ITRI Technology Completes AI Model Inspection in 90 Minutes, Helping Businesses Reduce Costs

Central News Agency, Hsinchu, May 14 (Reporter Chang Chien-chung) - The Industrial Technology Research Institute (ITRI) announced today that its developed Trustworthy AI Automated Evaluation Technology can significantly shorten the manual evaluation process, which originally took two weeks, to just 90 minutes. This helps businesses reduce the cost and barriers of AI implementation and testing, accelerating the deployment of trustworthy AI applications. To date, it has assisted in over 110 corporate AI model inspection cases.

ITRI released a press release today in which Li Yu-tai, Manager at the Center for Measurement Standards, stated that in the past, companies conducting language model testing relied heavily on manual, question-by-question checks. This was not only time-consuming and labor-intensive but also struggled to keep up with the rapid updates and iterations of models.

Li Yu-tai pointed out that ITRI's Center for Measurement Standards, combined with the technical capabilities of ITRI's Information and Communications Research Laboratories in developing testing tools and question banks, has established an automated AI evaluation mechanism. Through technologies like smart question selection, prompt design, and semantic matching, it automatically checks whether the model's answers comply with regulations, greatly improving testing efficiency and consistency.

Li said that the team has currently assisted in over 110 AI model inspection cases, with applications spanning large language models, smart customer service, business systems, and government documents. Over 90% of these cases come from private enterprises.

Tsai Shu-yu, Project Manager at ITRI's Information and Communications Research Laboratories, stated that most international open-source models are primarily trained on English data, which can easily lead to content that does not fit Taiwan's linguistic and cultural context. The team has specifically established a question bank for Traditional Chinese and Taiwanese cultural contexts to enhance the model's testing capabilities in language appropriateness, cultural sensitivity, and localized applications.

ITRI noted that the "Trustworthy AI Automated Evaluation Technology" received the Grand Award in the "Implementation and Transformation: Best Solution" category at the 2026 AI Award from the Taiwan AI Academy. The AI Award by the Taiwan AI Academy is a benchmark AI award in the country, focusing on the application of AI technology and industrial transformation achievements. (Editor: Lin Shu-yuan) 1150514