---
title: "Learning Shortcut Models for Efficient Recursive Reasoning (2026)"
description: "The method therefore offers explicit control over the compute–accuracy trade-off. 25% accuracy. 24 references."
authors: ["Shiv Shankar"]
year: 2026
venue: "Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop)"
source: "https://aclanthology.org/2026.acl-srw.120/"
doi: "10.18653/v1/2026.acl-srw.120"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/shankar-2026-learning-shortcut-models-efficient-recursive/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/shankar-2026-learning-shortcut-models-efficient-recursive/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/34/) / Learning Shortcut Models for Efficient Recursive Reasoning

# Learning Shortcut Models for Efficient Recursive Reasoning

**Authors:** [Shiv Shankar](https://md.research.elamur.ai/free/en/authors/shiv-shankar/) · **Year:** 2026 · **Venue:** [Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop)](https://md.research.elamur.ai/free/en/venues/acl/) ([ACL 2026](https://md.research.elamur.ai/free/en/venues/acl/2026/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2026.acl-srw.120/) · [PDF](https://aclanthology.org/2026.acl-srw.120.pdf) · [DOI 10.18653/v1/2026.acl-srw.120](https://doi.org/10.18653/v1/2026.acl-srw.120) · [OpenAlex](https://openalex.org/W7166818622)

**Citations:** 0 citations (OpenAlex, 2026-10-14)

## Summary
Recursive latent reasoners such as HRM and TRM refine hidden states through many small sequential passes. Their inference cannot be parallelised, they receive weak supervision for early steps, and they cannot adjust their compute budget at test time. This paper views the iterative refinement steps of models in the TRM/HRM family as a discrete dynamical system unfolding over latent time. It distils that system into a student "shortcut" model, which learns a scale-conditioned jump operator that predicts where the latent state will be after h teacher steps.

Training combines four objectives:
- matching teacher trajectories;
- self-consistency across jump sizes, so one large jump agrees with two smaller ones;
- stepwise improvement, which pushes each jump to raise the gold-answer likelihood;
- a repair term, under which a denoising variant of the base model pulls imperfect jump states back toward valid latent regions.

On ARC-AGI, a shortcut model distilled from a pretrained 7M-parameter TRM stays close to TRM accuracy while using half the inference FLOPs. Accuracy degrades gradually as jumps grow larger. The method therefore offers explicit control over the compute–accuracy trade-off.

## Key points
- ARC-1 / ARC-2 scores for the shortcut model (7.2M params) at 1-, 2-, 4- and 8-hop jumps, corresponding to 1×, 0.5×, 0.25× and 0.125× TRM inference FLOPs:

  | Model | ARC-1 | ARC-2 |
  |---|---|---|
  | Shortcut, 1-hop | 40.8 | 5.8 |
  | Shortcut, 2-hop | 39.7 | 3.0 |
  | Shortcut, 4-hop | 34.2 | 2.4 |
  | Shortcut, 8-hop | 28.6 | 1.7 |
  | TRM baseline (7M) | 40.4 | 3.3 |
  | DIS baseline (7M) | 41.3 | 6.0 |

- Large CoT LLMs are listed for reference. DeepSeek R1 scores 15.8 / 1.3, and the bespoke Grok-4 system scores 79.6 / 29.4.
- Without repair, the student's latent outputs stop tracking the teacher beyond jumps of about 4–5 steps. With repair, correlation with the teacher trajectory stays high.
- Per-step improvement supervision speeds up learning substantially: the model reaches 25% accuracy in about 50k training steps, versus roughly 300k without it.
- The author lists several limitations:
  - jump sizes are fixed globally rather than adapted per instance;
  - training adds overhead;
  - the extra shortcut network increases memory use;
  - experiments cover only ARC-AGI.

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

📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 16,039 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 15,891 works

🧭 Topics (1): [🧮 Reasoning](https://md.research.elamur.ai/free/en/topics/reasoning/) · 1,426 works

## Cite

Cite as: Shiv Shankar. 2026. Learning Shortcut Models for Efficient Recursive Reasoning. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop).

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@inproceedings{shankar2026learning,
  title = {{Learning Shortcut Models for Efficient Recursive Reasoning}},
  author = {Shiv Shankar},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-srw.120},
  url = {https://doi.org/10.18653/v1/2026.acl-srw.120}
}
```

- [← previous Learning reduplicative templates as hidden structures: the case of reduplication-phonology interactions](https://md.research.elamur.ai/free/en/papers/wang-2026-learning-reduplicative-templates-as-hidden/)
- [next → Learning Stress in Arabic Low-Resource Settings](https://md.research.elamur.ai/free/en/papers/qaddoumi-2026-learning-stress-arabic-low-resource/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/34/)

## References

24 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Alexia Jolicoeur-Martineau. 2025. Less is more: Recursive reasoning with tiny networks. Preprint, arXiv:2510.04871. [find](https://scholar.google.com/scholar?q=Alexia+Jolicoeur%2DMartineau%2E+2025%2E+Less+is+more%3A+Recursive+reasoning+with+tiny+networks%2E+Preprint%2C+arXiv%3A2510%2E04871%2E)
2. Arip Asadulaev, Rayan Banerjee, Fakhri Karray, and Martin Takac. 2025. Deep improvement supervision. arXiv preprint arXiv:2511.16886. — [arXiv:2511.16886](https://arxiv.org/abs/2511.16886) [find](https://scholar.google.com/scholar?q=Arip+Asadulaev%2C+Rayan+Banerjee%2C+Fakhri+Karray%2C+and+Martin+Takac%2E+2025%2E+Deep+improvement+supervision%2E+arXiv+preprint+arXiv%3A2511%2E16886%2E+%E2%80%94+%5BarXiv%3A2511%2E16886%5D%28https%3A%2F%2Farxiv%2Eorg%2Fabs%2F2511%2E16886%29)
3. Umberto Borso, Davide Paglieri, Jude Wells, and Tim Rocktäschel. 2025. Preference-based alignment of discrete diffusion models. Preprint, arXiv:2503.08295. — [arXiv:2503.08295](https://arxiv.org/abs/2503.08295) [find](https://scholar.google.com/scholar?q=Umberto+Borso%2C+Davide+Paglieri%2C+Jude+Wells%2C+and+Tim+Rockt%C3%A4schel%2E+2025%2E+Preference%2Dbased+alignment+of+discrete+diffusion+models%2E+Preprint%2C+arXiv%3A2503%2E08295%2E+%E2%80%94+%5BarXiv%3A2503%2E08295%5D%28https%3A%2F%2Farxiv%2Eorg%2Fabs%2F2)
4. Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, and David K. Duvenaud. 2018. Neural ordinary differential equations. In Advances in Neural Information Processing Systems, volume 31. [find](https://scholar.google.com/scholar?q=Ricky+T%2E+Q%2E+Chen%2C+Yulia+Rubanova%2C+Jesse+Bettencourt%2C+and+David+K%2E+Duvenaud%2E+2018%2E+Neural+ordinary+differential+equations%2E+In+Advances+in+Neural+Information+Processing+Systems%2C+volume+31%2E)
5. Subhabrata Dutta, Tanya Gautam, Soumen Chakrabarti, and Tanmoy Chakraborty. 2021. Redesigning the transformer architecture with insights from multi-particle dynamical systems. In Advances in Neural Information Processing Systems, volume 34, pages 5531–5544. [find](https://scholar.google.com/scholar?q=Subhabrata+Dutta%2C+Tanya+Gautam%2C+Soumen+Chakrabarti%2C+and+Tanmoy+Chakraborty%2E+2021%2E+Redesigning+the+transformer+architecture+with+insights+from+multi%2Dparticle+dynamical+systems%2E+In+Advances+in+Neural+In)
6. Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. 2021. [Perceiver: General perception with iterative attention](https://md.research.elamur.ai/free/en/papers/jaegle-2021-perceiver-general-perception-iterative-attention/). In Proceedings of the 38th International Conference on Machine Learning, pages 4651–4664.
7. Ahmadreza Jeddi, Marco Ciccone, and Babak Taati. 2026. Loopformer: Elastic-depth looped transformers for latent reasoning via shortcut modulation. arXiv preprint arXiv:2602.11451. [find](https://scholar.google.com/scholar?q=Ahmadreza+Jeddi%2C+Marco+Ciccone%2C+and+Babak+Taati%2E+2026%2E+Loopformer%3A+Elastic%2Ddepth+looped+transformers+for+latent+reasoning+via+shortcut+modulation%2E+arXiv+preprint+arXiv%3A2602%2E11451%2E)
8. Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2019. [ALBERT: A lite BERT for self-supervised learning of language representations](https://md.research.elamur.ai/free/en/papers/lan-2019-albert-lite-bert-self-supervised/). Preprint, arXiv:1909.11942.
9. Nanye Ma, Shangyuan Tong, Haolin Jia, Hexiang Hu, Yu-Chuan Su, Mingda Zhang, Xuan Yang, Yandong Li, Tommi Jaakkola, Xuhui Jia, et al. 2025. Inference-time scaling for diffusion models beyond scaling denoising steps. arXiv preprint arXiv:2501.09732. [find](https://scholar.google.com/scholar?q=Nanye+Ma%2C+Shangyuan+Tong%2C+Haolin+Jia%2C+Hexiang+Hu%2C+Yu%2DChuan+Su%2C+Mingda+Zhang%2C+Xuan+Yang%2C+Yandong+Li%2C+Tommi+Jaakkola%2C+Xuhui+Jia%2C+et+al%2E+2025%2E+Inference%2Dtime+scaling+for+diffusion+models+beyond+scaling+d)
10. Anh Nguyen, Viet Nguyen, Duc Vu, Trung Dao, Chi Tran, Toan Tran, and Anh Tran. 2025. Improved training technique for shortcut models. arXiv preprint arXiv:2510.21250. — [arXiv:2510.21250](https://arxiv.org/abs/2510.21250) [find](https://scholar.google.com/scholar?q=Anh+Nguyen%2C+Viet+Nguyen%2C+Duc+Vu%2C+Trung+Dao%2C+Chi+Tran%2C+Toan+Tran%2C+and+Anh+Tran%2E+2025%2E+Improved+training+technique+for+shortcut+models%2E+arXiv+preprint+arXiv%3A2510%2E21250%2E+%E2%80%94+%5BarXiv%3A2510%2E21250%5D%28https%3A%2F%2Farxi)
11. Kevin Frans, Danijar Hafner, Sergey Levine, and Pieter Abbeel. 2024. One step diffusion via shortcut models. Preprint, arXiv:2410.12557. — [arXiv:2410.12557](https://arxiv.org/abs/2410.12557) [find](https://scholar.google.com/scholar?q=Kevin+Frans%2C+Danijar+Hafner%2C+Sergey+Levine%2C+and+Pieter+Abbeel%2E+2024%2E+One+step+diffusion+via+shortcut+models%2E+Preprint%2C+arXiv%3A2410%2E12557%2E+%E2%80%94+%5BarXiv%3A2410%2E12557%5D%28https%3A%2F%2Farxiv%2Eorg%2Fabs%2F2410%2E12557%29)
12. Vignav Ramesh and Morteza Mardani. 2025. Test-time scaling of diffusion models via noise trajectory search. arXiv preprint arXiv:2506.03164. — [arXiv:2506.03164](https://arxiv.org/abs/2506.03164) [find](https://scholar.google.com/scholar?q=Vignav+Ramesh+and+Morteza+Mardani%2E+2025%2E+Test%2Dtime+scaling+of+diffusion+models+via+noise+trajectory+search%2E+arXiv+preprint+arXiv%3A2506%2E03164%2E+%E2%80%94+%5BarXiv%3A2506%2E03164%5D%28https%3A%2F%2Farxiv%2Eorg%2Fabs%2F2506%2E03164%29)
13. Khashayar Gatmiry, Nikunj Saunshi, Sashank J. Reddi, Stefanie Jegelka, and Sanjiv Kumar. 2024. On the role of depth and looping for in-context learning with task diversity. Preprint, arXiv:2410.21698. — [arXiv:2410.21698](https://arxiv.org/abs/2410.21698) [find](https://scholar.google.com/scholar?q=Khashayar+Gatmiry%2C+Nikunj+Saunshi%2C+Sashank+J%2E+Reddi%2C+Stefanie+Jegelka%2C+and+Sanjiv+Kumar%2E+2024%2E+On+the+role+of+depth+and+looping+for+in%2Dcontext+learning+with+task+diversity%2E+Preprint%2C+arXiv%3A2410%2E21698%2E)
14. Nikunj Saunshi, Nishanth Dikkala, Zhiyuan Li, Sanjiv Kumar, and Sashank J. Reddi. 2025. Reasoning with latent thoughts: On the power of looped transformers. Preprint, arXiv:2502.17416. [find](https://scholar.google.com/scholar?q=Nikunj+Saunshi%2C+Nishanth+Dikkala%2C+Zhiyuan+Li%2C+Sanjiv+Kumar%2C+and+Sashank+J%2E+Reddi%2E+2025%2E+Reasoning+with+latent+thoughts%3A+On+the+power+of+looped+transformers%2E+Preprint%2C+arXiv%3A2502%2E17416%2E)
15. Jonas Geiping, Sean 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. Preprint, arXiv:2502.05171. [find](https://scholar.google.com/scholar?q=Jonas+Geiping%2C+Sean+McLeish%2C+Neel+Jain%2C+John+Kirchenbauer%2C+Siddharth+Singh%2C+Brian+R%2E+Bartoldson%2C+Bhavya+Kailkhura%2C+Abhinav+Bhatele%2C+and+Tom+Goldstein%2E+2025%2E+Scaling+up+test%2Dtime+compute+with+latent+re)
16. Nikunj Saunshi, Stefani Karp, Shankar Krishnan, Sobhan Miryoosefi, Sashank J. Reddi, and Sanjiv Kumar. 2024. On the inductive bias of stacking towards improving reasoning. Preprint, arXiv:2409.19044. — [arXiv:2409.19044](https://arxiv.org/abs/2409.19044) [find](https://scholar.google.com/scholar?q=Nikunj+Saunshi%2C+Stefani+Karp%2C+Shankar+Krishnan%2C+Sobhan+Miryoosefi%2C+Sashank+J%2E+Reddi%2C+and+Sanjiv+Kumar%2E+2024%2E+On+the+inductive+bias+of+stacking+towards+improving+reasoning%2E+Preprint%2C+arXiv%3A2409%2E19044%2E+)
17. Zhengyang Geng, Mingyang Deng, Xingjian Bai, J Zico Kolter, and Kaiming He. 2025. Mean flows for one-step generative modeling. arXiv preprint arXiv:2505.13447. [find](https://scholar.google.com/scholar?q=Zhengyang+Geng%2C+Mingyang+Deng%2C+Xingjian+Bai%2C+J+Zico+Kolter%2C+and+Kaiming+He%2E+2025%2E+Mean+flows+for+one%2Dstep+generative+modeling%2E+arXiv+preprint+arXiv%3A2505%2E13447%2E)
18. Angeliki Giannou, Shashank Rajput, Jy-yong Sohn, Kangwook Lee, Jason D. Lee, and Dimitris Papailiopoulos. 2023. Looped transformers as programmable computers. In Proceedings of the 40th International Conference on Machine Learning, pages 11398–11442. [find](https://scholar.google.com/scholar?q=Angeliki+Giannou%2C+Shashank+Rajput%2C+Jy%2Dyong+Sohn%2C+Kangwook+Lee%2C+Jason+D%2E+Lee%2C+and+Dimitris+Papailiopoulos%2E+2023%2E+Looped+transformers+as+programmable+computers%2E+In+Proceedings+of+the+40th+International+)
19. Shiv Shankar and Tomas Geffner. 2025. Learning straight flows by learning curved interpolants. arXiv preprint arXiv:2503.20719. — [arXiv:2503.20719](https://arxiv.org/abs/2503.20719) [find](https://scholar.google.com/scholar?q=Shiv+Shankar+and+Tomas+Geffner%2E+2025%2E+Learning+straight+flows+by+learning+curved+interpolants%2E+arXiv+preprint+arXiv%3A2503%2E20719%2E+%E2%80%94+%5BarXiv%3A2503%2E20719%5D%28https%3A%2F%2Farxiv%2Eorg%2Fabs%2F2503%2E20719%29)
20. Charlie Victor Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. 2024. [Scaling LLM test-time compute optimally can be more effective than scaling model parameters](https://md.research.elamur.ai/free/en/papers/snell-2024-scaling-llm-test-time-compute/). Preprint, arXiv:2408.03314.
21. Yang Song and Prafulla Dhariwal. 2023. Improved techniques for training consistency models. arXiv preprint arXiv:2310.14189. [find](https://scholar.google.com/scholar?q=Yang+Song+and+Prafulla+Dhariwal%2E+2023%2E+Improved+techniques+for+training+consistency+models%2E+arXiv+preprint+arXiv%3A2310%2E14189%2E)
22. Using the likelihood of the answer as the reward, the above expression simplifies to A rep m,h (y), motivating Eq. (39), where the reward for y ⋆ is maximized while others are suppressed. Our deterministic shortcut-repair model approximates this by parameterizing the future latent through Guang Wang… [find](https://scholar.google.com/scholar?q=Using+the+likelihood+of+the+answer+as+the+reward%2C+the+above+expression+simplifies+to+A+rep+m%2Ch+%28y%29%2C+motivating+Eq%2E+%2839%29%2C+where+the+reward+for+y+%E2%8B%86+is+maximized+while+others+are+suppressed%2E+Our+determin)
23. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed H. Chi, Quoc V. Le, Denny Zhou, et al. 2022. [Chain-of-thought prompting elicits reasoning in large language models](https://md.research.elamur.ai/free/en/papers/wei-2022-chain-thought-prompting-elicits-reasoning/). In Advances in Neural Information Processing Systems, volume 35, pages 24824–24837. ξ m+h ≈ J θ (ξ m , m, h; x) ξ m+…
24. Aiqing Zhu, Pengzhan Jin, Beibei Zhu, and Yifa Tang. 2022. On numerical integration in neural ordinary differential equations. In International Conference on Machine Learning, pages 27527–27547. PMLR. [find](https://scholar.google.com/scholar?q=Aiqing+Zhu%2C+Pengzhan+Jin%2C+Beibei+Zhu%2C+and+Yifa+Tang%2E+2022%2E+On+numerical+integration+in+neural+ordinary+differential+equations%2E+In+International+Conference+on+Machine+Learning%2C+pages+27527%E2%80%9327547%2E+PMLR%2E)

*Updated 1 November 2026*