Elamur Research⌕
☰
home / papers / 2025 / Pangu Embedded: An Efficient Dual-system LLM Reasoner with Metacognition

Pangu Embedded: An Efficient Dual-system LLM Reasoner with Metacognition

Paper · 2025 · Hanting Chen, Yasheng Wang, Kai Han et al.

Summary

Reasoning-focused LLMs such as DeepSeek-R1 are expensive to deploy because they are very large and generate long chains of thought, sometimes tens of thousands of tokens per problem. The Huawei Pangu Team addresses this with Pangu Embedded, a 7B-parameter reasoner trained on Ascend NPUs and aimed at edge use. Training proceeds in two stages. First, a base reasoner is built through iterative distillation from a teacher model. In each round, training samples are chosen by a complexity score that depends on the student's current ability, and checkpoints are merged between rounds. Reinforcement learning on Ascend clusters follows, guided by the Multi-source Adaptive Reward System (MARS) and a scheduler designed to tolerate latency. Second, the model learns to work in a quick "fast" mode or a deliberate "slow" mode. Users can switch modes by hand, or the model can pick one based on query difficulty. On AIME 2024, GPQA and LiveCodeBench, the model beats Qwen3-8B and GLM4-9B. This suggests that compact, deployable reasoners can reach top-tier accuracy within one unified model.

Key points

  • Base model: It shares its tokenizer (153,376-token vocabulary) and pre-training data with Pangu Ultra. Pre-training ran in three phases: general, reasoning and annealing.
  • Data pool: An LLM annotated each sample with category, question type, answer verifiability and several difficulty dimensions. Near-duplicates were removed with n-gram MinHash-LSH, and the final set was chosen with the compression-based ZIP algorithm to maximize diversity.
  • Data difficulty pilot (AIME 2024):
    • Balanced easy/hard mix: 50.42%
    • Mostly hard: 48.75%
    • Mostly easy: 45.42%
    • No selection: 43.33%
  • RL infrastructure: MARS combines rule-based metrics with lightweight LLM judges to score math, code and general tasks. The scheduler pairs stale synchronous parallelism with prioritized data queues.
  • Lessons and extras: The authors find that SFT distillation works best with medium or easier data, while RL benefits from harder, broader data. The model also includes a mechanism that detects and repairs repetitive output.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksComputer Vision and Pattern Recognition · 2,869 works

🧭 Topics 1Research subjects and languages the work is about.Language models & pretraining · 3,320 works

Cite

Cite as Hanting Chen, Yasheng Wang, Kai Han, Dong Li, Lin Li, Zhenni Bi, Jinpeng Li, Haoyu Wang, Fei Mi, Mingjian Zhu, Bin Wang, Kaikai Song, Yifei Fu, Xu He, Yu Luo, Chong Zhu, Quan He, Xueyu Wu, Wei He, Hailin Hu, Yehui Tang, Dacheng Tao, Xinghao Chen and Yunhe Wang (and Other Contributors). 2025. Pangu Embedded: An Efficient Dual-system LLM Reasoner with Metacognition. arXiv:2505.22375.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2505.22375 · arXiv 2505.22375

@misc{chen2025pangu,
  title = {{Pangu Embedded: An Efficient Dual-system LLM Reasoner with Metacognition}},
  author = {Hanting Chen, Yasheng Wang, Kai Han, Dong Li, Lin Li, Zhenni Bi, Jinpeng Li, Haoyu Wang, Fei Mi, Mingjian Zhu, Bin Wang, Kaikai Song, Yifei Fu, Xu He, Yu Luo, Chong Zhu, Quan He, Xueyu Wu, Wei He, Hailin Hu, Yehui Tang, Dacheng Tao, Xinghao Chen and Yunhe Wang (and Other Contributors)},
  year = {2025},
  doi = {10.48550/arxiv.2505.22375},
  eprint = {2505.22375},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2505.22375}
}

References

61 references, 32 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Huawei atlas 800t atlas-800t-a2. a2. https://e.huawei.com/cn/products/computing/ascend/ ↗
  2. Huawei atlas 800t a2 technical specifications. https://support.huawei.com/enterprise/en/doc/EDOC1100349804/2bf2c017/technical-specifications?idPath=23710424\| 251366513\|22892968\|252309113\|254184887. ↗
  3. Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.
  4. Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. find
  5. Akhiad Bercovich, Itay Levy, Izik Golan, Mohammad Dabbah, Ran El-Yaniv, Omri Puny, Ido Galil, Zach Moshe, Tomer Ronen, Najeeb Nabwani, et al. Llama-nemotron: Efficient reasoning models. arXiv preprint arXiv:2505.00949, 2025.
  6. Zhenni Bi, Kai Han, Chuanjian Liu, Yehui Tang, and Yunhe Wang. Forest-of-thought: Scaling test-time compute for enhancing llm reasoning. In ICML, 2025.
  7. Qianlong Du, Chengqing Zong, and Jiajun Zhang. Mods: Model-oriented data selection for instruction tuning. arXiv preprint arXiv:2311.15653, 2023. — [arXiv:2311.15653](https://arxiv.org/abs/2311.15653) find
  8. Zhiwei Fei, Xiaoyu Shen, Dawei Zhu, Fengzhe Zhou, Zhuo Han, Songyang Zhang, Kai Chen, Zongwen Shen, and Jidong Ge. Lawbench: Benchmarking legal knowledge of large language models. arXiv preprint arXiv:2309.16289, 2023.
  9. Sicheng Feng, Gongfan Fang, Xinyin Ma, and Xinchao Wang. Efficient reasoning models: A survey. arXiv preprint arXiv:2504.10903, 2025.
  10. Aryo Pradipta Gema, Joshua Ong Jun Leang, Giwon Hong, Alessio Devoto, Alberto Carlo Maria Mancino, Rohit Saxena, Xuanli He, Yu Zhao, Xiaotang Du, Mohammad Reza Ghasemi Madani, et al. Are we done with mmlu? arXiv preprint arXiv:2406.04127, 2024. find
  11. Alexandros V Gerbessiotis and Leslie G Valiant. Direct bulk-synchronous parallel algorithms. Journal of parallel and distributed computing, 22(2):251–267, 1994. find
  12. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025.
  13. Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason Weston, and Yuandong Tian. Training large language models to reason in a continuous latent space. arXiv preprint arXiv:2412.06769, 2024. find
  14. Qirong Ho, James Cipar, Henggang Cui, Seunghak Lee, Jin Kyu Kim, Phillip B Gibbons, Garth A Gibson, Greg Ganger, and Eric P Xing. More effective distributed ml via a stale synchronous parallel parameter server. In Advances in Neural Information Processing Systems, volume 26, 2013. find
  15. Zhen Huang, Haoyang Zou, Xuefeng Li, Yixiu Liu, Yuxiang Zheng, Ethan Chern, Shijie Xia, Yiwei Qin, Weizhe Yuan, and Pengfei Liu. O1 replication journey–part 2: Surpassing o1-preview through simple distillation, big progress or bitter lesson? arXiv preprint arXiv:2411.16489, 2024. find
  16. Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card. arXiv preprint arXiv:2412.16720, 2024.
  17. Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Armando Solar-Lezama, Koushik Sen, and Ion Stoica. Livecodebench: Holistic and contamination free evaluation of large language models for code. arXiv preprint arXiv:2403.07974, 2024.
  18. Woosuk Kwon, Zhuohan Li, Lianmin Xie, Mu Zheng, Cody Youxiang Yu, Joseph E Gonzalez, Ion Stoica, and Hao Zhang. Efficient memory management for large language model serving with vllm. arXiv preprint arXiv:2309.06180, 2023.
  19. Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. In Proceedings of the 29th Symposium on Operating Systems Principles, pages 611–626, 2023.
  20. ByteDance AI Lab. Hybridflow: A flexible and efficient rlhf framework. In USENIX Symposium on Operating Systems Design and Implementation, 2023.
  21. Ming Li, Yong Zhang, Zhitao Li, Jiuhai Chen, Lichang Chen, Ning Cheng, Jianzong Wang, Tianyi Zhou, and Jing Xiao. From quantity to quality: Boosting LLM performance with self-guided data selection for instruction tuning. In Kevin Duh, Helena Gomez, and Steven Bethard, editors, Proceedings of the 202…
  22. Tianle Li, Wei-Lin Chiang, Evan Frick, Lisa Dunlap, Banghua Zhu, Joseph E. Gonzalez, and Ion Stoica. From live data to high-quality benchmarks: The arena-hard pipeline, April 2024. find
  23. Yuetai Li, Xiang Yue, Zhangchen Xu, Fengqing Jiang, Luyao Niu, Bill Yuchen Lin, Bhaskar Ramasubramanian, and Radha Poovendran. Small models struggle to learn from strong reasoners. arXiv preprint arXiv:2502.12143, 2025. find
  24. Eric Liang, Richard Liaw, Robert Nishihara, Philipp Moritz, Roy Fox, Ken Goldberg, Joseph E Gonzalez, Michael I Jordan, and Ion Stoica. Rllib: Abstractions for distributed reinforcement learning. International Conference on Machine Learning, 2018. find
  25. Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s verify step by step. In The Twelfth International Conference on Learning Representations, 2023. — Let's Verify Step by Step
  26. Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437, 2024. find
  27. Wei Liu, Weihao Zeng, Keqing He, Yong Jiang, and Junxian He. What makes good data for alignment? a comprehensive study of automatic data selection in instruction tuning. In The Twelfth International Conference on Learning Representations.
  28. Haotian Luo, Li Shen, Haiying He, Yibo Wang, Shiwei Liu, Wei Li, Naiqiang Tan, Xiaochun Cao, and Dacheng Tao. O1-pruner: Length-harmonizing fine-tuning for o1-like reasoning pruning. arXiv preprint arXiv:2501.12570, 2025. find
  29. MAA. Codeforces. American Invitational Mathematics Examination - AIME 2024, 2024. https://maa.org/math-competitions/american-invitational-mathematics-examination-aime. ↗
  30. Lucie Charlotte Magister, Jonathan Mallinson, Jakub Dominik Adamek, Eric Malmi, and Aliaksei Severyn. Teaching small language models to reason. In The 61st Annual Meeting Of The Association For Computational Linguistics, 2023.
  31. Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès, and Tatsunori Hashimoto. s1: Simple test-time scaling. arXiv preprint arXiv:2501.19393, 2025.
  32. OpenAI. Introducing openai o1, 2024. find
  33. Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730–27744, 2022.
  34. David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R Bowman. Gpqa: A graduate-level google-proof q&a benchmark. In First Conference on Language Modeling, 2024.
  35. Mengjie Ren, Boxi Cao, Hongyu Lin, Cao Liu, Xianpei Han, Ke Zeng, Guanglu Wan, Xunliang Cai, and Le Sun. Learning or Self-aligning? Rethinking Instruction Fine-tuning, August 2024. arXiv:2402.18243 [cs]. find
  36. Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaoqing Ellen Tan, Yossi Adi, Jingyu Liu, Romain Sauvestre, Tal Remez, et al. Code llama: Open foundation models for code. arXiv preprint arXiv:2308.12950, 2023.
  37. John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. find
  38. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024.
  39. Sharath Turuvekere Sreenivas, Saurav Muralidharan, Raviraj Joshi, Marcin Chochowski, Ameya Sunil Mahabaleshwarkar, Gerald Shen, Jiaqi Zeng, Zijia Chen, Yoshi Suhara, Shizhe Diao, et al. Llm pruning and distillation in practice: The minitron approach. arXiv preprint arXiv:2408.11796, 2024.
  40. Hanshi Sun, Momin Haider, Ruiqi Zhang, Huitao Yang, Jiahao Qiu, Ming Yin, Mengdi Wang, Peter Bartlett, and Andrea Zanette. Fast best-of-n decoding via speculative rejection. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. find
  41. Chenxia Tang, Jianchun Liu, Hongli Xu, and Liusheng Huang. Top-nσ: Not all logits are you need, 2024. find
  42. Qwen Team. Qwq: Reflect deeply on the boundaries of the unknown. https://qwenlm.github.io/blog/qwq-32b-preview/. Accessed 28-11-2024. ↗
  43. Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023.
  44. Songjun Tu, Jiahao Lin, Qichao Zhang, Xiangyu Tian, Linjing Li, Xiangyuan Lan, and Dongbin Zhao. Learning when to think: Shaping adaptive reasoning in r1-style models via multi-stage rl. arXiv preprint arXiv:2505.10832, 2025. — [arXiv:2505.10832](https://arxiv.org/abs/2505.10832) find
  45. Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, et al. A survey on large language model based autonomous agents. Frontiers of Computer Science, 18(6):186345, 2024.
  46. Yunhe Wang, Hanting Chen, Yehui Tang, Tianyu Guo, Kai Han, Ying Nie, Xutao Wang, Hailin Hu, Zheyuan Bai, Yun Wang, et al. Pangu-pi: Enhancing language model architectures via nonlinearity compensation. arXiv preprint arXiv:2312.17276, 2023.
  47. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837, 2022.
  48. Wenyi Xiao, Leilei Gan, Weilong Dai, Wanggui He, Ziwei Huang, Haoyuan Li, Fangxun Shu, Zhelun Yu, Peng Zhang, Hao Jiang, et al. Fast-slow thinking for large vision-language model reasoning. arXiv preprint arXiv:2504.18458, 2025. — [arXiv:2504.18458](https://arxiv.org/abs/2504.18458) find
  49. Xiaohan Xu, Ming Li, Chongyang Tao, Tao Shen, Reynold Cheng, Jinyang Li, Can Xu, Dacheng Tao, and Tianyi Zhou. A survey on knowledge distillation of large language models. arXiv preprint arXiv:2402.13116, 2024. — [arXiv:2402.13116](https://arxiv.org/abs/2402.13116) find
  50. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388, 2025.
  51. Chuanpeng Yang, Yao Zhu, Wang Lu, Yidong Wang, Qian Chen, Chenlong Gao, Bingjie Yan, and Yiqiang Chen. Survey on knowledge distillation for large language models: methods, evaluation, and application. ACM Transactions on Intelligent Systems and Technology, 2024. find
  52. Yixin Ye, Zhen Huang, Yang Xiao, Ethan Chern, Shijie Xia, and Pengfei Liu. Limo: Less is more for reasoning. arXiv preprint arXiv:2502.03387, 2025.
  53. Huifeng Yin, Yu Zhao, Minghao Wu, Xuanfan Ni, Bo Zeng, Hao Wang, Tianqi Shi, Liangying Shao, Chenyang Lyu, Longyue Wang, et al. Towards widening the distillation bottleneck for reasoning models. arXiv preprint arXiv:2503.01461, 2025.
  54. Mingjia Yin, Chuhan Wu, Yufei Wang, Hao Wang, Wei Guo, Yasheng Wang, Yong Liu, Ruiming Tang, Defu Lian, and Enhong Chen. Entropy law: The story behind data compression and llm performance. arXiv preprint arXiv:2407.06645, 2024. find
  55. Yichun Yin, Wenyong Huang, Kaikai Song, Yehui Tang, Xueyu Wu, Wei Guo, Peng Guo, Yaoyuan Wang, Xiaojun Meng, Yasheng Wang, et al. Pangu ultra: Pushing the limits of dense large language models on ascend npus. arXiv preprint arXiv:2504.07866, 2025.
  56. Ping Yu, Jing Xu, Jason Weston, and Ilia Kulikov. Distilling system 2 into system 1. arXiv preprint arXiv:2407.06023, 2024. find
  57. Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Tiantian Fan, Gaohong Liu, Lingjun Liu, Xin Liu, et al. Dapo: An open-source llm reinforcement learning system at scale. arXiv preprint arXiv:2503.14476, 2025.
  58. Weihao Zeng, Yuzhen Huang, Qian Liu, Wei Liu, Keqing He, Zejun Ma, and Junxian He. Simplerl-zoo: Investigating and taming zero reinforcement learning for open base models in the wild. arXiv preprint arXiv:2503.18892, 2025.
  59. Jiajie Zhang, Nianyi Lin, Lei Hou, Ling Feng, and Juanzi Li. Adaptthink: Reasoning models can learn when to think. arXiv preprint arXiv:2505.13417, 2025.
  60. Han Zhao, Haotian Wang, Yiping Peng, Sitong Zhao, Xiaoyu Tian, Shuaiting Chen, Yunjie Ji, and Xiangang Li. Am deepseek r1 distilled 1.4m, 2025. find
  61. Chunting Zhou, Pengfei Liu, Puxin Xu, Srini Iyer, Jiao Sun, Yuning Mao, Xuezhe Ma, Avia Efrat, Ping Yu, Lili Yu, Susan Zhang, Gargi Ghosh, Mike Lewis, Luke Zettlemoyer, and Omer Levy. LIMA: Less Is More for Alignment, May 2023. arXiv:2305.11206 [cs]. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/chen-2025-pangu-embedded-efficient-dual/