Russian scientists refine AI to accelerate research with 73% less memory

Researchers at HSE University have refined a large language model to better handle Russian scientific terminology. Their work aims to create AI tools that act as partners for scientists, automating complex analysis. The improvements include faster processing and far lower memory demands than the original system. The team adapted the AI using the iFORA-QA dataset, a curated collection of reports and materials in science, technology, and innovation. This fine-tuning allowed the system to answer highly specialised questions with greater accuracy.

After adaptation, the AI’s response speed increased by 2.7 times. Memory requirements also fell sharply, dropping by nearly three-quarters compared to the original open-source model. The researchers now plan to add extra tools to further reduce the risk of AI-generated inaccuracies, known as hallucinations. The long-term goal is to build a full system of intelligent agents tailored to Russian scientific needs. These agents would assist researchers by handling data analysis and providing reliable insights.

The adapted AI now runs 2.7 times faster while using 73% less memory. Its improved precision in specialised fields makes it more practical for research applications. Future updates will focus on strengthening factual reliability and expanding its scientific capabilities.