New LongZero method enhances LLM long-context reasoning without external data

Researchers have introduced LongZero, a novel method designed to enhance the long-context reasoning capabilities of large language models (LLMs) without requiring any external training data. This approach enables LLMs to generate their own tasks, including multi-hop reasoning and multi-round retrieval exercises, by synthesizing questions and answers with supporting evidence. LongZero then expands this evidence into long, multi-document contexts and trains the model using GRPO. Experiments on two LLMs across five benchmarks demonstrate that LongZero outperforms baselines that rely on externally synthesized data, indicating a promising direction for improving LLM long-context reasoning. AI

IMPACT This method could significantly reduce the cost and complexity of training LLMs for tasks requiring long-context understanding.

RANK_REASON The cluster contains a research paper detailing a new method for improving LLM capabilities. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• GRPO
• Hugging Face
• LLMs
• LongZero
• reinforcement learning

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →