A New Method to Improve Mental Execution Skills of Artificial Intelligence Models
Large language models (LLM) are the main goal of researchers for a long time to improve the ability to solve complex problems. A new research reveals an innovative framework called Self-Play Search Distillation (SPSD) against this difficulty. This method is possible to produce premium quality synthetic data that models can learn.
Current problems can be summarized as follows:
-Data quality problem: Because synthetic data is low quality, models learning is limited
- Depression cost: Very expensive and time-consuming to achieve high quality data labeled by human
- Included execution: Available methods are difficult decisions, alternative options and insufficient to provide data with their results
- MuZero-like networks: Frame is able to transfer using MuZero style networks trained in board games such as chess
On the basis of the SPSD technique, artificial intelligence networks play with themselves (self-play) lies the idea of learning themselves. This approach adapts the methods used successfully in board games to language mode. The data obtained is a compelling and varied, up to models need to do difficult interests.
This development can revolutionize the training of the models of the **dil, which can make smarter and better mind. Especially if the quality of synthetic data increases, the dependence with human strength will be reduced, and the logical thinking skills of models can significantly convey. Research is promising to show how data failure in the field of artificial intelligence can be exceeded.