Research map / Large language models / Language, safety and interpretability
LongWriter-Zero: Mastering Ultra-Long Text Generation via Reinforcement Learning
Yuhao Wu, Yushi Bai, Zhiqiang Hu, Roy Ka-Wei Lee, Juanzi Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add a think step (Think prompt) to RL training (ablation) | LongWriter-Zero (Qwen2.5-32B base) · Arena-Write | Elo: 700 -> 1,200 (Base-nothink vs Base-think) | Improved |
| add continual pretraining on 30B writing tokens before RL (ablation) | LongWriter-Zero · Arena-Write | Elo: 1221 -> 1447 | Improved |
| add continual pretraining before RL (ablation) | LongWriter-Zero · WritingBench | average critic score: 8.12 -> 8.69 | Improved |
| add think step (ablation) | LongWriter-Zero · WritingBench | average critic score: 8.04 -> 8.12 | Improved |
| add think step (ablation) | LongWriter-Zero · Arena-Write | Elo: 668 -> 1221 | Improved |
| use RL instead of SFT on the continual-pretrained base (ablation) | Qwen2.5-32B continual-pretrained · Arena-Write | Elo: 971 -> 1447 | Improved |
| add continual pretraining before SFT (ablation) | Qwen2.5-32B SFT · Arena-Write | Elo: 964 -> 971 | No clear effect |
| RL with full pipeline instead of LongWriter synthetic-data SFT (ablation) | Qwen2.5-32B · WritingBench | average critic score: 8.08 -> 8.69 | Improved |
| apply the RL pipeline at 14B scale | Qwen2.5-14B · WritingBench | average critic score: 7.60 -> 8.49 | Improved |
| apply the RL pipeline at 14B scale | Qwen2.5-14B · Arena-Write | Elo: 744 -> 1198 | Improved |
| train with composite length, writing and format rewards from base model | Base-nothink · Arena-Write | Elo: 200 -> over 600 | Improved |
| normalize each reward component within group before averaging advantages | final reward · RL training | training stability | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- LongWriter-8B — LongWriter: Unleashing 10,000+ Word Generation from Long Context LLMs (ICLR 2025) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- LongWriter: Unleashing 10,000+ Word Generation from Long Context LLMs (ICLR 2025 · similarity 0.93)
- LongGenBench: Benchmarking Long-Form Generation in Long Context LLMs (ICLR 2025 · similarity 0.92)
- Successor Features for Efficient Multi-Subject Controlled Text Generation (ICML 2024 · similarity 0.89)
- RewriteLM: An Instruction-Tuned Large Language Model for Text Rewriting (AAAI 2024 · similarity 0.89)
- Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation (ICLR 2025 · similarity 0.89)
- WritingBench: A Comprehensive Benchmark for Generative Writing (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
