Abstract: Focusing on the Xianshuihe-Xiaojiang fault zone, this study integrates historical seismicity, instrumental records (1970-2025), and paleoseismic data to systematically evaluate magnitude distribution characteristics using maximum likelihood estimation and Monte Carlo simulations. The results reveal that while the fault zone as a whole approximates a standard Gutenberg-Richter (G-R) distribution (with b-values stabilizing between 0.95 and 1.05), individual fault segments exhibit significant non-linear deviations in the moderate-to-strong earthquake range, with frequencies exceeding those extrapolated from small earthquakes. This scale dependence correlates with slip rates; segments with high slip rates deviate earlier and more drastically from linearity, implying that high strain loading may suppress small-earthquake nucleation and promote stress release through larger ruptures. Furthermore, the overall statistical pattern is susceptible to "contamination" by highly active internal segments, leading to significant local specificity in composite models. Notably, certain segments are currently in an active window of accelerated stress release, where observed rates of moderate-to-strong earthquakes significantly deviate from G-R extrapolations and exceed long-term paleoseismic averages. This suggests that traditional G-R extrapolation methods based on steady-state assumptions may severely underestimate large earthquake probabilities during such periods. This study confirms that the G-R relationship is essentially a macroscopic statistical law of multi-fault composite systems, whereas individual segments are governed by local physical processes, emphasizing the necessity of segment-specific parameter calibration constrained by multi-source data for refining seismic zoning and hazard models.