LongWriter-Zero: Mastering Ultra-Long Text Generation via Reinforcement Learning

Researchers introduce LongWriter-Zero, a reinforcement learning approach that enables large language models to generate ultra-long, high-quality text without relying on synthetic training data. The 32B parameter model outperforms traditional supervised fine-tuning methods and even surpasses larger 100B+ models on long-form writing benchmarks.