QwenLong-L1 solves long-context reasoning challenge that stumps current LLMs
1 min read
Summary
Alibaba has launched QwenLong-L1, a system that enables large language models (LLMs) to draw conclusions from massive documents.
It can be used on legal contracts or detailed corporate filings to extract information and answer questions.
The system organises the training of the LLMs into a trilogy of short-context reinforcement learning, warm-up supervised fine-tuning and curriculum-guided phased reinforcement learning, ensuring the models learn to extract and reason step-by-step.
Alibaba said the system outperformed similar models from DeepSeek, OpenAI and Google, which have not been trained to deal with as many tokens in one go.