LongZero: Enhancing Long-Context Reasoning in LLMs from Zero Data

Forum topic · 小凯 · 2026-10-23

Summary

LongZero (arXiv:2610.30132, by Zhennan Wan and Jianfei Chen) is a zero-data reinforcement learning method that improves the long-context reasoning ability of large language models without any external data. RL is effective for long-context reasoning but requires training data that are long, diverse, challenging, and verifiable; such data are expensive to annotate, scarce in nature, and existing synthesis approaches depend on external corpora or stronger teacher models that become unavailable as the trained model improves. LongZero addresses this by first synthesizing questions, answers, and supporting evidence in a compact structured form, then expanding the evidence separately into long multi-document contexts. It combines self-generated multi-hop reasoning tasks with multi-round retrieval tasks and trains models using the GRPO algorithm. Experiments on two LLMs and five long-context reasoning benchmarks show that LongZero generalizes to real tasks and outperforms baselines built on externally synthesized corpus data, offering a practical and promising path toward better long-context reasoning in LLMs.

Paper Overview

Research Area: NLP Authors: Zhennan Wan, Jianfei Chen Published: 2026-10-21 arXiv: 2610.30132

Abstract

Reinforcement 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. Human annotation of such data is expensive, naturally occurring data are too scarce to sustain RL, and existing synthesis methods depend on external corpora or stronger teacher models, which become harder to obtain as the trained models get stronger. Recently, RL from zero data, in which a model learns from tasks it proposes itself, has achieved excellent performance, but existing methods target short-context tasks. Therefore, we propose LongZero, a method that lets an LLM improve its own long-context reasoning without any external data.

How It Works

LongZero first synthesizes questions, answers, and their supporting evidence in a compact structured form.
It then expands the evidence separately into long multi-document contexts.
It combines self-generated multi-hop reasoning tasks with multi-round retrieval tasks.
Models are trained with the GRPO algorithm.

Results

Experiments across two LLMs and five long-context reasoning benchmarks show that LongZero generalizes to real-world tasks and outperforms baselines that rely on data synthesized from external corpora, providing a viable and promising path for improving LLM long-context reasoning.