Learning in the 21st century has shifted to using technology and various approaches that facilitate students' ability to solve problems independently. One substitute method for assessing students' scientific problem-solving skills and self-efficacy was an AR-based STEM approach. This study aimed to determine the effect and correlation between self-efficacy and students' problem-solving skills in learning science using AR. A quasi-experiment with a pretest-posttest non-equal control group design model was employed. The number of samples in this study consisted of 50 students spread across the experimental (VA) and control (VB) classes. The results of the value of the problem-solving ability in the control class were 58 on the pretest and 77 on the post-test, while the experimental class yielded 61 on the pretest and 82  on the post-test. The average self-efficacy of the control class was 53 on the pretest and 60 on the post-test, while the experimental class yielded 56 on the pretest and 79 on the post-test. The study's results showed that using the AR-based STEM approach affected self-efficacy and problem-solving ability, with a percentage of 21%. The T-test score obtained by tcount>ttable (3.630 > 2.010) indicates a significant correlation between self-efficacy and students' problem-solving ability. Students with higher levels of self-efficacy were better at solving problems, as shown by the regression value for every 1% increase in the self-efficacy variable, contributing 64.13 to the problem-solving ability. Thus, there was a significant correlation between self-efficacy and problem-solving ability in science learning using the AR-based STEM approach.