The purpose of this study is to explore the impact of problem solving instruction on 126 middle school students’ learning of physical science in terms of their scientific knowledge, scientific concept dependent reasoning and problem solving ability. This study used a quasi-experiment with one factorial design of instructional approaches (problem solving and traditional hands-on learning). Sixty-one students participated in problem solving while sixty-five students participated in traditional hands-on learning. Results indicate that the problem solving group significantly outperformed the traditional hands-on learning group for both immediate and retaining effect, regardless of scientific knowledge, scientific concept dependent reasoning and scientific problem solving abilities. The regression results also indicated that the scientific concept dependent reasoning test is the best predictor for scientific problem solving ability, followed by scientific knowledge itself. Our study demonstrates that students’ scientific knowledge, reasoning and problem solving all are successfully improved after receiving six weeks scientific problem solving.