Recently, Examstar released the first vertical large model in the field of test assessment, which has been tested in multiple projects before this release. The large model has the characteristics of verticalization, scenario-based and high precision, and integrates large language models, image large models, and multi-modal large models, which can help enterprises and institutions solve scenarios such as recruitment, certification, and talent ability evaluation to provide more intelligent and efficient solutions.
2024 is the 10th year of the establishment of Examstar, which has been focusing on the field of digital exams for 10 years, and at present, there are 550,000+ registered enterprises on the Examstar platform, with a total of 1.57 million ** exams delivered, serving nearly 100 million candidates. In 2021, Exam Star innovatively put forward the concept of "serious examination" in the industry, and realized ** exam and ** invigilation through pure online methods to ensure the fairness of ** exam; In 2023, Examstar has successively released a number of functions and solutions based on large models to further improve the efficiency and accuracy of the whole process of various types of exam scenarios.
The release of the industry's first vertical test assessment model is another milestone in the field of test assessment based on years of industry insight and service experience, and on the basis of embracing leading large model technology. In the entire examination and evaluation process, the large model can solve the intelligent solutions of the whole process such as proposition, invigilation, scoring and evaluation.
In terms of propositions, LLM intelligent question solving, through the fine-tuning training of 530,000 test questions, can already support the question needs of most general scenarios, as well as the question needs of 171 professional qualification examinations and professional field certifications. After the test of the certification project, the use of LLM intelligent question solving can reduce the cost of solving questions by 95%, and the efficiency of solving questions can be increased by more than 14 times.
In the invigilation scenario, multimodal AI-assisted invigilation is realized through gesture recognition, facial recognition, voice content monitoring, etc., and the tendency to cheat in the exam is divided into levels based on a certain algorithm, so as to achieve hierarchical and accurate invigilation, which greatly improves the invigilation efficiency of the first exam. Based on the database of cheating behaviors generated by more than 2 million exams, the invigilation efficiency is increased by 10 times.
In the LLM scoring scenario, after data training and model fine-tuning of 11 million test papers, and after project measurement, it is now possible to achieve a speed of 90% and 95% with manual scoring, achieving a credible and usable state. It can be applied to the scoring of various subjective test papers, improving the scoring efficiency by more than 7 times. In 2024, we will also carry out in-depth cooperation with a number of institutions to create a pilot and benchmark for large-scale LLM scoring in the industry.
In addition to general large model technology manufacturers, Examstar will also carry out strategic cooperation with Tsinghua University on multi-modal large models to promote large models into the industry. In the future, ExamStar will carry out in-depth cooperation with more and more enterprises and institutions on large-scale model business, explore and optimize multiple scenarios in the field of examination and assessment, and promote the high-quality development of the field of examination and evaluation.