With the promotion of STEM education and active practice, experiential learning has become a crucial instructional design strategy. Experiential learning emphasizes a student-centered learning model, encouraging students to explore unknown fields through individual and team collaborative efforts. Through practical activities, it promotes active exploration and deep understanding, thereby stimulating students' creativity and independent thinking abilities. However, there are still some challenges confronting the learners. However, due to the limitations of short-term memory and the ability to organize key information, students often struggle to recall all past dialogues during reflection and observation processes. This hinders effective discussions and communication, impacting the application of collaborative strategies. To address this challenge in STEM collaborative discussions, this study explores the application of generative AI technology in STEM education. By integrating experiential learning with collaborative strategies, the goal is to enhance students' learning outcomes and reflective thinking. This research introduces a GPT-based large language model tool, the GPT-Assisted Summarization Aid (GASA), which is applied in collaborative learning activities within the framework of experiential learning. GASA supports learners in their reflective observation and conceptualization processes by providing feedback and assistance tailored to different phases of the experiential learning curriculum. The experimental results indicate significant differences in learning outcomes and reflective thinking between the experimental group using GASA and the control group. Furthermore, students in the experimental group were more effective in sharing information, establishing a common understanding, and solving problems during the collaborative learning process. This study provides empirical evidence for teaching strategies in STEM education and reveals the potential of generative AI technology in educational technology. Future research should explore larger sample sizes and long-term learning impacts to further validate the broad applicability of GASA in different educational contexts.