This approach aims to train reasoning models without relying on human-curated data, allowing models to learn by proposing tasks, solving them, and learning from

This approach aims to train reasoning models without relying on human-curated data, allowing models to learn by proposing tasks, solving them, and learning from both stages through self-play with the aid of an environment. The core of this research is the Absolute Zero Reasoner (AZR), which focuses on proposing and solving coding tasks, utilizing a code executor for verifiable feedback. Key Findings and Contributions: State-of-the-Art Performance: AZR has demonstrated state-of-the-art performance in coding and mathematical reasoning tasks, outperforming models trained on traditional human-curated datasets. Enhanced Reasoning Capabilities: The study suggests that coding capabilities developed through AZR training may amplify overall improvements in reasoning. Models trained with AZR showed stronger gains in generalized reasoning compared to those trained with expert code. Scalability: The performance improvements observed with AZR appear to scale with the size of the model. Cognitive Behaviors: AZR exhibits emergent cognitive behaviors such as step-by-step reasoning and trial-and-error. The research also noted that token counts grow with training and vary depending on the type of task. (Summarized by Gemini)