Elamur Research⌕
☰
home / papers / 2026 / Efficient Learned Data Compression via Dual-Stream Feature Decoupling

Efficient Learned Data Compression via Dual-Stream Feature Decoupling

Paper · 2026 · Huidong Ma, Xinyan Shi, Sun Hui et al.

Summary

Learned Data Compression (LDC) produces smaller files than classic tools like Gzip or zstd, but it is slow. The authors identify two causes. First, single-stream models, usually deep MLP stacks, try to capture short-range regularities and long-range context with one shared set of parameters, which forces deep serial layers and raises decoding latency. Second, in GPU–CPU systems the GPU's probability prediction and the CPU's arithmetic coding run at different speeds, so the pipeline stalls and autoregressive decoding stays serial.

Their compressor, FADE, addresses both. A Dual-Stream Multi-Scale Decoupler runs a CNN branch for local N-gram patterns alongside an MLP branch with a GeGLU rolling cache for global context, and a Content-Adaptive Router merges the two. A Hierarchical Gated Refiner then sharpens the probability estimates. A Concurrent Stream-Parallel Pipeline combines data parallelism with thread-safe, double-buffered temporal parallelism. On standard benchmarks, including Silesia, FADE reports state-of-the-art compression ratio and throughput, with the lowest latency and GPU memory use. This makes neural lossless compression more practical to deploy.

Key points

  • Dual-stream design: two shallow parallel branches (local CNN and global MLP) replace a deep serial MLP stack. Ablations measured in NLL show the combination beats both single-stream baselines.
  • Throughput scaling: with batch size 512, throughput peaks at 4347 KB/min using 8 workers. At batch size 8192 it saturates at 11,216 KB/min.
  • Portable pipeline: adding CSPP to other LDC methods (TRACE, PAC, MSDZip, SEP, EDPC) raises their throughput by 14.88–28.38%, including a 23.00% gain for the already parallel-optimized EDPC.
  • Trade-off: FADE needs slightly more FLOPs and parameters than most LDC baselines (DZip is the exception), yet no compared method matches its inference latency or uses less GPU memory at peak.
  • Code: released at github.com/huidong-ma/FADE.

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

🎓Nanyang Technological UniversitySingapore · Singapore156 works · 25,644 citations🎓Beijing Institute of TechnologyBeijing · China34 works · 3,652 citations🎓Nankai UniversityTianjin · China24 works · 762 citations

 Fields, topics and concepts

📚 Shelves 3The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksComputer Vision and Pattern Recognition · 2,869 works

Cite

Cite as Huidong Ma, Xinyan Shi, Sun Hui, Xiaofei Yue, Xiaoguang Liu, Gang Wang and Wentong Cai. 2026. Efficient Learned Data Compression via Dual-Stream Feature Decoupling. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-long.324

@inproceedings{ma2026efficient,
  title = {{Efficient Learned Data Compression via Dual-Stream Feature Decoupling}},
  author = {Huidong Ma, Xinyan Shi, Sun Hui, Xiaofei Yue, Xiaoguang Liu, Gang Wang and Wentong Cai},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.324},
  url = {https://doi.org/10.18653/v1/2026.acl-long.324}
}

References

46 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Gene M Amdahl. 1967. Validity of the single processor approach to achieving large scale computing capabilities. In Proceedings of the April 18-20, 1967, spring joint computer conference, pages 483–485. find
  2. Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. 2016. Layer normalization. arXiv preprint arXiv:1607.06450.
  3. Jeff Gilchrist. 2003. pbzip2 - parallel bzip2 file compressor. https://compression.ca/pbzip2/. ↗
  4. Mohit Goyal, Kedar Tatwawadi, Shubham Chandak, and Idoia Ochoa. 2018. Deepzip: Lossless data compression using recurrent neural networks. arXiv preprint arXiv:1811.08162. — [arXiv:1811.08162](https://arxiv.org/abs/1811.08162) find
  5. Shaojie Bai, J Zico Kolter, and Vladlen Koltun. 2018. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arXiv preprint arXiv:1803.01271. find
  6. Mohit Goyal, Kedar Tatwawadi, Shubham Chandak, and Idoia Ochoa. 2021. Dzip: improved general-purpose lossless compression based on novel neural network modeling. In 2021 data compression conference (DCC), pages 153–162. IEEE. find
  7. Maximilian Beck, Korbinian Pöppel, Markus Spanring, Andreas Auer, Oleksandra Prudnikova, Michael Kopp, Günter Klambauer, Johannes Brandstetter, and Sepp Hochreiter. 2024. xlstm: Extended long short-term memory. arXiv preprint arXiv:2405.04517. find
  8. Dan Hendrycks and Kevin Gimpel. 2016. Gaussian error linear units (GELUs). arXiv preprint arXiv:1606.08415. find
  9. F Bellard. 2019. Nncp: Lossless data compression with neural networks. https://bellard.org/nncp/. ↗
  10. David A Huffman. 1952. A method for the construction of minimum-redundancy codes. Proceedings of the IRE, 40(9):1098–1101. find
  11. Fabrice Bellard. 2021. Nncp v2: Lossless data compression with transformer. Preprint at Fabrice Bellard https://bellard.org/nncp/nncp_v2. pdf. ↗
  12. Keith Ito and Linda Johnson. 2017. The lj speech dataset. https://keithito.com/LJ-Speech-Dataset. ↗
  13. Urvashi Khandelwal, He He, Peng Qi, and Dan Jurafsky. 2018. Sharp nearby, fuzzy far away: How neural language models use context. arXiv preprint arXiv:1805.04623. find
  14. Byron Knoll. 2016. https://github.com/byronknoll/cmix. ↗
  15. Byron Knoll. 2017. lstm-compress. https://github.com/byronknoll/lstm-compress. ↗
  16. Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. 2002. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324. find
  17. Valerii Likhosherstov, Krzysztof M Choromanski, Jared Quincy Davis, Xingyou Song, and Adrian Weller. 2021. Sub-linear memory: How to make performers slim. Advances in Neural Information Processing Systems, 34:6707–6719. find
  18. Qian Liu, Yiling Xu, and Zhu Li. 2019. Decmac: A deep context model for high efficiency arithmetic coding. In 2019 International Conference on Artificial Intelligence in Information and Communication (ICAIIC), pages 438–443. IEEE. find
  19. Zeyi Lu, Xiaoxiao Ma, Yujun Huang, Minxiao Chen, Bin Chen, Baoyi An, and Shu-Tao Xia. 2025. Edpc: Accelerating lossless compression via lightweight probability models and decoupled parallel dataflow. In Proceedings of the 33rd ACM International Conference on Multimedia, pages 7268–7276. find
  20. Huidong Ma, Hui Sun, Liping Yi, Yanfeng Ding, Xiaoguang Liu, and Gang Wang. 2025a. Msdzip: Universal lossless compression for multi-source data via stepwise-parallel and learning-based prediction. In Proceedings of the ACM on Web Conference 2025, pages 3543–3551. find
  21. Matt Mahoney. 2006. Large text compression benchmark. https://www.mattmahoney.net/dc/textdata.html. ↗
  22. Yu Mao, Yufei Cui, Tei-Wei Kuo, and Chun Jason Xue. 2022a. Accelerating general-purpose lossless compression via simple and scalable parameterization. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3205–3213. find
  23. Yu Mao, Yufei Cui, Tei-Wei Kuo, and Chun Jason Xue. 2022b. Trace: A fast transformer-based general-purpose lossless compressor. In Proceedings of the ACM Web Conference 2022, pages 1829–1838. find
  24. Yu Mao, Jingzong Li, Yufei Cui, and Chun Jason Xue. 2023. Faster and stronger lossless compression with optimized autoregressive framework. In 2023 60th ACM/IEEE Design Automation Conference (DAC), pages 1–6. IEEE. find
  25. Alexandre Mercat, Marko Viitanen, and Jarno Vanne. 2020. Uvg dataset: 50/120fps 4k sequences for video codec analysis and development. In Proceedings of the 11th ACM multimedia systems conference, pages 297–302. find
  26. Igor Pavlov. 1999. 7z official website. https://www.7-zip.org/. ↗
  27. Diogo Pratas and Armando J Pinho. 2019. A dna sequence corpus for compression benchmark. In Practical Applications of Computational Biology and Bioinformatics, 12th International Conference, pages 208–215. Springer. find
  28. Rawzor. 2008. Image compression benchmark. http://imagecompression.info/test_images/. ↗
  29. Hochreiter Sepp and Schmidhuber Jürgen. 1997. Long short-term memory. Neural Computation MIT-Press. find
  30. Julian Seward. 1996. The official website of the xz compressor. find
  31. Huidong Ma, Hui Sun, Liping Yi, Xiaoguang Liu, and Gang Wang. 2025b. Multi-source data lossless compression via parallel expansion mapping and xlstm. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE. find
  32. Claude E Shannon. 1948. A mathematical theory of communication. The Bell system technical journal, 27(3):379–423. find
  33. Huidong Ma, Cheng Zhong, Danyang Chen, Haofa He, and Feng Yang. 2023a. cnnlsv: detecting structural variants by encoding long-read alignment information and convolutional neural network. BMC bioinformatics, 24(1):119. find
  34. Hui Sun, Yanfeng Ding, Liping Yi, Huidong Ma, Gang Wang, Xiaoguang Liu, Cheng Zhong, and Wentong Cai. 2025a. Pmklc: Parallel multi-knowledge learning-based lossless compression for large-scale genomics database. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining… find
  35. Huidong Ma, Cheng Zhong, Hui Sun, Danyang Chen, and Haixiang Lin. 2023b. ricme: Long-read based mobile element variant detection using sequence re-alignment and identity calculation. In International Symposium on Bioinformatics Research and Applications, pages 165–177. Springer. find
  36. Noam Shazeer. 2020. GLU variants improve transformer. arXiv preprint arXiv:2002.05202. find
  37. Hui Sun, Huidong Ma, Feng Ling, Haonan Xie, Yongxia Sun, Liping Yi, Meng Yan, Cheng Zhong, Xiaoguang Liu, and Gang Wang. 2025b. A survey and benchmark evaluation for neural-network-based lossless universal compressors toward multi-source data. Frontiers of Computer Science, 19(7):1–16. find
  38. Hui Sun, Huidong Ma, Yingfeng Zheng, Haonan Xie, Meng Yan, Cheng Zhong, Xiaoguang Liu, and Gang Wang. 2024a. Lrcb: A comprehensive benchmark evaluation of reference-free lossless compression tools for genomics sequencing long reads data. In 2024 Data Compression Conference (DCC), pages 584–584. IEEE… find
  39. Hui Sun, Liping Yi, Huidong Ma, Yongxia Sun, Yingfeng Zheng, Wenwen Cui, Meng Yan, Gang Wang, and Xiaoguang Liu. 2025c. Genomics data lossless compression with (s, k)-mer encoding and deep neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 12577–12585… find
  40. Hui Sun, Yingfeng Zheng, Haonan Xie, Huidong Ma, Cheng Zhong, Meng Yan, Xiaoguang Liu, and Gang Wang. 2024b. Pqsdc: a parallel lossless compressor for quality scores data via sequences partition and run-length prediction mapping. Bioinformatics, 40(5):btae323. find
  41. Chandra Shekhara Kaushik Valmeekam, Krishna Narayanan, Dileep Kalathil, Jean-Francois Chamber-land, and Srinivas Shakkottai. 2023. Llmzip: Lossless text compression using large language models. arXiv preprint arXiv:2306.04050. — [arXiv:2306.04050](https://arxiv.org/abs/2306.04050) find
  42. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, and Łukasz Kaiser. 2017. Attention is all you need. Advances in Neural Information Processing Systems.
  43. Meng Wan, Rongqiang Cao, Yanghao Li, Jue Wang, Zijian Wang, Qi Su, Lei Qiu, Peng Shi, Yangang Wang, and Chong Li. 2025. Sep: A general lossless compression framework with semantics enhancement and multi-stream pipelines. In Proceedings of the Thirty-Fourth International Joint Conference on Artificia… find
  44. Ian H Witten, Radford M Neal, and John G Cleary. 1987. Arithmetic coding for data compression. Communications of the ACM, 30(6):520–540. find
  45. Kai Zhao, Sheng Di, Xin Lian, Sihuan Li, Dingwen Tao, Julie Bessac, Zizhong Chen, and Franck Cappello. 2020. Sdrbench: Scientific data reduction benchmark for lossy compressors. In 2020 IEEE international conference on big data (Big Data), pages 2716–2724. IEEE. find
  46. Jacob Ziv and Abraham Lempel. 1977. A universal algorithm for sequential data compression. IEEE Transactions on information theory, 23(3):337–343. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/ma-2026-efficient-learned-data-compression-dual/