Russian researchers from the T-Bank AI Research laboratory have presented a new approach to training large language models. Response accuracy reached 15%. The work of Russian specialists was presented at the international ICRL conference in Singapore.
The method of Russian researchers is based on advanced Trust Region technologies used in various fields of artificial intelligence. With their help, the quality of AI responses across five key indicators is significantly improved.
The development can be used in a variety of fields, including intelligent assistants and medical diagnostic systems. Scientists are confident that thanks to their approach, the probability of errors in responses will decrease.
In AlpacaEval 2.0 and Arena Hard tests, the new method showed an improvement of 2.3 to 15.1 percentage points compared to traditional methods. Language models have become 10–15% more efficient in summarizing texts and more accurately follow user instructions.
Read more on the topic:
Scientists create AI to process data about the Sun
Russian scientists have trained a neural network to search for malicious programs
Saratov scientists have found an alternative to neural networks
