Scientists from T-Bank AI Research have improved the safety and accuracy of AI responses by up to 15%

Scientists from the T-Bank AI Research Laboratory for Artificial Intelligence (AI) have developed a method for teaching large language models (LLM), which increases the quality of AI responses by up to 15% in five indicators. This was reported to Izvestia on April 25 by the organization's press service.

It is specified that the created methodology is based on the existing Trust Region methods. The results of the study have gained recognition in the global community, including being presented at the International Conference on Representation Training (ICRL), which is taking place in Singapore on April 24-28.

The new method can be applied in various directions, including the creation of virtual assistants and chatbots in areas ranging from education to medicine. Among its advantages, scientists named improved text generation quality, reduced effect of excessive optimization, and ease of implementation.

"Our new approach allows us to maintain a balance between the model's ability to solve a new narrow task and a general understanding of the worldview, which opens up opportunities for creating more flexible and adaptive models. This area is far from exhausted — scientists still have a lot of space for further AI research and improvements that can lead to new breakthroughs in optimizing language models and their application in the real world," said Boris Shaposhnikov, head of the AI Alignment research group at the T—Bank AI Research Laboratory.

It is noted that in the future, the Trust Region method will play a significant role in creating more effective language models and laying the foundation for a new paradigm in the development of artificial intelligence.

The experts tested the method on Alpaca Eval 2.0 and Arena Hard metrics. The Alpaca Eval 2.0 test showed an improvement in the quality of AI responses from 2.3% to 15.1%, with useful and relevant responses more often observed.

It is noted that in the learning process, the language model can deviate from the settings. Subsequently, this provokes a sharp decline in the quality of responses, and also affects the likelihood of "misleading" the model on the learning path. As a counter measure, scientists from T-Bank AI Research suggested periodically updating the "default settings". According to them, this method allows the model to mark key "landmarks" and avoid deviations on the way to the target point.

It clarifies that the Trust Region takes into account the dynamic change of this point through a soft update with minor changes at each stage of training, or a hard update entirely with a certain frequency. According to experiments, both types of updates lead to clearer and safer responses. Thus, the responses of models trained on the task of shortening long texts have improved by 10-15%, the AI is also less confused in complex tasks and follows instructions better.

In geometric progress: Russia should become a leader in the scale of AI application

At the AI Journey conference, Putin predicted a revolution in several areas and learned how GigaChat sped up the processing of direct line appeals

On April 22, Roman Koposov, an expert in the field of digital transformation, Deputy director of ARB Pro Strategic planning company, shared with Izvestia that in the coming years, AI-based multi-agent systems will take the place of traditional management teams in both business and the Russian public sector. The analyst emphasized that the world is moving from the concept of "one smart algorithm" to the architecture of collective AI.