AI That Teaches Itself: Tsinghua University’s ‘Absolute Zero’ Trains LLMs With Zero External Data
Researchers developed Absolute Zero Reasoner (AZR), enhancing language models' reasoning without external data. AZR generates and solves its own tasks, creating a self-evolving curriculum. It uses a code executor for task validation and answer verification. AZR-Coder-7B achieved state-of-the-art performance in overall and coding averages. Larger models benefit more from AZR, showing improvements beyond 200 training...