AI Models Achieve Breakthrough By Learning Through Self-Generated Questions

In a groundbreaking development, researchers from Tsinghua University, the Beijing Institute for General Artificial Intelligence (BIGAI), and Pennsylvania State University have successfully enabled artificial intelligence (AI) models to learn by posing questions to themselves. This innovative approach, dubbed “self-play,” allows AI to transcend traditional learning methods, which rely heavily on human instruction and example-based training. The team devised a system called Absolute Zero Reasoner (AZR), which leverages a large language model to generate challenging yet solvable Python coding problems.

The model then attempts to solve these problems, checks its work by running the code, and refines its abilities based on successes and failures. This cyclical process enables the model to augment its capacity for both problem-posing and problem-solving. According to Zhao, a researcher involved in the project, this approach mirrors human learning, which evolves beyond mere imitation and rote memorization.

“In the beginning, you imitate your parents and do as your teachers instruct, but then you basically have to ask your own questions,” he explained. “And eventually, you can surpass those who taught you back in school.” The researchers tested their approach using the open-source language model Qwen, with 7 billion and 14 billion parameter versions.