuppressive molecules including FOXP3, CD38, HLA-DR, CD39, and PD-1, whose abundance correlated with treatment outcome. The frequency of these Ki67+ Tregs was significantly higher at baseline and increased during treatment in patients who relapsed compared to non-relapsers. Further validation through flow cytometry (n = 120) confirmed the predictive value of this Treg subset. Multiplex immunohistochemistry (n = 249) demonstrated that Ki67+ Tregs in tumors could predict immunotherapy benefit, with aPD1 improving EFS only in patients with low baseline levels of Ki67+ Tregs. These findings were further validated in the multicenter phase 3 DIPPER trial (n = 262, NCT03427827) and the phase 3 OAK trial of anti-PD-L1 immunotherapy in NSCLC, underscoring the predictive value of Ki67+ Treg frequency in identifying the beneficiaries of immunotherapy and potentially guiding personalized treatment strategies.