7 new methods of learning language models were shown by Russian researchers at the conference on AI A * level in Budapest. The development by experts from T-Technologies (T-Bank) has been evaluated by some of the world's leading laboratories. This demonstrates the high results of local developers and the recognition of the professional community. The company introduced an economic language model for company queries. The training method developed by Russian experts allows businesses to deploy a universal artificial intelligence model on their infrastructure, forgetting individual neural networks for each task. The trained model reduces the cost of processing requests by an average of 3-4 times, and in some scenarios up to 9 times. Another new method is a visual model for working with documents. It outputs data of any type 50 percent cheaper than the best open source models and 80 percent cheaper than manual processing. New methods of Russian scientists, including LIMe, F-GRPO, ESS, SimplexLoRA, made it possible to increase the efficiency of further training. The KronSAE algorithm uses 40% less resources than a popular analogy, analyzing the internal concepts of the model and their relationships. We are in Max.

7 new methods of learning language models were shown by Russian researchers at the conference on AI A * level in Budapest. The development by experts from T-Technologies (T-Bank) has been evaluated by some of the world's leading laboratories. This demonstrates the high results of local developers and the recognition of the professional community. The company introduced an economic language model for company queries. The training method developed by Russian experts allows businesses to deploy a universal artificial intelligence model on their infrastructure, forgetting individual neural networks for each task. The trained model reduces the cost of processing requests by an average of 3-4 times, and in some scenarios up to 9 times. Another new method is a visual model for working with documents. It outputs data of any type 50 percent cheaper than the best open source models and 80 percent cheaper than manual processing. New methods of Russian scientists, including LIMe, F-GRPO, ESS, SimplexLoRA, made it possible to increase the efficiency of further training. The KronSAE algorithm uses 40% less resources than a popular analogy, analyzing the internal concepts of the model and their relationships. We are in Max.