Absolute Zero Reasoner. AI News – #2 May

Absolute Zero Reasoner (AZR), developed by researchers from Tsinghua University, the Beijing Institute for General Artificial Intelligence, and Pennsylvania State University, is an artificial intelligence model designed to autonomously develop reasoning abilities. Its fundamental feature is the ability to independently generate tasks for itself and then solve them. Importantly, the learning process of AZR is based on verifying the correctness of these solutions through an objective, external mechanism – in this case, a code executor – and does not require any pre-prepared training data by humans. This model operates within the RLVR (Reinforcement Learning from Verifiable Reward) paradigm, called “Absolute Zero,” meaning its development is driven by a reward system based on verifiable results of its own work. Sounds complicated? Let me explain!

AZR can autonomously generate tasks and then solve them, maximizing its own learning progress. Most importantly – it does this without the need to use any external data prepared by humans. It’s a bit like giving AI a sandbox and a shovel, and it starts building increasingly complex castles, learning from every grain of sand.

At the heart of AZR is a clever mechanism in which the AI model plays two roles simultaneously:

The entire system learns through interaction with an environment that provides verifiable feedback. In the case of AZR, this environment is the code executor. It can objectively check whether the generated code works correctly and produces the expected results. It’s a bit like a judge at a competition who fairly evaluates the performance.

This creates a loop of continuous improvement:

And so on, in a loop, with the AI becoming better and better at inventing useful tasks and solving them, starting from absolute zero – hence the name.

The creators of AZR focused on the domain of coding. Why?

AZR learns three fundamental types of reasoning about code: deduction (predicting outcomes), abduction (inferring inputs from outcomes), and induction (creating a program based on examples).

AZR’s results are impressive. The model can outperform other systems in certain tasks that were trained on huge, human-prepared datasets. It excels particularly well at generalizing skills to new, previously unknown domains. Sounds like a revolution, right?

However, some temper the enthusiasm. They point out that although AZR is impressive, it does not eliminate the fundamental problems of large language models (LLMs). It is rather a very advanced automation of synthetic data generation. Here, AI does not learn to “think” in a completely new way but rather becomes a master at solving specific tests and tasks it sets for itself. The concept of “self-play” is also not new.

However, the fact that AZR achieves such good results in coding and math tasks training completely without external data is extremely interesting. This is where the potential paradigm shift lies – instead of flooding AI with tons of data, we allow it to explore and discover on its own.

Let’s think about how humans learn. Sure, we use books and teachers (which resembles supervised learning in AI), but a huge part of our knowledge comes from interacting with the world, experimenting, setting challenges for ourselves, and learning from successes and failures. AZR tries to imitate this very process, driven by curiosity and exploration.

Is Absolute Zero Reasoner a true breakthrough and milestone? It is certainly a fascinating research direction that shows the path to more autonomous and, perhaps, more “intelligent” artificial intelligence might run through self-driven discovery of knowledge. As always, the future will show how much these promising concepts will change the AI landscape. One thing is certain – it’s worth staying tuned, so keep up with AI news with Delante and subscribe to our newsletter!