AI Models Learn by Generating and Solving Their Own Coding Problems

New Self‑Play Learning Framework

A collaborative team from Tsinghua University, the Beijing Institute for General Artificial Intelligence (BIGAI), and Pennsylvania State University introduced a system named Absolute Zero Reasoner (AZR). The system leverages a large language model to first generate solvable yet challenging Python coding tasks, then uses the same model to attempt solutions, and finally checks the code by executing it. Successes and failures feed back into the model, refining its ability to both create better problems and solve them.

Performance Gains

Testing the method on open‑source language models Qwen with 7 billion and 14 billion parameters revealed significant improvements in coding and reasoning performance. In some cases, the refined models outperformed larger models that had been trained on human‑curated data.

Human‑Like Learning

The researchers liken the process to how humans move beyond imitation, first copying teachers and then formulating their own questions to surpass prior instruction. This self‑play concept has roots in earlier work by AI pioneers and aligns with recent efforts at other institutions to use self‑generated tasks for model improvement.

Future Directions

While currently limited to problems with clear, automatic verification such as coding or math, the team envisions extending the approach to broader agentic tasks like web browsing or office automation. Successful expansion could bring AI systems closer to autonomous learning capabilities that require less human‑provided data.