Paper summary: Absolute Zero: Reinforced Self-play Reasoning with Zero Data

Problem

RL with verifiable rewards (“zero” settings) avoids labelling reasoning steps, but still needs curated question–answer sets. The authors argue that relying on human-made examples won’t scale, and might limit a future system that’s better than humans.

Method

Absolute Zero: a single model learns to propose tasks that maximise its own learning progress, and to solve them. The Absolute Zero Reasoner (AZR) uses a code executor to:

• validate that a proposed code-reasoning task is well-formed
• verify the solver’s answers

The executor is the one source of verifiable reward, which keeps the open-ended curriculum grounded.

Key results

From the abstract: trained entirely without external data, AZR reaches overall state-of-the-art performance on coding and mathematical reasoning, beating zero-setting models that use tens of thousands of human-curated examples. It works across model scales and model classes. Figure 1 summarises the results.

Limitations

There’s no dedicated limitations section in the extracted text. The approach relies on tasks that a code executor can verify.