Elamur Research⌕
☰
home / papers / 2026 / Amadeus: Autoregressive Model with Bidirectional Attribute Modelling for Symbolic Music

Amadeus: Autoregressive Model with Bidirectional Attribute Modelling for Symbolic Music

Paper · 2026 · Hongju Su, Ke Li, Lan Yang et al.

Summary

Most symbolic music generators decode a note's attributes (pitch, duration, velocity, instrument) one by one in a fixed order. This either makes sequences very long or imposes an artificial dependency chain that music theory does not support. The authors instead treat a note's attributes as an unordered set and build Amadeus, a two-level model. A Transformer decoder autoregressively produces one latent vector per note. A bidirectional discrete diffusion decoder then fills in that note's attributes through iterative masked prediction. A Conditional Information Enhancement Module (CIEM) adds global context and sharpens the note latents. For training they release AMD (Amadeus MIDI Dataset), which has 1.9M pre-training pieces and 320K text-annotated fine-tuning pieces. In text-to-MIDI experiments, Amadeus outperforms T2M-InferAlign, MuseCoco and Text2Midi on objective metrics and in listening tests, and it decodes several times faster. Because any subset of attributes can be masked, users can fix some attributes and generate the rest. They can also trade quality for speed by changing the number of denoising steps.

Key points

  • Architecture and models: Amadeus combines a note-level autoregressive generator with an attribute-level discrete diffusion decoder and CIEM. It comes in two sizes: Amadeus (200M parameters) and Amadeus-M (500M).
  • Dataset: AMD is described as the largest open-source symbolic music dataset. It contains 1.9 million pre-training samples and 320,000 fine-tuning samples with rich text annotations.
  • Human evaluation: On a 5-point scale, Amadeus scored 3.82 for musical quality, compared with 3.47 for T2M-InferAlign. It scored 3.73 for mood matching, compared with 2.96. All comparisons were significant (p ≤ 0.0023), and inter-rater agreement was Fleiss' κ = 0.68. The results table reports 20 raters, but the Limitations section mentions 15.
  • Speed: At its highest-quality setting (8 steps), Amadeus generates 16.23 notes/s, compared with 4.02 notes/s for the baselines. With a single denoising step it reaches 32.12 notes/s, at lower quality. Amadeus-M runs at 10.51 notes/s.
  • Ablation: Removing CIEM lowers most metrics, for example TBT drops from 73.93 to 70.01. Using 4 denoising steps sometimes beats 8, for instance on CTS and CM top3.

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

🎓Beijing University of Posts and TelecommunicationsBeijing · China57 works · 32,843 citations🎓University of SurreyGuildford · United Kingdom21 works · 677 citations

 Fields, topics and concepts

📚 Shelves 3The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksComputer Vision and Pattern Recognition · 2,869 worksSignal Processing · 452 works

Cite

Cite as Hongju Su, Ke Li, Lan Yang, Honggang Zhang and Yi-Zhe Song. 2026. Amadeus: Autoregressive Model with Bidirectional Attribute Modelling for Symbolic Music. 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.1898

@inproceedings{su2026amadeus,
  title = {{Amadeus: Autoregressive Model with Bidirectional Attribute Modelling for Symbolic Music}},
  author = {Hongju Su, Ke Li, Lan Yang, Honggang Zhang and Yi-Zhe Song},
  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.1898},
  url = {https://doi.org/10.18653/v1/2026.acl-long.1898}
}

References

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

  1. Andrea Agostinelli, Timo I. Denk, Zalán Borsos, Jesse Engel, Mauro Verzetti, Antoine Caillon, Qingqing Huang, Aren Jansen, Adam Roberts, Marco Tagliasacchi, Matt Sharifi, Neil Zeghidour, and Christian Frank. 2023. Musiclm: Generating music from text. Cite arxiv:2301.11325Comment: Supplementary mater… — [arXiv:2301.11325](https://arxiv.org/abs/2301.11325) find
  2. Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne Van Den Berg. 2021. Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems, 34:17981–17993.
  3. Keshav Bhandari, Abhinaba Roy, Kyra Wang, Geeta Puri, Simon Colton, and Dorien Herremans. 2025. Text2midi: Generating symbolic music from captions. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 23478–23486. find
  4. Jamshed Bharucha and Carol L. Krumhansl. 1983. The representation of harmonic structure in music: Hierarchies of stability as a function of context. Cognition, 13(1):63–102. find
  5. Louis Bradshaw and Simon Colton. 2025. Aria-midi: A dataset of piano midi files for symbolic music modeling. Preprint, arXiv:2504.15071. — [arXiv:2504.15071](https://arxiv.org/abs/2504.15071) find
  6. Tom Collins, Robin Laney, Alistair Willis, and Paul H Garthwaite. 2011. Chopin, mazurkas and markov: Making music in style with statistics. Significance, 8(4):154–159. find
  7. David Cope. 2000. The algorithmic composer, volume 16. AR Editions, Inc. find
  8. Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc Le, and Ruslan Salakhutdinov. 2019. Transformer-XL: Attentive language models beyond a fixed-length context. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2978–2988, Florence, Italy. Assoc… find
  9. Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, and Ilya Sutskever. 2020. Jukebox: A generative model for music. arXiv preprint arXiv:2005.00341. find
  10. Hao-Wen Dong, Ke Chen, Shlomo Dubnov, Julian McAuley, and Taylor Berg-Kirkpatrick. 2023. Multi-track music transformer. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). find
  11. Zach Evans, Julian D. Parker, CJ Carr, Zack Zukowski, Josiah Taylor, and Jordi Pons. 2024. Stable audio open. Preprint, arXiv:2407.14358. — [arXiv:2407.14358](https://arxiv.org/abs/2407.14358) find
  12. Zixun Guo and Simon Dixon. 2025. Moonbeam: A midi foundation model using both absolute and relative music attributes. Preprint, arXiv:2505.15559. — [arXiv:2505.15559](https://arxiv.org/abs/2505.15559) find
  13. Lejaren Arthur Hiller and Leonard M Isaacson. 1957. Experimental Music; Composition with an electronic computer. Greenwood Publishing Group Inc. find
  14. Wen-Yi Hsiao, Jen-Yu Liu, Yin-Cheng Yeh, and Yi-Hsuan Yang. 2021. Compound word transformer: Learning to compose full-song music over dynamic directed hypergraphs. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 178–186. find
  15. Cheng-Zhi Anna Huang, Ashish Vaswani, Jakob Uszkoreit, Noam Shazeer, Ian Simon, Curtis Hawthorne, Andrew M Dai, Matthew D Hoffman, Monica Dinculescu, and Douglas Eck. 2018. Music transformer. arXiv preprint arXiv:1809.04281. find
  16. Yu-Siang Huang and Yi-Hsuan Yang. 2020. Pop music transformer: Beat-based modeling and generation of expressive pop piano compositions. In Proceedings of the 28th ACM International Conference on Multimedia, MM ’20, page 1180–1188, New York, NY, USA. Association for Computing Machinery. find
  17. Jongmin Jung Jiwoo Ryu, Hao-Wen Dong and Dasaem Jeong. 2024. Nested music transformer. In 25th International Society for Music Information Retrieval Conference (ISMIR). find
  18. Keon Ju Maverick Lee, Jeff Ens, Sara Adkins, Pedro Sarmento, Mathieu Barthet, and Philippe Pasquier. 2025. The gigamidi dataset with features for expressive music performance detection. Transactions of the International Society for Music Information Retrieval, 8(1). find
  19. Jiafeng Liu, Yuanliang Dong, Zehua Cheng, Xinran Zhang, Xiaobing Li, Feng Yu, and Maosong Sun. 2022. Symphony generation with permutation invariant language model. Preprint, arXiv:2205.05448. — [arXiv:2205.05448](https://arxiv.org/abs/2205.05448) find
  20. Peiling Lu, Xin Xu, Chenfei Kang, Botao Yu, Chengyi Xing, Xu Tan, and Jiang Bian. 2023. Musecoco: Generating symbolic music from text. Preprint, arXiv:2306.00110. — [arXiv:2306.00110](https://arxiv.org/abs/2306.00110) find
  21. Jan Melechovsky, Zixun Guo, Deepanway Ghosal, Navonil Majumder, Dorien Herremans, and Soujanya Poria. 2024. Mustango: Toward controllable text-to-music generation. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tec… find
  22. Artidoro Pagnoni, Ram Pasunuru, Pedro Rodriguez, John Nguyen, Benjamin Muller, Margaret Li, Chunting Zhou, Lili Yu, Jason Weston, Luke Zettlemoyer, Gargi Ghosh, Mike Lewis, Ari Holtzman, and Srinivasan Iyer. 2024. Byte latent transformer: Patches scale better than tokens. Preprint, arXiv:2412.09871. find
  23. Matthias Plasser, Silvan Peter, and Gerhard Widmer. 2023. Discrete diffusion probabilistic models for symbolic music generation. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI ’23. find
  24. Xingwei Qu, Yuelin Bai, Yinghao Ma, Ziya Zhou, Ka Man Lo, Jiaheng Liu, Ruibin Yuan, Lejun Min, Xueling Liu, Tianyu Zhang, and 1 others. 2024. Mupt: A generative symbolic music pretrained transformer. arXiv preprint arXiv:2404.06393. — [arXiv:2404.06393](https://arxiv.org/abs/2404.06393) find
  25. Colin Raffel. 2016. Learning-Based Methods for Comparing Sequences, with Applications to Audio-to-MIDI Alignment and Matching. Ph.D. thesis, Columbia University, USA. find
  26. Abhinaba Roy, Geeta Puri, and Dorien Herremans. 2025. Text2midi-inferalign: Improving symbolic music generation with inference-time alignment. arXiv preprint arXiv:2505.12669. — [arXiv:2505.12669](https://arxiv.org/abs/2505.12669) find
  27. Sida Tian, Can Zhang, Wei Yuan, Wei Tan, and Wenjie Zhu. 2025. Xmusic: Towards a generalized and controllable symbolic music generation framework. IEEE Transactions on Multimedia. find
  28. Yashan Wang, Shangda Wu, Jianhuai Hu, Xingjian Du, Yueqi Peng, Yongxin Huang, Shuai Fan, Xiaobing Li, Feng Yu, and Maosong Sun. 2025. Notagen: Advancing musicality in symbolic music generation with large language model training paradigms. Preprint, arXiv:2502.18008. — [arXiv:2502.18008](https://arxiv.org/abs/2502.18008) find
  29. Olof Mogren. 2016. C-rnn-gan: A continuous recurrent neural network with adversarial training. In Constructive Machine Learning Workshop (CML) at NIPS 2016, page 1. find
  30. Shangda Wu, Yashan Wang, Xiaobing Li, Feng Yu, and Maosong Sun. 2024. Melodyt5: A unified score-to-score transformer for symbolic music processing. Preprint, arXiv:2407.02277. — [arXiv:2407.02277](https://arxiv.org/abs/2407.02277) find
  31. Eita Nakamura, Philippe Cuvillier, Arshia Cont, Nobu-taka Ono, and Shigeki Sagayama. 2015. Autoregressive hidden semi-markov model of symbolic music performance for score following. In 16th International Society for Music Information Retrieval Conference (ISMIR). find
  32. Yusong Wu*, Ke Chen*, Tianyu Zhang*, Yuchen Hui*, Taylor Berg-Kirkpatrick, and Shlomo Dubnov. 2023. Large-scale contrastive language-audio pretraining with feature fusion and keyword-to-caption augmentation. In IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP. find
  33. Shen Nie, Fengqi Zhu, Chao Du, Tianyu Pang, Qian Liu, Guangtao Zeng, Min Lin, and Chongxuan Li. 2025a. Scaling up masked diffusion models on text. Preprint, arXiv:2410.18514. — [arXiv:2410.18514](https://arxiv.org/abs/2410.18514) find
  34. Iannis Xenakis. 1992. Formalized music: thought and mathematics in composition. 6. Pendragon Press. find
  35. Shen Nie, Fengqi Zhu, Zebin You, Xiaolu Zhang, Jingyang Ou, Jun Hu, Jun Zhou, Yankai Lin, Ji-Rong Wen, and Chongxuan Li. 2025b. Large language diffusion models. arXiv preprint arXiv:2502.09992.
  36. Lili Yu, Dániel Simig, Colin Flaherty, Armen Aghajanyan, Luke Zettlemoyer, and Mike Lewis. 2023. Megabyte: Predicting million-byte sequences with multiscale transformers. Advances in Neural Information Processing Systems, 36:78808–78823. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/su-2026-amadeus-autoregressive-model-bidirectional-attribute/