InftyThink: Breaking the Length Limits of Long-Context Reasoning in Large Language Models

Yuchen Yan^1,2,*, Yongliang Shen^1,†, Yang Liu ², Jin Jiang ^2,3,
Mengdi zhang ², Jian Shao^1,†, Yueting Zhuang¹

¹Zhejiang University ²Meituan Group ³Peking university
Preprint. Under review.
^*Contribution during internship at Meituan Group, ^†Corresponding Author

🤗 Dataset | Arxiv | 📑 WebPage

News 🔥🔥

• 2025.07.12: We release our re-implemented dataset.
• 2025.05.24: We release our HomePage and Code examples.
• 2025.03.09: We release our paper.

Overview 🦾🦾

In this paper, we propose a fundamentally different approach to long-context reasoning. Rather than viewing reasoning as a single extended process, we introduce InftyThink, a novel paradigm that divides complex reasoning into multiple interrelated short reasoning segments. Each segment remains within a computationally efficient context length while maintaining the coherent flow of thought across iterations. This approach draws inspiration from human cognitive processes, where complex problem-solving frequently involves breaking problems into manageable parts and summarizing intermediate progress.

Our contributions can be summarized as follows:

• We introduce InftyThink, which transforms monolithic long-form reasoning into iterative reasoning with summarization, mimicking human working memory patterns and reducing the quadratic computational complexity of transformer-based models to a more manageable form.
• We develop a technique to reconstruct existing long-context reasoning datasets (demonstrated on OpenR1-Math) into our iterative format, preserving reasoning quality while enabling more efficient computation without specialized architectures.
• Across multiple model architectures, our approach achieves significant improvements while substantially reducing computational costs, challenging the assumed trade-off between reasoning depth and efficiency.

QuickStart 🎯🎯

InftyThink
├── data_preprocess  # Generate InftyThink-style data
├── inference        # An example for using InftyThink-style models
├── docs
└── readme.md

Generate InftyThink-style Data

Step 1: Thinking process segmentation

cd data_preprocess
python3 segmentation.py --dataset_name open-r1/OpenR1-Math-220k \
    --tokenizer Qwen/Qwen2.5-Math-7B \
    --eta 4096

Step 2: Generate summary and form InftyThink-style data

cd data_preprocess
python3 generate_data.py --model meta-llama/Llama-3.3-70B-Instruct

After code finished, InftyThink-style data is available.

Inference

We provide an example for InftyThink-style reasoning, after your SFT on InftyThink-style data, feel free to try it!

cd inference
python3 infer_single.py

Citation

If you find our work helpful, feel free to give us a cite.

@misc{yan2025inftythink,
    title={InftyThink: Breaking the Length Limits of Long-Context Reasoning in Large Language Models}, 
    author={Yuchen Yan and Yongliang Shen and Yang Liu and Jin Jiang and Mengdi Zhang and Jian Shao and Yueting Zhuang},
    year={2025},
    eprint={2503.06692},
    archivePrefix={arXiv},
    primaryClass={cs.CL},
    url={https://arxiv.org/abs/2503.06692}, 
}

Contact Us

If you have any questions, please contact us by email: yanyuchen@zju.edu.cn