LongZero teaches LLMs long-context reasoning from zero data
TL;DR
- LongZero trains LLMs for long-context reasoning using only self-generated data, with no human annotation or external corpora required.
- The model first synthesizes a question, answer, and supporting evidence in compact form, then expands that evidence into a long multi-document context.
- Tested on two LLMs across five long-context reasoning benchmarks, the method outperforms RL baselines trained on data synthesized from external corpora.
Reinforcement learning can lift an LLM's long-context reasoning without any external training data, according to LongZero, a new preprint from Zhennan Wan and Jianfei Chen. The authors note the usual pipeline depends on human annotation that is expensive, naturally occurring data that are 'too scarce to sustain RL,' or synthesis built on outside corpora and stronger teacher models.
Instead, LongZero has the model first 'synthesize the question and answer with its supporting evidence in a compact structured form,' then separately expand that evidence into a long multi-document context. It combines self-written multi-hop reasoning with self-written multi-round retrieval, and trains with GRPO.
The authors evaluate on two LLMs across five long-context reasoning benchmarks and compare against RL on data synthesized from external corpora. Their result: 'LongZero can generalize to real tasks, and outperforms both baselines, suggesting a viable and promising path to improving long-context reasoning of LLMs.' The abstract names neither the base models nor the benchmarks nor any per-task scores. Two researchers we follow surfaced the preprint shortly after it went live.
Shared on Bluesky by 2 AI experts
Originally reported by arxiv.org
Read the original article →Original headline: LongZero: Enhancing Long-Context Reasoning in LLMs from Zero Data