1
Absolute Zero Reasoner
Current reasoning models trained with Reinforcement Learning with Verifiable Rewards (RLVR) often rely on manually curated datasets, raising scalability concerns and potentially limiting future AI growth beyond human-defined tasks. To address this, we propose the Absolute Zero paradigm, where a single model…