Abstract
Computerized adaptive testing (CAT) systems face major challenges at the beginning of test administration, when limited response data produces unstable ability estimates and poor item selection. This cold-start problem reduces measurement precision and testing efficiency, especially for students whose abilities diverge from population norms. This study introduces a hybrid ability-estimation model that dynamically integrates neural network predictions with classical item response theory (IRT) estimation throughout the adaptive testing process. The neural component uses auxiliary student information-including demographics, prior performance, and early response patterns-to generate accurate initial ability estimates, while the IRT component preserves psychometric validity as response data accumulate. A dynamic fusion mechanism gradually shifts estimation weight from the neural model to the IRT model as more items are administered. Experimental validation on 2847 students across four subject domains shows that the hybrid approach reduces RMSE in ability estimation by 34.2% during the first five items compared with traditional CAT methods, while maintaining equivalent precision in later stages. The system also decreases the number of items required to reach target precision (SE < 0.3) by 28.7% on average, with the largest gains observed for students at ability extremes.