This project investigates how different mixed-reality (MR) visualization methods—in-situ and ex-situ—affect learning performance in arthroscopic training across varying task complexities and training repetitions. While in-situ overlays anatomical structures directly onto the physical knee model, ex-situ presents the virtual model spatially offset, requiring learners to integrate separated visual cues. Thirty medical trainees performed multi-level arthroscopy targeting tasks (low/medium/high visibility) under MR guidance or traditional apprenticeship teaching. Objective metrics (accuracy, completion time, spatial deviation, errors, landmark-judgment time) and subjective metrics (NASA-TLX, preference) were collected. By examining the interaction between visualization method, task complexity, and training frequency, the study aims to identify optimal applicability zones for each MR method—where in-situ supports early learning and feature-point recognition, and where ex-situ enhances visuospatial reasoning and psychomotor coordination. The findings aim to inform adaptive MR-based curricula for arthroscopic education.