cfDNA fragmentomic features have been used in cancer detection models; however, the generalizability of the models needs to be tested. We proposed a type of cfDNA fragmentomic feature named chromosomal arm-level fragment size distribution (ARM-FSD), evaluated and compared its performance and generalizability for lung cancer and pan-cancer detection with existing cfDNA fragmentomic features (as reference) by using cohorts from different institutions. The ARM-FSD lung cancer model outperformed the reference model by ~10% when being tested by two external cohorts (AUC: 0.97 vs. 0.86; 0.87 vs. 0.76). For pan-cancer detection, the performance of the ARM-FSD based model is consistently higher than the reference (AUC: 0.88 vs. 0.75, 0.98 vs. 0.63) in a pan-cancer and a lung cancer external validation cohort, indicating that ARM-FSD model produces stable performance across multiple cohorts. Our study reveals ARM-FSD based models have a higher generalizability, and highlights the necessity of cross-study validation for predictive model development.