--- license: mit ---

👉 LightThinker 👈

LightThinker Family: From Thinking Compression to Adaptive Memory Management [![Awesome](https://awesome.re/badge.svg)](https://github.com/zjunlp/LightThinker) [![License: MIT](https://img.shields.io/badge/License-MIT-green.svg)](https://opensource.org/licenses/MIT) ![](https://img.shields.io/github/last-commit/zjunlp/LightThinker?color=green)

Github • 📄 LightThinker arXiv • 📄 LightThinker++ arXiv • 🤗 Hugging Face

## Table of Contents - 🔔 [News](#news) - 👀 [Overview](#overview) - 🤗 [Resources](#resources) - 📁 [Repository Structure](#repository-structure) - 🚀 [LightThinker++ — General Reasoning](#lightthinker--general-reasoning) - 🌐 [LightThinker++ — Agentic Reasoning](#lightthinker--agentic-reasoning) - 💡 [LightThinker](#lightthinker) - 🎁 [Acknowledgement](#acknowledgement) - 🚩 [Citation](#citation) ## 🔔 News - **[2026-03]** We release a new paper: "[LightThinker++: From Reasoning Compression to Memory Management](http://arxiv.org/abs/2604.03679)". - **[2025-08]** Our paper "[LightThinker: Thinking Step-by-Step Compression](https://arxiv.org/abs/2502.15589)" has been accepted to EMNLP 2025. - **[2025-02]** We release "[LightThinker: Thinking Step-by-Step Compression](https://arxiv.org/abs/2502.15589)". ## 👀 Overview LLMs can solve increasingly complex reasoning tasks, but long thought traces make inference expensive because the model must retain and attend to large contexts. The **LightThinker family** addresses this by compressing or actively managing intermediate reasoning states. - **LightThinker** compresses intermediate thoughts into compact gist-token representations. The implementation and AnLLM baseline are preserved under [`lightthinker_v1/`](./lightthinker_v1/). - **LightThinker++** introduces explicit adaptive memory management. The model can use memory primitives such as `commit`, `expand`, `fold`, and `final_answer` to control its scratchpad while reasoning. This top-level README is a quick-start guide. Detailed model links, data links, environment variables, script parameters, and evaluation options live in the module README files. ## 🤗 Resources - **LightThinker++ General Reasoning:** models, SFT data, and full configuration are in [`general_reasoning/README.md`](./general_reasoning/README.md). - **LightThinker++ Agentic Reasoning:** models, SFT data, and full configuration are in [`agentic_reasoning/README.md`](./agentic_reasoning/README.md). - **LightThinker:** models, data archive, and full configuration are in [`lightthinker_v1/README.md`](./lightthinker_v1/README.md). ## 📁 Repository Structure ``` LightThinker/ ├── general_reasoning/ # LightThinker++ for math and general reasoning tasks ├── agentic_reasoning/ # LightThinker++ for agentic deep-research tasks ├── lightthinker_v1/ # LightThinker and AnLLM baseline code ├── assets/ # README assets └── README.md ``` | Method | Folder | Full guide | |--------|--------|------------| | LightThinker++ for general reasoning | [`general_reasoning/`](./general_reasoning/) | [`general_reasoning/README.md`](./general_reasoning/README.md) | | LightThinker++ for agentic reasoning | [`agentic_reasoning/`](./agentic_reasoning/) | [`agentic_reasoning/README.md`](./agentic_reasoning/README.md) | | LightThinker | [`lightthinker_v1/`](./lightthinker_v1/) | [`lightthinker_v1/README.md`](./lightthinker_v1/README.md) | ## 🚀 LightThinker++ — General Reasoning Quick start for math and general reasoning tasks such as GSM8K, MMLU, GPQA, and BBH: ```bash cd general_reasoning conda env create -f environment.yml conda activate lt_plus_general_reasoning cp .env.example ../.env # Generate synthetic trajectories after setting an input JSONL dataset. DATASET_PATH=/path/to/input.jsonl bash scripts/synthetic.sh # Run inference with a trained or downloaded model. MODEL=/path/to/model bash scripts/infer.sh 0 gsm8k ``` See [`general_reasoning/README.md`](./general_reasoning/README.md) for configuration, datasets, and full arguments. ## 🌐 LightThinker++ — Agentic Reasoning Quick start for long-horizon deep-research and multi-hop QA tasks: ```bash cd agentic_reasoning bash setup.sh cp .env.synthetic.example .env # Run synthetic data generation with the included sample dataset. bash scripts/synthetic.sh ``` For inference, use `cp .env.infer.example .env`, configure your model endpoint, and run `bash scripts/inference.sh`. See [`agentic_reasoning/README.md`](./agentic_reasoning/README.md) for configuration, datasets, and full arguments. ## 💡 LightThinker Quick start for the LightThinker implementation: ```bash cd lightthinker_v1 conda create -n lightthinker python=3.9 -y conda activate lightthinker pip install -r requirements.txt # Train and run inference. bash train.sh bash inference.sh ``` See [`lightthinker_v1/README.md`](./lightthinker_v1/README.md) for configuration, datasets, and full arguments. ## 🎁 Acknowledgement We use the [`ms-swift`](https://github.com/modelscope/ms-swift) framework for full-parameter SFT of the LightThinker++ models. The general reasoning evaluation also builds on ideas from [TokenSkip](https://github.com/carriex/TokenSkip). The LightThinker evaluation includes baseline code inspired by H2O from [Meta-llama](https://github.com/meta-llama/llama-cookbook) and SepLLM from [HKUDS](https://github.com/HKUDS/SepLLM). ## 🚩 Citation If this work is helpful, please kindly cite: ```bibtex @article{lightthinker++, author = {Yuqi Zhu and Jintian Zhang and Zhenjie Wan and Yujie Luo and Shuofei Qiao and Zhengke Gui and Da Zheng and Lei Liang and Huajun Chen and Ningyu Zhang}, title = {LightThinker++: From Reasoning Compression to Memory Management}, journal = {CoRR}, volume = {abs/2604.03679}, year = {2026}, url = {https://doi.org/10.48550/arXiv.2604.03679}, doi = {10.48550/ARXIV.2604.03679}, eprinttype = {arXiv}, eprint = {2604.03679}, timestamp = {Fri, 08 May 2026 07:40:46 +0200}, biburl = {https://dblp.org/rec/journals/corr/abs-2604-03679.bib}, bibsource = {dblp computer science bibliography, https://dblp.org} } @inproceedings{lightthinker, author = {Jintian Zhang and Yuqi Zhu and Mengshu Sun and Yujie Luo and Shuofei Qiao and Lun Du and Da Zheng and Huajun Chen and Ningyu Zhang}, editor = {Christos Christodoulopoulos and Tanmoy Chakraborty and Carolyn Rose and Violet Peng}, title = {LightThinker: Thinking Step-by-Step Compression}, booktitle = {Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, {EMNLP} 2025, Suzhou, China, November 4-9, 2025}, pages = {13307--13328}, publisher = {Association for Computational Linguistics}, year = {2025}, url = {https://doi.org/10.18653/v1/2025.emnlp-main.673}, doi = {10.18653/V1/2025.EMNLP-MAIN.673}, timestamp = {Mon, 02 Feb 2026 09:39:37 +0100}, biburl = {https://dblp.org/rec/conf/emnlp/ZhangZSLQDZCZ25.bib}, bibsource = {dblp computer science bibliography, https://dblp.org} } ```