Approach to forming a set of test tasks for personnel assessment using chatbots

Abstract

The article proposes an approach to forming a set of test tasks for personnel assessment using the OpenAI ChatGPT-5.2 Plus chatbot. A mechanism for selecting questions from a task bank has been developed, ensuring the simultaneous achievement of three key requirements: the variability of test sessions, the absence of repetitions, and the preservation of the predefined assessment structure. It is shown that the use of a probability distribution for selecting difficulty intervals makes it possible to control the sequence of tasks, guiding the generator toward a higher share of medium-level questions while reducing the frequency of extreme values. The conducted simulation confirmed the statistical consistency of the generated samples, which indicates the stability and reproducibility of the approach. Two alternative methods for generating task numbers are considered: the inverse transform method and the acceptance–rejection method. Experimental results show that both approaches produce statistically homogeneous samples and can be effectively used in practical testing systems. The study also justifies integrating the task generation module into the chatbot structure. In this case, the chatbot manages the test session by defining the allowed set of questions, preventing repetition, supporting dialogue interaction with the user, recording answers, and adapting the further testing process. The average share of correct answers was 0.71, indicating a balance between task difficulty and the testing trajectory. For profiles with a higher level of competence, the average task difficulty increased to 1.17–1.21, while in less prepared scenarios it decreased to 0.79–0.81. The average duration of a test session was 418.8 seconds. Final scores ranged from 63.7 to 88.4. Simulation experiments showed that full uniqueness of tasks is ensured within each session. Таbl.: 3. Fig.: 1. Refs.: 11 titles.