Forest soil organic carbon (SOC) stability is influenced by its relative composition of particulate organic carbon (POC) and mineral-associated organic carbon (MAOC) fractions. However, conventional SOC fractionation methods are labor-intensive and restrict large-scale monitoring of SOC dynamics. Visible–near infrared (Vis–NIR) spectroscopy offers a rapid alternative, yet its applicability for predicting SOC fractions in forest soils under field conditions remains poorly understood. This study developed an integrated framework to evaluate the feasibility of in-situ Vis–NIR spectroscopy for predicting SOC fractions by comparing four in-situ application workflows, including direct laboratory-to-field transfer, EPO-assisted transfer, direct in-situ modeling, and EPO-assisted in-situ modeling. Direct transfer of laboratory models to in-situ spectra resulted in substantial performance degradation due to moisture-driven spectral domain shifts (POC: R² = 0.80; MAOC: R² = 0.59). In contrast, direct in-situ modeling. The highest accuracy for POC was achieved using EPO-corrected in-situ spectra (R² = 0.90), whereas MAOC prediction performed best using uncorrected in-situ spectra (R² = 0.71). Independent cross-year validation further demonstrated that environmental variability, particularly soil moisture, constrained model robustness. The analysis of the fitted models revealed distinct spectral mechanisms controlling SOC fraction predictions, linking POC to shortwave infrared organo–clay absorption features (~2200 nm) and MAOC to visible wavelengths associated with iron oxides. These findings highlight the conditional feasibility of in-situ Vis–NIR spectroscopy for forest SOC fraction prediction and guide field-based soil carbon monitoring.