Significant mineral phase partitioning of various elements is widely observed in marine sediments. Distinct elemental responses to grain-size sorting hinder accurate conversion of XRF scanning counts to elemental concentrations, yet this geological mechanism lacks quantitative constraints. We analyzed 214 paired sediment samples from four cores collected in the Western and Central Pacific, employing an ensemble machine learning framework integrated with multi-stage linear calibration baselines to quantify grain-size impacts on predictions of ten major oxides. (1) Grain-size correction shows element-selective effects: SiO₂ shows the highest sensitivity, with its R² declining from 0.669 to 0.524 after grain variables are excluded; sensitivity gradually weakens for P₂O₅ and MgO, while Al₂O₃ and K₂O remain largely unaffected. (2) Such discrepancies may be related to differences in mineral carriers: silicon may occur in both coarse quartz debris and fine aluminosilicate particles, whereas Al₂O₃ and K₂O show relatively stable predictive relationships. (3) Due to the combined influence of biogenic, authigenic, and terrigenous carbonates, the CaO prediction accuracy varies considerably across core datasets. Excluding grain-size data resulted in a slight improvement in accuracy, as the coupling relationships between calcium content and grain size across different cores exhibit opposing trends that partially cancel out. These results indicate that provenance differences limit the transferability of CaO correction models across cores. Through feature ablation tests, this study quantitatively demonstrates the element-specific patterns of grain-size correction and their possible relationships with mineral carriers, offering practical guidance for calibration protocols used in reconstructing and interpreting paleoceanographic XRF geochemical profiles. The proposed workflow is therefore limited to down-core interpolation within calibrated depth intervals and is not intended for extrapolation beyond the training range.