Elamur Research⌕
☰
home / papers / 2026 / SpiralThinker: Latent Reasoning through an Iterative Process with Text–Latent Interleaving

SpiralThinker: Latent Reasoning through an Iterative Process with Text–Latent Interleaving

Paper · 2026 · Shengmin Piao and Sanghyun Park

Summary

SpiralThinker is a method for latent reasoning, where a language model reasons inside its hidden representations instead of writing out every step as text. The authors identify two gaps in earlier latent methods. First, latent tokens are usually handled in a single forward pass with no control over how they evolve. Second, there is no principled way to mix textual and latent steps. SpiralThinker addresses both. It first fine-tunes a model on explicit step-by-step reasoning. It then replaces alternating textual steps with blocks of <latent> tokens and refines those representations over several iterations without emitting extra tokens. A progressive alignment objective keeps each iteration's latent states consistent with the textual step they replace. Evaluated on GSM8K-Aug, ProsQA and StrategyQA (math, logic and commonsense), SpiralThinker outperforms prior latent reasoning baselines. The work suggests that supervised iterative refinement makes latent reasoning more stable while retaining some interpretability.

Key points

  • Training pipeline: Training has two stages.
    • An explicit stage fine-tunes the model on reasoning steps marked with <bot>/<eot> boundary tokens.
    • An implicit stage replaces odd- or even-positioned steps with N <latent> tokens enclosed in <bol>/<eol>.
    • A latent adapter maps the latent representations into the token-embedding space.
  • Optimal iteration counts: The best number of iterations depends on the dataset: K=5 for GSM8K-Aug, K=4 for ProsQA and K=3 for StrategyQA. The best number of latent tokens is also dataset-specific.
  • Alignment weighting: Three ways of weighting the alignment loss across iterations were compared.
    • Uniform weighting performed worst, especially on ProsQA and StrategyQA.
    • Weighting only the last iteration improved results.
    • A softmax-based schedule was generally best.
  • Probing the latent tokens: The authors matched each latent state to its nearest token embeddings on a GSM8K-Aug example.
    • Some latent tokens captured operands and the "*" operator.
    • Another tracked intermediate results that converged to the correct value across iterations.
    • Others signalled the end of the step (<eol>).
  • Limitations and availability: The number of iterations is fixed for every step, and interleaving is applied whether or not a step needs it. Adaptive versions of both are left for future work. Code is available at github.com/shengminp/SpiralThinker.

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

 Institutions

🎓Yonsei UniversitySeoul · South Korea17 works · 501 citations

 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.Reasoning · 1,426 works

Cite

Cite as Shengmin Piao and Sanghyun Park. 2026. SpiralThinker: Latent Reasoning through an Iterative Process with Text–Latent Interleaving. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.1605

@inproceedings{piao2026spiralthinker,
  title = {{SpiralThinker: Latent Reasoning through an Iterative Process with Text–Latent Interleaving}},
  author = {Shengmin Piao and Sanghyun Park},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.1605},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.1605}
}

References

33 references, 13 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Xinghao Chen, Anhao Zhao, Heming Xia, Xuan Lu, Hanlin Wang, Yanjun Chen, Wei Zhang, Jian Wang, Wenjie Li, and Xiaoyu Shen. 2025a. Reasoning beyond language: A comprehensive survey on latent chain-of-thought reasoning. arXiv preprint arXiv:2505.16782. — [arXiv:2505.16782](https://arxiv.org/abs/2505.16782) find
  2. Xingyu Chen, Jiahao Xu, Tian Liang, Zhiwei He, Jianhui Pang, Dian Yu, Linfeng Song, Qiuzhi Liu, Mengfei Zhou, Zhuosheng Zhang, Rui Wang, Zhaopeng Tu, Haitao Mi, and Dong Yu. 2025b. Do NOT think that much for 2+3=? on the overthinking of long reasoning models. In Forty-second International Conference… find
  3. Yilong Chen, Junyuan Shang, Zhenyu Zhang, Yanxi Xie, Jiawei Sheng, Tingwen Liu, Shuohuan Wang, Yu Sun, Hua Wu, and Haifeng Wang. 2025c. Inner thinking transformer: Leveraging dynamic depth scaling to foster adaptive internal thinking. In Proceedings of the 63rd Annual Meeting of the Association for…
  4. Jeffrey Cheng and Benjamin Van Durme. 2024. Compressed chain of thought: Efficient reasoning through dense representations. arXiv preprint arXiv:2412.13171. find
  5. Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, and 1 others. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168.
  6. Mostafa Dehghani, Stephan Gouws, Oriol Vinyals, Jakob Uszkoreit, and Lukasz Kaiser. 2019. Universal transformers. In International Conference on Learning Representations. find
  7. Yuntian Deng, Yejin Choi, and Stuart Shieber. 2024. From explicit cot to implicit cot: Learning to internalize cot step by step. arXiv preprint arXiv:2405.14838. find
  8. Yuntian Deng, Kiran Prasad, Roland Fernandez, Paul Smolensky, Vishrav Chaudhary, and Stuart Shieber. 2023. Implicit chain of thought reasoning via knowledge distillation. arXiv preprint arXiv:2311.01460. find
  9. Jonas Geiping, Sean Michael McLeish, Neel Jain, John Kirchenbauer, Siddharth Singh, Brian R. Bartoldson, Bhavya Kailkhura, Abhinav Bhatele, and Tom Goldstein. 2025. Scaling up test-time compute with latent reasoning: A recurrent depth approach. In ES-FoMo III: 3rd Workshop on Efficient Systems for F… find
  10. Mor Geva, Daniel Khashabi, Elad Segal, Tushar Khot, Dan Roth, and Jonathan Berant. 2021. Did aristotle use a laptop? a question answering benchmark with implicit reasoning strategies. Transactions of the Association for Computational Linguistics, 9:346– 361. find
  11. Ian Goodfellow, Yoshua Bengio, Aaron Courville, and Yoshua Bengio. 2016. Deep learning, volume 1. MIT press Cambridge. find
  12. Sachin Goyal, Ziwei Ji, Ankit Singh Rawat, Aditya Krishna Menon, Sanjiv Kumar, and Vaishnavh Nagarajan. 2024. Think before you speak: Training language models with pause tokens. In The Twelfth International Conference on Learning Representations. find
  13. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783.
  14. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, and 1 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948.
  15. Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason E Weston, and Yuandong Tian. 2024. Training large language model to reason in a continuous latent space. find
  16. David Herel and Tomas Mikolov. 2024. Thinking tokens for language modeling. arXiv preprint arXiv:2405.08644. — [arXiv:2405.08644](https://arxiv.org/abs/2405.08644) find
  17. Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations.
  18. Jindong Li, Yali Fu, Li Fan, Jiahong Liu, Yao Shu, Chengwei Qin, Menglin Yang, Irwin King, and Rex Ying. 2025. Implicit reasoning in large language models: A comprehensive survey. arXiv preprint arXiv:2509.02350.
  19. Tengxiao Liu, Qipeng Guo, Xiangkun Hu, Cheng Jiayang, Yue Zhang, Xipeng Qiu, and Zheng Zhang. 2024. Can language models learn to skip steps? Advances in Neural Information Processing Systems, 37:45359–45385. find
  20. Jacob Pfau, William Merrill, and Samuel R. Bowman. 2024. Let’s think dot by dot: Hidden computation in transformer language models. In First Conference on Language Modeling. find
  21. Nikunj Saunshi, Nishanth Dikkala, Zhiyuan Li, Sanjiv Kumar, and Sashank J. Reddi. 2025. Reasoning with latent thoughts: On the power of looped transformers. In The Thirteenth International Conference on Learning Representations. find
  22. Zhenyi Shen, Hanqi Yan, Linhai Zhang, Zhanghao Hu, Yali Du, and Yulan He. 2025. Codi: Compressing chain-of-thought into continuous space via self-distillation. arXiv preprint arXiv:2502.21074.
  23. Charlie Victor Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. 2025. Scaling LLM test-time compute optimally can be more effective than scaling parameters for reasoning. In The Thirteenth International Conference on Learning Representations. find
  24. DiJia Su, Hanlin Zhu, Yingchen Xu, Jiantao Jiao, Yuandong Tian, and Qinqing Zheng. 2025. Token assorted: Mixing latent and text tokens for improved language model reasoning. In Forty-second International Conference on Machine Learning. find
  25. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30.
  26. Xuezhi Wang and Denny Zhou. 2024. Chain-of-thought reasoning without prompting. Advances in Neural Information Processing Systems, 37:66383–66409. find
  27. Yue Wang, Qiuzhi Liu, Jiahao Xu, Tian Liang, Xingyu Chen, Zhiwei He, Linfeng Song, Dian Yu, Juntao Li, Zhuosheng Zhang, and 1 others. 2025. Thoughts are all over the place: On the underthinking of o1-like llms. arXiv preprint arXiv:2501.18585. find
  28. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian ichter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824– 24837.
  29. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pier-ric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, and 1 others. 2019. Huggingface’s transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771.
  30. Yige Xu, Xu Guo, Zhiwei Zeng, and Chunyan Miao. 2025. SoftCoT: Soft chain-of-thought for efficient reasoning with LLMs. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 23336– 23351, Vienna, Austria. Association for Computation…
  31. Qifan Yu, Zhenyu He, Sijie Li, Xun Zhou, Jun Zhang, Jingjing Xu, and Di He. 2025. Enhancing autoregressive chain-of-thought through loop-aligned reasoning. arXiv preprint arXiv:2502.08482. — [arXiv:2502.08482](https://arxiv.org/abs/2502.08482) find
  32. Zhen Zhang, Xuehai He, Weixiang Yan, Ao Shen, Chenyang Zhao, Shuohang Wang, Yelong Shen, and Xin Eric Wang. 2025. Soft thinking: Unlocking the reasoning potential of llms in continuous concept space. arXiv preprint arXiv:2505.15778.
  33. Rui-Jie Zhu, Tianhao Peng, Tianhao Cheng, Xingwei Qu, Jinfa Huang, Dawei Zhu, Hao Wang, Kaiwen Xue, Xuanliang Zhang, Yong Shan, and 1 others. 2025. A survey on latent reasoning. arXiv preprint arXiv:2507.06203.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/piao-2026-spiralthinker-latent-reasoning-through-iterative/