LongZero

🌐 Github | 📖 Paper | 🤗 Data

Introduction

LongZero lets an LLM improve its own long-context reasoning without any external data. The model first writes the question, its supporting evidence and the answer in a compact structured form, and only then expands the evidence into a long multi-document context. These self-written multi-hop reasoning tasks are combined with self-written multi-round retrieval tasks in the style of MRCR, and the model is trained on them with GRPO.

This dataset contains the complete training and validation data used in the paper. Each base model synthesizes and filters its own data: Qwen3-4B-Thinking-2507 and Qwen3-8B. The code is in the Github repository.

Data

You can download the data into the root of the Github repository:

hf download zhennan1/LongZero --repo-type dataset --local-dir . --include "data/*"

The data is organized as follows:

./
└── data/
    ├── qwen3_4b/
    │   ├── train.parquet  # 2,304 reasoning + 1,152 retrieval samples
    │   ├── val.parquet    # 256 reasoning + 128 retrieval samples
    │   └── examples.txt   # one readable validation example of each task
    └── qwen3_8b/
        └── ...

Each row follows the verl schema:

• data_source: custom_longcontextqa_synthesized for reasoning data, custom_mrcr_* for retrieval data.
• prompt: Chat messages given to the model. A reasoning prompt contains eight shuffled documents of one fictional world and a multi-hop question; a retrieval prompt is a long conversation of the model's own responses followed by a request for the k-th response to one request.
• ability, reward_model: Reward rule and ground truth. Reasoning ground truth is the answer and its accepted aliases; retrieval ground truth is the random prefix followed by the requested response.
• extra_info: Metadata. For reasoning data: question id, domain, number of hops, evidence documents, context length, and the number of correct answers during difficulty filtering (k_passes of k_total). For retrieval data: needle positions, requested index and random prefix.

The median context length is 19.4K tokens (Qwen3-4B-Thinking-2507) and 14.6K tokens (Qwen3-8B) for reasoning data, and 11.4K and 11.1K tokens for retrieval data, counted with each model's tokenizer.

All entities, facts, documents and conversations are fictional and written by the models. Names may coincidentally match real people or organizations, so the documents must not be treated as factual information. The retrieval requests use the formats and nouns of OpenAI MRCR; all responses are written by the base models.

Citation

If you find this project helpful, please cite it as follows:

@article{wan2026longzero,
  title={LongZero: Enhancing Long-Context Reasoning in LLMs from Zero Data},
  author={Zhennan Wan and Jianfei Chen},
  year={2026}
}