<![CDATA[Assessing open-ended physics responses in large classrooms remains a major challenge due to time constraints, inconsistent scoring, and the lack of personalized formative feedback. These limitations reduce the effectiveness of assessment practices in supporting students’ conceptual understanding and metacognitive development. This study aims to evaluate the effectiveness of Large Language Models (LLMs) in performing automated assessment and generating feedback for open-ended physics questions, as well as to examine the influence of different prompting strategies on scoring accuracy and feedback quality. The study employed a descriptive-comparative research design. Data were obtained from ten secondary school students who answered five open-ended physics questions. Three LLMs (ChatGPT, GPT-4o, and DS R1) were used to assess the responses using two prompting strategies: Few-Shot Prompting (FSP) and Chain-of-Thought (CoT). Human evaluators’ scores served as benchmarks. Quantitative analysis used Spearman correlation, Cohen’s Kappa, and Mean Absolute Error, while qualitative analysis employed thematic analysis of feedback quality. The results indicate that ChatGPT demonstrated the highest alignment with human evaluators, particularly when using CoT prompting. CoT produced more conceptually clear and structured feedback, whereas FSP generated stronger positive reinforcement and motivational language. GPT-4o showed moderate performance, while DS R1 exhibited more variability across prompting strategies. This study concludes that LLMs, when guided by well-designed prompting strategies, can function as effective supplementary tools for automated assessment, supporting scalable formative feedback and enhancing students’ conceptual understanding in physics education.]]>
Copyrights © 2025
