LongZero improves LLM long-context reasoning without external data
Original title: LongZero: Enhancing Long-Context Reasoning in LLMs from Zero Data
The method self-writes multi-hop and retrieval tasks, trains with GRPO, and beat external-corpus RL on five benchmarks.
Researchers propose LongZero, a reinforcement-learning method that lets a model improve long-context reasoning without external documents or a stronger teacher. The model first writes a question, answer, and compact evidence, then expands that evidence into a long multi-document context. Combined with self-written multi-round retrieval tasks and trained with GRPO, it generalized to real tasks and outperformed RL on externally synthesized data across two LLMs and five benchmarks.
Unconfirmed: Benchmark gains are reported by the authors and are not independently confirmed here.
- Existing zero-data RL methods target short-context tasks, the paper says.
- Evidence is synthesized in compact form before being expanded into long context.
- Training mixes multi-hop reasoning tasks with multi-round retrieval tasks.
- The paper compares LongZero with RL on data from external corpora.
From the original article
1 min read at arxiv.orgReinforcement learning (RL) is an effective way to strengthen the long-context reasoning of large language models (LLMs), but it needs training data that are long, diverse, challenging, and verifiable.
By Zhennan Wan · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.