[PDG 450] Absolute Zero: Reinforced Self-play Reasoning with Zero Data

Link to article: https://arxiv.org/pdf/2505.03335
Title: Absolute Zero: Reinforced Self-play Reasoning with Zero Data
Content: The Absolute Zero paradigm introduces AZR (Absolute Zero Reasoner), a system that generates its own training tasks and improves reasoning abilities without any external data, using a code executor to validate tasks and verify answers as a unified reward source. This approach addresses scalability concerns of current reinforcement learning methods that still depend on human-curated question-answer datasets, even when they avoid direct supervision of reasoning processes. Despite training entirely without external data, AZR achieves state-of-the-art performance on coding and mathematical reasoning benchmarks, outperforming existing zero-setting models that rely on tens of thousands of human examples.
Slack link: ml-ka.slack.com, channel: #pdg. Please join us -- if you cannot join, please message us here or to mlpaperdiscussiongroupka@gmail.com.

In the Paper Discussion Group (PDG) we discuss recent and fundamental papers in the area of machine learning on a weekly basis. If you are interested, please read the paper beforehand and join us for the discussion. If you have not fully understood the paper, you can still participate – everyone is welcome! You can join the discussion or simply listen in. The discussion is in German or English depending on the participants.