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.

• →LongWriter-Zero uses reinforcement learning instead of synthetic data to train models for ultra-long text generation.
• →The approach starts from scratch without annotated or synthetic data, using specialized reward models for length control and quality.
• →The 32B model outperforms much larger models including DeepSeek R1 and Qwen3-235B on writing benchmarks.
• →Traditional supervised fine-tuning approaches suffer from costly, artificial, and structurally monotonous synthetic data.
• →The model and data have been open-sourced for research community use.