With population aging and increasing awareness of oral health, dental caries and tooth defects remain conditions with a high clinical burden. Tooth preparation is a critical procedure in restorative dentistry and requires high geometric accuracy and operational consistency. Robot-assisted tooth preparation can reduce the influence of clinician fatigue and hand tremor, thereby improving procedural stability and standardization. However, cantilevered end-effector instruments are prone to elastic deformation under preparation loads, and the resulting errors vary with posture and working conditions. Therefore, compensation based solely on geometric errors is difficult to achieve effectively. To address this problem, this study proposes a load-driven deformation error modeling and mirror-iteration trajectory optimization method for cantilevered end-effector instruments. First, a deformation error model of the cantilevered end-effector system is established based on cantilever beam theory. A finite element-assisted equivalent interface stiffness identification method is then introduced, in which equivalent interface parameters are identified from the loading response of local connection regions to improve the prediction accuracy of the model. Second, a load-deformation sensitivity-driven mirror-iteration feedforward compensation method is proposed. In the iterative update process, a sensitivity factor is introduced to correct the error induced by changes in preparation depth after compensation, and the compensated trajectory is optimized through constraint and convergence control. Quasi-static single-point loading experiments show that the cantilevered end-effector system exhibits good linear force-displacement response characteristics within the typical preparation load range of 1.0~2.0 N, supporting the assumptions of small-deformation linear elasticity and approximate three-directional decoupling. Ablation experiments on extracted teeth demonstrate that the proposed method reduces the MAE, RMSE, and P95 of the prepared surface to 0.044, 0.058, and 0.087 mm, respectively, with an average error suppression rate of 59.3%. These results verify the error suppression capability and trajectory optimization stability of the proposed method. Future work will integrate online force-position sensing and adaptive parameter updating strategies to further improve adaptability to individual tooth differences, bur wear, and variations in end-effector configuration.