Russian Researchers Teach AI Models to Learn From a Single Glance
A new training method based on tracking human eye movement across text could cut learning time for AI models nearly in half.
Scientists at Innopolis University have developed a technique that reduces the training time of large language models by 1.5 to 2 times. The method is built on analyzing how humans read and process text.
According to the university’s press service, researchers have for the first time applied data on human eye movement while reading to fine-tune large language models. The findings make it possible to align AI responses with user expectations more quickly, significantly shortening overall training time.
Feedback at the Micro Level
The development team refined the RLHF method – reinforcement learning from human feedback. Instead of receiving a general evaluation of its output, the model is given detailed information about which parts of the generated text attract the most user attention.
The system analyzes attention distribution and incorporates it into the reward mechanism used to train the AI. This allows the model to fine-tune its responses more precisely and reduce unnecessary training iterations.
Energy Efficiency
Researchers estimate that the new approach lowers computational resource requirements, making the development of AI assistants more energy-efficient. Faster training could pave the way for more accurate and adaptive digital assistants.
The technology could be applied in healthcare, education, and corporate services – areas where fast and accurate interaction between humans and AI systems is essential.
