A group of Chinese scientists from Microsoft Research Asia, the University of Beijing and the University of Sian Jiaotun presented innovative method of language models (Large Language Model, Language Model. LLM). Based on the principles of human training, experts have developed a strategy called “Learning from Mistakes, Lema), which allows artificial intelligence to independently study on their own mistakes and thus significantly improve the quality of solving mathematical tasks.
The study is based on a unique idea: if a person, faced with a mistake in a mathematical task, analyzes and corrects it, then the machine, studying on the same principle, will be able to increase his effectiveness. Applying their approach, the team used incorrect discussions of models such as LLAMA-2, then the GPT-4 analyzed these errors, teaching the model to provide the right solutions.
This approach was extremely effective. On two mathematical tasks and using five different LLMs, the LEMA method showed a stable improvement in performance compared to previous techniques based only on data on standard solutions. Moreover, specialized LLM, including WizardMath and Metamath, also demonstrated an improvement, reaching accuracy of 85.4% at the GSM8K data set and 27.1% on MATH, which is a record for open source models.
The discovery has important prospects for the development of artificial intelligence. The source code, data and models used in the study, are now available on github , which allows the research community around the world to continue working in this direction , deepening and expanding knowledge in the field of machine learning.
Complex areas of the AI, such as healthcare, finance and development of autonomous vehicles, can potentially benefit from the introduction of such systems that are capable of continuous self -improvement through analysis and correction of their own errors.
The study opens up new horizons: we are approaching the era in which artificial intelligence not only performs tasks, but is also able to study effectively, increasing our ability to solve more and more difficult problems, which was previously considered the prerogative of a person.