A joint study by researchers from China and the United States has unveiled a new system called “AbsoluteZero Reasoner,” which enables artificial intelligence models to improve their programming and reasoning capabilities through “self-play,” posing complex problems to themselves and solving them, thereby bypassing traditional training methods reliant on human data.
Article contents:
- Introduction
- The concept of learning through questioning
- The “AbsoluteZero Reasoner” system
- Results and superior performance
- Human simulation in learning
- Major laboratories adopting the approach
- The path toward superintelligence
- Frequently asked questions
Introduction
The smartest artificial intelligence models have long relied primarily on imitation; they learn either by consuming massive amounts of human-produced examples or by attempting to solve problems set for them by human trainers. However, AI appears to be starting to learn in a more human-like manner, by discovering interesting questions to ask itself and independently trying to find the correct answer.
The concept of learning through questioning
A joint research project from Tsinghua University, the Beijing Institute for General Artificial Intelligence, and Pennsylvania State University demonstrates that artificial intelligence can learn logical thinking and reasoning in this way by handling computer code. The core idea is that the model does not merely stop at input data, but begins generating new challenges for itself to transcend traditional training limits.
The “AbsoluteZero Reasoner” system
Researchers created a system they named “Absolute Zero Reasoner.” The system initially operates by using a large language model to generate Python programming problems that are challenging yet solvable. Afterward, the system uses the same model to solve those problems before verifying its work by attempting to run the code.
In the final step, the system uses successes and failures as a signal to refine the original model, simultaneously enhancing its ability to pose better problems and solve them. This closed-loop self-improvement allows the model to evolve without the constant need for new external data.
Results and superior performance
The research team found that their approach significantly improved the programming and reasoning skills for both the 7-billion and 14-billion parameter versions of the open-source language model “Qwen.” Most impressively, the model outperformed some models trained using human-curated and carefully selected data, proving the effectiveness of self-learning.
Human simulation in learning
Andrew Zhao, a PhD student at Tsinghua University and the original creator of the project, explained that this approach resembles how human learning transcends rote memorization or imitation. Zhao said: “At first, you imitate your parents and do as your teachers do, but then you essentially have to ask your own questions. Eventually, you can surpass those who taught you in school.”
Researchers noted that the idea of AI learning in this manner, sometimes referred to as “self-play,” dates back years and was previously explored by pioneers in the field such as Jürgen Schmidhuber and Pierre-Yves Oudeyer.
Major laboratories adopting the approach
There are early indicators that the “absolute zero” approach is gaining traction in some major AI laboratories. For instance, a project called “Agent 0” from Salesforce, Stanford, and the University of North Carolina involves a software agent that improves itself through self-play. As with the Chinese project, the model improves its general reasoning through trial-and-error problem-solving.
Additionally, a recent paper by researchers at Meta, the University of Illinois, and Carnegie Mellon University introduced a system utilizing a similar type of self-play for software engineering, indicating that this technique may be the first step toward training models for super-intelligent software agents.
The path toward superintelligence
One of the most exciting elements of the project, according to researcher Zilong Zheng, is the way the model’s problem-posing and problem-solving skills scale. The challenges become more difficult as the model grows more powerful. Nevertheless, the primary challenge remains that the system currently operates only on easily verifiable problems, such as mathematics or coding.
As the project advances, it may become possible to use it for agentic AI tasks such as web browsing or office work. The staggering prospect of an approach like this is that it could, theoretically, allow models to surpass human teaching. Zheng stated: “Once we have that, it’s sort of a path to artificial superintelligence.”
Frequently asked questions
What is the “AbsoluteZero Reasoner” system?
Answer: It is an artificial intelligence system developed by Tsinghua University that generates programming problems for itself and solves them, then uses the results to improve its capabilities without human intervention.
How does this approach differ from traditional AI training?
Answer: Traditional training relies on feeding the model human data and examples to imitate, whereas this approach relies on “self-play” and posing new questions to enhance independent reasoning capacity.
Did this model outperform human-trained models?
Answer: Yes, results showed that the “Qwen” model trained with this system outperformed in programming and reasoning skills some models that received human-curated and selected data.
What are the future applications of this technology?
Answer: Researchers hope to expand this system to encompass more complex tasks such as web browsing and office work, which could pave the way toward developing artificial superintelligence that exceeds human capabilities.
