Robots will be able to capture humanity by doing homework with your kids.
A group of Chinese scientists from Microsoft Research Asia, Peking University and Xi'an Jiaotong University presented an innovative method for teaching language models (Large Language Model, LLM). Based on the principles of human learning, experts have developed a strategy called "Learning from Mistakes" (LeMa), which allows artificial intelligence to independently learn from its own mistakes and thus significantly improve the quality of solving mathematical problems.
The study is based on a unique idea: if a person, faced with an error in a mathematical problem, analyzes and corrects it, then the machine, learning by the same principle, will be able to increase its efficiency. Applying their approach, the team used incorrect reasoning models such as LLaMA-2, then GPT-4 analyzed these errors, training the model to provide correct solutions.
This approach has been extremely effective. On two mathematical problems and using five different LLMs, the LeMa method showed a steady improvement in performance compared to previous techniques based only on data on standard solution paths. Moreover, specialized LLMs, including WizardMath and MetaMath, also showed improvements, achieving 85.4% accuracy on the GSM8K dataset and 27.1% on MATH, which is a record for open source models.
The discovery holds 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 their knowledge in the field of machine learning.
Complex AI applications, such as healthcare, finance, and autonomous vehicle development, have the potential to benefit enormously from the introduction of such systems, which are capable of continuous self-improvement through analysis and correction of their own mistakes.
The research opens up new horizons: we are approaching an era in which artificial intelligence not only performs tasks, but is also able to learn effectively, improving its ability to solve increasingly complex problems, which was previously considered the prerogative of humans.
A group of Chinese scientists from Microsoft Research Asia, Peking University and Xi'an Jiaotong University presented an innovative method for teaching language models (Large Language Model, LLM). Based on the principles of human learning, experts have developed a strategy called "Learning from Mistakes" (LeMa), which allows artificial intelligence to independently learn from its own mistakes and thus significantly improve the quality of solving mathematical problems.
The study is based on a unique idea: if a person, faced with an error in a mathematical problem, analyzes and corrects it, then the machine, learning by the same principle, will be able to increase its efficiency. Applying their approach, the team used incorrect reasoning models such as LLaMA-2, then GPT-4 analyzed these errors, training the model to provide correct solutions.
This approach has been extremely effective. On two mathematical problems and using five different LLMs, the LeMa method showed a steady improvement in performance compared to previous techniques based only on data on standard solution paths. Moreover, specialized LLMs, including WizardMath and MetaMath, also showed improvements, achieving 85.4% accuracy on the GSM8K dataset and 27.1% on MATH, which is a record for open source models.
The discovery holds 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 their knowledge in the field of machine learning.
Complex AI applications, such as healthcare, finance, and autonomous vehicle development, have the potential to benefit enormously from the introduction of such systems, which are capable of continuous self-improvement through analysis and correction of their own mistakes.
The research opens up new horizons: we are approaching an era in which artificial intelligence not only performs tasks, but is also able to learn effectively, improving its ability to solve increasingly complex problems, which was previously considered the prerogative of humans.
