BAKU, Azerbaijan, October 15. Azerbaijan Technical University is developing an approach to detect "hallucinations" in large language models (LLMs) by utilizing several indicators, including response generation time and temperature fluctuations in graphics processing units (GPUs), said Executive Director of the Institute of Defense Technologies and Cybersecurity at Azerbaijan Technical University, Rahid Alakbarli, during the 20th IEEE International Conference on the Application of Information and Communication Technologies (AICT2026) in Baku on Thursday, Trend’s correspondent reports from the event.
According to Alakbarli, the proliferation of large language models and chatbots has brought about the issue of "hallucinations", instances where artificial intelligence (AI) generates inaccurate or erroneous information. This raises concerns regarding the reliability of such systems, particularly when applied within the defense industry.
"We have concluded that without proper management of large language models, we cannot place our trust in AI," he noted.
To investigate this issue, researchers are analyzing how large language models respond to queries, which includes measuring temperature changes in GPUs while questions are being processed.
Alakbarli explained that, much like humans who require more time to think when addressing complex issues, the researchers are exploring whether similar patterns exist in the operation of large language models, and whether these patterns can be used to identify potential hallucinations.
He said that the study employs five theoretical approaches to evaluate model performance, including metrics based on Fisher information, Kullback-Leibler divergence, and methods derived from information geometry.
Alakbarli added that the research also accounts for response generation time, the number of tokens used, and various performance indicators related to the GPUs.
He emphasized that an increase in the graphics processor's temperature can be attributed not only to the specifics of request processing but also to other computational factors, including an increased workload. Therefore, this metric is considered in conjunction with other data.
