This study investigates computational thinking (CT) proficiency in programming-based learning environments using a unified analytical framework that combines statistical analysis and predictive modeling. CT proficiency is examined across five pedagogically grounded dimensions—abstraction, decomposition, algorithmic thinking, debugging, and pattern recognition—derived from rubric-based assessment of student work. First, descriptive and inferential statistical analyses are conducted to examine overall CT performance and gender-related patterns, providing correlational insight into group-level differences. Building on this analysis, a transformer-based model (BERT) is employed to predict continuous CT proficiency from students’ code comments and reflective journals, enabling semantic modeling of higher-order reasoning expressed in natural language. The predictive performance of BERT is benchmarked against traditional machine learning and lightweight deep learning baselines. Results show that transformer-based semantic modeling improves predictive accuracy while maintaining interpretability through post hoc explanation methods. Explainable AI techniques are used to identify linguistic and behavioral indicators associated with CT proficiency, and gender-related interpretations are derived through subsequent comparative analysis rather than direct prediction. Overall, the study positions deep learning as a complementary tool to statistical analysis for understanding and predicting CT proficiency.

In the era of Artificial Intelligence (AI), where technology is transforming industries, education stands at a pivotal juncture. With an increasing emphasis on critical thinking and problem-solving, there is a growing need for innovative tools that can foster these essential skills among students. Traditional education methods need help making personalized scalable and interesting experiences for students at this task type which this research aims to solve. The research uses AI and deep learning tools to build an effective framework that enables better riddle solving for students by proposing state of the art deep features including sentence embeddings and ULMfit to be applied as input to deep learning models. In contrast, this study examines different traditional machine learning and deep learning models including ensemble learning models, used as baseline models for comparing the performance of the proposed transformer architectures based on RoBERTa-Large to determine which approach works best, achieving highest accuracy of 96% to effectively handle riddle complexity. The research studies used text data patterns using TF-IDF, Count Vectorization, and word embedding techniques which apply in the form of Roberta. Our research findings help educators, technology experts and scientific teams design educational tools with an easy-to-deploy AI solution.