Аналіз та імплементація ідеї навчання великих мовних моделей за аналогією з дитячим когнітивним розвитком

Abstract

The study is devoted to the problem of not enough efficiency of large language models in solving mathematical problems, sequential multi-step analysis and reasoning about possible solutions. The article reveals the imperfection of modern LLMs in the issues of sequential thinking, despite their ability to generate natural and meaningful text, they still remain significantly limited in solving problems that require logical reasoning or justification. It was found that such limitations are often a consequence of the structure of the learning process of the language model, because it has nothing to do with the processes that occur during the formation of cognitive and mental models of a person. This work considers the approach of learning by similarity to human cognitive development, which is based on a combination of the following concepts: Curriculum learning – learning with a gradual increase in the complexity of tasks, Scaffolding – segmented learning with a decrease in the level of care after mastering a certain level of knowledge, and progressive strengthening of mental structures by creating generalized patterns for solving problems. The main attention is focused on the implementation of the learning process with a gradual complication of the input material – from basic arithmetic equations to large complex logical-verbal problems. The effectiveness of this approach is evaluated on the example of Phi-2 and Mistral 7B models, trained in two modes: standard random learning and the proposed supervised learning. The results of the study prove that the proposed cognitive learning approach provides better results when solving basic tasks and complex multi-component tasks that require intermediate generalized knowledge. The analysis of similar studies confirms the promising development of training with imitation of cognitive development to increase the efficiency of processing input data under conditions of a limited amount of training information.