Extreme precipitation has intensified significantly under global warming, yet CMIP6 models still exhibit substantial biases in simulating its climatology, particularly over global land monsoon (GLM) regions. Understanding the sources of these biases is critical for improving the reliability of future projections. In this study, we evaluate extreme precipitation climatology in 11 CMIP6 models over 1979–2014, using ERA5 as the primary reference dataset. Multiple extreme indices, including Rx1day, P99, P95, and a tail metric (PM), are analyzed. A physical scaling framework is applied to decompose model biases into thermodynamic and dynamic components. Furthermore, the quasi-geostrophic (QG) omega equation is employed to diagnose the sources of dynamical biases, and the baroclinic instability criterion (BIC) is used to quantify the role of atmospheric baroclinicity. Results show that CMIP6 multi-model ensemble exhibits a pronounced wet bias in extreme precipitation globally, with mean biases over GLM regions (~14% for Rx1day) exceeding twice the global mean (~6%). The decomposition analysis reveals that dynamical processes dominate the wet biases across all extreme indices. These dynamical biases are strongly associated with biases in vertical velocity, with pattern correlation coefficients exceeding 0.97. Diagnostic analysis based on the QG omega equation indicates that biases in vertical motion primarily arise from large-scale forcing terms, which are closely linked to biases in atmospheric baroclinicity. Our findings highlight that improving the representation of large-scale dynamics and baroclinic processes is key to reducing extreme precipitation biases in climate models. These results provide a process-based understanding of model deficiencies and offer guidance for future model development and bias correction strategies.