LongWriter: Unleashing 10,000+ Word Generation from Long Context LLMs

TL;DR

This paper introduces LongWriter, a method to extend the output length of long context LLMs beyond 10,000 words by creating specialized datasets and benchmarks, demonstrating that existing models have untapped potential for ultra-long generation.

Contribution

We propose AgentWrite and LongWriter-6k to significantly increase output lengths of LLMs, and develop LongBench-Write for evaluating ultra-long generation, achieving state-of-the-art results.

Findings

Models can generate over 10,000 words with proper training data.

AgentWrite decomposes tasks to enable ultra-long output generation.

State-of-the-art performance on ultra-long generation benchmark.

Abstract

Current long context large language models (LLMs) can process inputs up to 100,000 tokens, yet struggle to generate outputs exceeding even a modest length of 2,000 words. Through controlled experiments, we find that the model's effective generation length is inherently bounded by the sample it has seen during supervised fine-tuning (SFT). In other words, their output limitation is due to the scarcity of long-output examples in existing SFT datasets. To address this, we introduce AgentWrite, an agent-based pipeline that decomposes ultra-long generation tasks into subtasks, enabling off-the-shelf LLMs to generate coherent outputs exceeding 20,000 words. Leveraging AgentWrite, we construct LongWriter-6k, a dataset containing 6,000 SFT data with output lengths ranging from 2k to 32k words. By incorporating this dataset into model training, we successfully scale the output length of existing models to over 10,000 words while maintaining output quality. We also develop LongBench-Write, a comprehensive benchmark for evaluating ultra-long generation capabilities. Our 9B parameter model, further improved through DPO, achieves state-of-the-art performance on this benchmark, surpassing even much larger proprietary models. In general, our work demonstrates that existing long context LLM already possesses the potential for a larger output window--all you need is data with extended output during model alignment to unlock this capability. Our code & models are at: https://github.com/THUDM/LongWriter.