Views
No views yet
lightthinker_v1/.commit, expand, fold, and final_answer to control its scratchpad while reasoning.general_reasoning/README.md.agentic_reasoning/README.md.lightthinker_v1/README.md.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/README.md |
| LightThinker++ for agentic reasoning | agentic_reasoning/ | agentic_reasoning/README.md |
| LightThinker | lightthinker_v1/ | lightthinker_v1/README.md |
1cd general_reasoning
2conda env create -f environment.yml
3conda activate lt_plus_general_reasoning
4cp .env.example ../.env
5
6# Generate synthetic trajectories after setting an input JSONL dataset.
7DATASET_PATH=/path/to/input.jsonl bash scripts/synthetic.sh
8
9# Run inference with a trained or downloaded model.
10MODEL=/path/to/model bash scripts/infer.sh 0 gsm8kgeneral_reasoning/README.md for configuration, datasets, and full arguments.1cd agentic_reasoning
2bash setup.sh
3cp .env.synthetic.example .env
4
5# Run synthetic data generation with the included sample dataset.
6bash scripts/synthetic.shcp .env.infer.example .env, configure your model endpoint, and run bash scripts/inference.sh.agentic_reasoning/README.md for configuration, datasets, and full arguments.1cd lightthinker_v1
2conda create -n lightthinker python=3.9 -y
3conda activate lightthinker
4pip install -r requirements.txt
5
6# Train and run inference.
7bash train.sh
8bash inference.shlightthinker_v1/README.md for configuration, datasets, and full arguments.ms-swift framework for full-parameter SFT of the LightThinker++ models. The general reasoning evaluation also builds on ideas from TokenSkip. The LightThinker evaluation includes baseline code inspired by H2O from Meta-llama and SepLLM from HKUDS.1@article{lightthinker++,
2 author = {Yuqi Zhu and
3 Jintian Zhang and
4 Zhenjie Wan and
5 Yujie Luo and
6 Shuofei Qiao and
7 Zhengke Gui and
8 Da Zheng and
9 Lei Liang and
10 Huajun Chen and
11 Ningyu Zhang},
12 title = {LightThinker++: From Reasoning Compression to Memory Management},
13 journal = {CoRR},
14 volume = {abs/2604.03679},
15 year = {2026},
16 url = {https://doi.org/10.48550/arXiv.2604.03679},
17 doi = {10.48550/ARXIV.2604.03679},
18 eprinttype = {arXiv},
19 eprint = {2604.03679},
20 timestamp = {Fri, 08 May 2026 07:40:46 +0200},
21 biburl = {https://dblp.org/rec/journals/corr/abs-2604-03679.bib},
22 bibsource = {dblp computer science bibliography, https://dblp.org}
23}
24
25@inproceedings{lightthinker,
26 author = {Jintian Zhang and
27 Yuqi Zhu and
28 Mengshu Sun and
29 Yujie Luo and
30 Shuofei Qiao and
31 Lun Du and
32 Da Zheng and
33 Huajun Chen and
34 Ningyu Zhang},
35 editor = {Christos Christodoulopoulos and
36 Tanmoy Chakraborty and
37 Carolyn Rose and
38 Violet Peng},
39 title = {LightThinker: Thinking Step-by-Step Compression},
40 booktitle = {Proceedings of the 2025 Conference on Empirical Methods in Natural
41 Language Processing, {EMNLP} 2025, Suzhou, China, November 4-9, 2025},
42 pages = {13307--13328},
43 publisher = {Association for Computational Linguistics},
44 year = {2025},
45 url = {https://doi.org/10.18653/v1/2025.emnlp-main.673},
46 doi = {10.18653/V1/2025.EMNLP-MAIN.673},
47 timestamp = {Mon, 02 Feb 2026 09:39:37 +0100},
48 biburl = {https://dblp.org/rec/conf/emnlp/ZhangZSLQDZCZ25.bib},
49 bibsource = {dblp computer science bibliography, https://dblp.org}
50}