---
title: "FlowHN — paper by Moradi et al. (2026) — summary and ideas"
description: "FlowHN replaces learned routing with a fixed, hardware-aligned scheduling policy. 32.7% accuracy. 23 references."
authors: ["Mohammad Mahdi Moradi", "Walid Ahmed", "Shuangyue Wen", "Sudhir Mudur", "Weiwei Zhang", "Yang Liu"]
year: 2026
venue: "Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track)"
source: "https://aclanthology.org/2026.acl-industry.83/"
doi: "10.18653/v1/2026.acl-industry.83"
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/moradi-2026-flowhn-adaptive-token-routing-efficient/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/moradi-2026-flowhn-adaptive-token-routing-efficient/)

[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/33/) / FlowHN: Adaptive Token Routing for Efficient Parallel Hybrid Networks

# FlowHN: Adaptive Token Routing for Efficient Parallel Hybrid Networks

**Authors:** Mohammad Mahdi Moradi, Walid Ahmed, Shuangyue Wen, Sudhir Mudur, [Weiwei Zhang](https://md.research.elamur.ai/free/en/authors/weiwei-zhang/) and [Yang Liu](https://md.research.elamur.ai/free/en/authors/yang-liu/) · **Year:** 2026 · **Venue:** [Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track)](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-industry.83/) · [PDF](https://aclanthology.org/2026.acl-industry.83.pdf) · [DOI 10.18653/v1/2026.acl-industry.83](https://doi.org/10.18653/v1/2026.acl-industry.83) · [OpenAlex](https://openalex.org/W7166816281)

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

## Summary
Hybrid language models pair attention, which handles precise token interactions but scales quadratically, with state-space models (SSMs), which scale linearly but are weaker at fine-grained retrieval. Running both branches in parallel avoids idle time, but the branches have different FLOP costs. The faster branch then waits for the slower one, and learned routing (e.g., mixture-of-experts) makes latency less predictable. FlowHN replaces learned routing with a fixed, hardware-aligned scheduling policy. Each parallel block gets a share of tokens sized to each branch's per-token FLOP cost. These assignments then rotate across layers so every token eventually passes through both attention and SSM. Branch outputs are merged by concatenation followed by a learned projection. The authors compare four allocation variants against LLaMA, Mamba-1/2, Jamba and Hymba at 135M, 350M and 1B parameters. The circulating variant, FAC_Split, gives the best balance of efficiency and accuracy, with the highest MFU at every scale. Because FlowHN needs no changes to optimizers, training stacks or serving infrastructure, it suits latency-sensitive production workloads.

## Key points
- **Four token-allocation strategies were tested:**
  - No_Split runs every token through both branches and is the most accurate but costliest.
  - AE_Split divides tokens evenly and alternates them between blocks.
  - FA_Split assigns tokens by FLOP cost but never moves them.
  - FAC_Split assigns by FLOP cost and rotates tokens across layers.
- **Throughput:** FAC_Split reaches about 34K TPS at 135M, 14K at 350M and 7.9K at 1B. The abstract reports up to 4× throughput and 15% higher MFU than baselines. The one faster model is Jamba-1B, whose MoE design activates only 483M parameters.
- **Long context:** On LongFQA (8K), NQ (16K) and CUAD (32K), evaluated 3-shot on 8×V100, FAC_Split-1B averages 32.7% accuracy, 3781 TPS and 53.4% MFU. Its speed advantage over Jamba grows from 1.91× to 2.94× as context length increases, and over Hymba from 2.47× to 3.37×.
- **Other tasks at 1B scale:** FAC_Split-1B scores 22.2 on GSM8K, the best among the tested models, at roughly 10.7K TPS. On the GovReport and SQuALITY summarization benchmarks it matches baseline ROUGE-L while running over 2× faster than Jamba and Hymba.
- **Stated limitations:** The FLOP estimates used for scheduling are static. Larger-scale and distributed-serving evaluations remain future work.

*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/) · 18,614 works · [👁️ Computer Vision and Pattern Recognition](https://md.research.elamur.ai/free/en/shelves/computer-science/computer-vision-and-pattern-recognition/) · 3,486 works · [🔌 Hardware and Architecture](https://md.research.elamur.ai/free/en/shelves/computer-science/hardware-and-architecture/) · 167 works · [🌐 Computer Networks and Communications](https://md.research.elamur.ai/free/en/shelves/computer-science/computer-networks-and-communications/) · 100 works

## Cite

Cite as: Mohammad Mahdi Moradi, Walid Ahmed, Shuangyue Wen, Sudhir Mudur, Weiwei Zhang and Yang Liu. 2026. FlowHN: Adaptive Token Routing for Efficient Parallel Hybrid Networks. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track).

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

```bibtex
@inproceedings{moradi2026flowhn,
  title = {{FlowHN: Adaptive Token Routing for Efficient Parallel Hybrid Networks}},
  author = {Mohammad Mahdi Moradi, Walid Ahmed, Shuangyue Wen, Sudhir Mudur, Weiwei Zhang and Yang Liu},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-industry.83},
  url = {https://doi.org/10.18653/v1/2026.acl-industry.83}
}
```

- [← previous Flow-Based Page Unique Semantic Mapping Architecture for Document Visual Question Answering](https://md.research.elamur.ai/free/en/papers/wang-2026-flow-based-page-unique-semantic/)
- [next → FlowRAG: Synergizing Explicit Reasoning via Frequency-Aware Multi-Granularity Graph Flow](https://md.research.elamur.ai/free/en/papers/zhan-2026-flowrag-synergizing-explicit-reasoning-frequency/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/33/)

## References

23 references, 12 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie Cai, Michael Terry, Quoc Le, and 1 others. 2021. [Program synthesis with large language models](https://md.research.elamur.ai/free/en/papers/austin-2021-program-synthesis-large-language-models/). arXiv preprint arXiv:2108.07732.
2. Mohammad Bavarian, Heewoo Jun, Nikolas Tezak, John Schulman, Christine McLeavey, Jerry Tworek, and Mark Chen. 2022. [Efficient training of language models to fill in the middle](https://md.research.elamur.ai/free/en/papers/bavarian-2022-efficient-training-language-models-fill/). arXiv preprint arXiv:2207.14255.
3. Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, and 1 others. 2023. [Palm: Scaling language modeling with pathways](https://md.research.elamur.ai/free/en/papers/chowdhery-2022-palm-scaling-language-modeling-pathways/). Journal of Machine Learning Research, 24(240):1–113.
4. 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](https://md.research.elamur.ai/free/en/papers/cobbe-2021-training-verifiers-solve-math-word/). arXiv preprint arXiv:2110.14168.
5. Xin Dong, Yonggan Fu, Shizhe Diao, Wonmin Byeon, Zijia Chen, Ameya Sunil Mahabaleshwarkar, Shih-Yang Liu, Matthijs Van Keirsbilck, Min-Hung Chen, Yoshi Suhara, and 1 others. 2024. [Hymba: A hybrid-head architecture for small language models](https://md.research.elamur.ai/free/en/papers/dong-2024-hymba-hybrid-head-architecture-small/). arXiv preprint arXiv:2411.13676.
6. Daniel Y Fu, Tri Dao, Khaled K Saab, Armin W Thomas, Atri Rudra, and Christopher Ré. 2022. Hungry hungry hippos: Towards language modeling with state space models. arXiv preprint arXiv:2212.14052. [find](https://scholar.google.com/scholar?q=Daniel+Y+Fu%2C+Tri+Dao%2C+Khaled+K+Saab%2C+Armin+W+Thomas%2C+Atri+Rudra%2C+and+Christopher+R%C3%A9%2E+2022%2E+Hungry+hungry+hippos%3A+Towards+language+modeling+with+state+space+models%2E+arXiv+preprint+arXiv%3A2212%2E14052%2E)
7. Paolo Glorioso, Quentin Anthony, Yury Tokpanov, and 1 others. 2024. Zamba: A compact 7b ssm hybrid model. arXiv preprint arXiv:2405.XXXXX. [find](https://scholar.google.com/scholar?q=Paolo+Glorioso%2C+Quentin+Anthony%2C+Yury+Tokpanov%2C+and+1+others%2E+2024%2E+Zamba%3A+A+compact+7b+ssm+hybrid+model%2E+arXiv+preprint+arXiv%3A2405%2EXXXXX%2E)
8. Albert Gu and Tri Dao. 2023. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752. [find](https://scholar.google.com/scholar?q=Albert+Gu+and+Tri+Dao%2E+2023%2E+Mamba%3A+Linear%2Dtime+sequence+modeling+with+selective+state+spaces%2E+arXiv+preprint+arXiv%3A2312%2E00752%2E)
9. Dan Hendrycks, Steven Basart, Saurav Kadavath, Mantas Mazeika, Akul Arora, Ethan Guo, Collin Burns, Samir Puranik, Horace He, Dawn Song, and 1 others. 2021. [Measuring coding challenge competence with apps](https://md.research.elamur.ai/free/en/papers/hendrycks-2021-measuring-coding-challenge-competence-apps/). arXiv preprint arXiv:2105.09938.
10. Luyang Huang, Shuyang Cao, Nikolaus Parulian, Heng Ji, and Lu Wang. 2021. [Efficient attentions for long document summarization](https://md.research.elamur.ai/free/en/papers/huang-2021-efficient-attentions-long-document-summarization/). arXiv preprint arXiv:2104.02112.
11. Hamel Husain, Ho-Hsiang Wu, Tiferet Gazit, Miltiadis Allamanis, and Marc Brockschmidt. 1909. [Codesearchnet challenge: Evaluating the state of semantic code search](https://md.research.elamur.ai/free/en/papers/husain-2020-codesearchnet-challenge-evaluating-state-semantic/).(2019). arXiv preprint arXiv:1909.09436.
12. Arkil Patel, Satwik Bhattamishra, and Navin Goyal. 2021. [Are nlp models really able to solve simple math word problems?](https://md.research.elamur.ai/free/en/papers/patel-2021-are-nlp-models-really-able/) arXiv preprint arXiv:2103.07191.
13. Liliang Ren, Yang Liu, Yadong Lu, Yelong Shen, Chen Liang, and Weizhu Chen. 2024. [Samba: Simple hybrid state space models for efficient unlimited context language modeling](https://md.research.elamur.ai/free/en/papers/ren-2024-samba-simple-hybrid-state-space/). arXiv preprint arXiv:2406.07522.
14. Eva Sharma, Chen Li, and Lu Wang. 2019. [Bigpatent: A large-scale dataset for abstractive and coherent summarization](https://md.research.elamur.ai/free/en/papers/sharma-2019-bigpatent-large-scale-dataset-abstractive/). arXiv preprint arXiv:1906.03741.
15. Daria Soboleva, Faisal Al-Khateeb, Robert Myers, Jacob R Steeves, Joel Hestness, and Nolan Dey. 2023. SlimPajama: A 627B token cleaned and deduplicated version of RedPajama. [https://www.cerebras.net/blog/slimpajama-a-627b-token-cleaned-a](https://www.cerebras.net/blog/slimpajama-a-627b-token-cleaned-a) nd-deduplicated-version-of-redpajama. [↗](https://www.cerebras.net/blog/slimpajama-a-627b-token-cleaned-a)
16. Alex Wang, Richard Yuanzhe Pang, Angelica Chen, Jason Phang, and Samuel R Bowman. 2022. Squality: Building a long-document summarization dataset the hard way. arXiv preprint arXiv:2205.11465. [find](https://scholar.google.com/scholar?q=Alex+Wang%2C+Richard+Yuanzhe+Pang%2C+Angelica+Chen%2C+Jason+Phang%2C+and+Samuel+R+Bowman%2E+2022%2E+Squality%3A+Building+a+long%2Ddocument+summarization+dataset+the+hard+way%2E+arXiv+preprint+arXiv%3A2205%2E11465%2E)
17. Songlin Yang, Bailin Wang, Yikang Shen, Rameswar Panda, and Yoon Kim. 2023. Gated linear attention transformers with hardware-efficient training. arXiv preprint arXiv:2312.06635. [find](https://scholar.google.com/scholar?q=Songlin+Yang%2C+Bailin+Wang%2C+Yikang+Shen%2C+Rameswar+Panda%2C+and+Yoon+Kim%2E+2023%2E+Gated+linear+attention+transformers+with+hardware%2Defficient+training%2E+arXiv+preprint+arXiv%3A2312%2E06635%2E)
18. Qiuhao Zeng, Jerry Huang, Peng Lu, Gezheng Xu, Boxing Chen, Charles Ling, and Boyu Wang. 2025. Zeta: Leveraging z-order curves for efficient top-k attention. arXiv preprint arXiv:2501.14577. — [arXiv:2501.14577](https://arxiv.org/abs/2501.14577) [find](https://scholar.google.com/scholar?q=Qiuhao+Zeng%2C+Jerry+Huang%2C+Peng+Lu%2C+Gezheng+Xu%2C+Boxing+Chen%2C+Charles+Ling%2C+and+Boyu+Wang%2E+2025%2E+Zeta%3A+Leveraging+z%2Dorder+curves+for+efficient+top%2Dk+attention%2E+arXiv+preprint+arXiv%3A2501%2E14577%2E+%E2%80%94+%5BarXiv%3A)
19. Danyang Zhang, Junhao Song, Ziqian Bi, Xinyuan Song, Yingfang Yuan, Tianyang Wang, Joe Yeong, and Junfeng Hao. 2025. Mixture of experts in large language models. arXiv preprint arXiv:2507.11181. — [arXiv:2507.11181](https://arxiv.org/abs/2507.11181) [find](https://scholar.google.com/scholar?q=Danyang+Zhang%2C+Junhao+Song%2C+Ziqian+Bi%2C+Xinyuan+Song%2C+Yingfang+Yuan%2C+Tianyang+Wang%2C+Joe+Yeong%2C+and+Junfeng+Hao%2E+2025%2E+Mixture+of+experts+in+large+language+models%2E+arXiv+preprint+arXiv%3A2507%2E11181%2E+%E2%80%94+%5Bar)
20. Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. 2020. Transformers are rnns: Fast autoregressive transformers with linear attention. In International conference on machine learning, pages 5156–5165. PMLR. [find](https://scholar.google.com/scholar?q=Angelos+Katharopoulos%2C+Apoorv+Vyas%2C+Nikolaos+Pappas%2C+and+Fran%C3%A7ois+Fleuret%2E+2020%2E+Transformers+are+rnns%3A+Fast+autoregressive+transformers+with+linear+attention%2E+In+International+conference+on+machine+l)
21. Opher Lieber, Barak Lenz, Hofit Bata, Gal Cohen, Jhonathan Osin, Itay Dalmedigos, Erez Safahi, Shaked Meirom, Yonatan Belinkov, Shai Shalev-Shwartz, and 1 others. 2024. [Jamba: A hybrid transformer-mamba language model](https://md.research.elamur.ai/free/en/papers/lieber-2024-jamba-hybrid-transformer-mamba-language/). arXiv preprint arXiv:2403.19887.
22. Wang Ling, Dani Yogatama, Chris Dyer, and Phil Blunsom. 2017. Program induction by rationale generation: Learning to solve and explain algebraic word problems. arXiv preprint arXiv:1705.04146. [find](https://scholar.google.com/scholar?q=Wang+Ling%2C+Dani+Yogatama%2C+Chris+Dyer%2C+and+Phil+Blunsom%2E+2017%2E+Program+induction+by+rationale+generation%3A+Learning+to+solve+and+explain+algebraic+word+problems%2E+arXiv+preprint+arXiv%3A1705%2E04146%2E)
23. Shen-Yun Miao, Chao-Chun Liang, and Keh-Yih Su. 2021. A diverse corpus for evaluating and developing english math word problem solvers. arXiv preprint arXiv:2106.15772. [find](https://scholar.google.com/scholar?q=Shen%2DYun+Miao%2C+Chao%2DChun+Liang%2C+and+Keh%2DYih+Su%2E+2021%2E+A+diverse+corpus+for+evaluating+and+developing+english+math+word+problem+solvers%2E+arXiv+preprint+arXiv%3A2106%2E15772%2E)

*Updated 1 November 2026*