Abstract: To reveal the multi-dimensional complex correlations among the physical and mechanical parameters of cohesive soils, a structural equation model(SEM)was established based on the engineering investigation data of 1051 valid cohesive soil samples at a depth of 2~20m in the central urban area of Changchun City. The single prediction-oriented paradigm of previous studies was broken through and the deficiency of insufficient analysis on internal action mechanisms was remedied. Bootstrap resampling and sensitivity analysis with phantom variables were employed to quantitatively assess parameter estimation stability and the impact of confounding variables. Generalized Additive Model(GAM)were integrated to identify non-linear characteristics of model paths. The results demonstrate that the constructed SEM exhibits satisfactory goodness-of-fit(CFI=0.991, TLI=0.954). The liquidity index(I_L)acts as the core physical indicator governing the compression coefficient(a), cohesion(c), and tangent of internal friction angle(tanφ) of cohesive soils; it is significantly positively correlated with the compression coefficient a, and significantly negatively correlated with cohesion c and tanφ, with stable path coefficients and an explicit physical mechanism. Most paths are obviously affected by omitted confounding variables, suggesting that indicators reflecting"mineral composition" and"microstructure" should be incorporated in subsequent research. Distinct nonlinear characteristics are observed in the c-I_L and tanφ-w paths, which results in poor prediction performance of the model for cohesion c and tangent of internal friction angle (tanφ) of new samples. In the future, the integration of SEM with machine learning techniques can be considered to develop a hybrid model with both theoretical interpretability and nonlinear fitting ability. The research results provide a theoretical basis for the optimization of engineering parameters and the improvement of mechanical parameter prediction models for cohesive soils, enrich the research methods for multi-parameter correlations of cohesive soils, and can serve as a reference for geotechnical engineering investigation and design in similar areas.