REVIEW 3 major objections 5 minor 19 references
Amadeus: Autoregressive Model with Bidirectional Attribute Modelling for Symbolic Music
T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash
Pith's one-line read A symbolic-music system that decodes note attributes in parallel, not one by one, claims better quality, control, and at least 4x speed.
desk verdict A genuinely new two-level architecture (AR note generator + bidirectional discrete diffusion for attributes) with a large dataset and real speed gains, but the central 'attributes are unordered' claim is not actually tested and the text-conditioned comparison is confounded by pretraining data. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the masked discrete diffusion note decoder, which treats each note's attributes as tokens that are independently masked during a forward process and jointly recovered in the reverse process; no attribute order is assumed. The note generator is a Transformer decoder that autoregressively outputs a latent vector for each note; the Conditional Information Enhancement Module sharpens this latent via self- and cross-attention to the first note's latent, and the Music Latent Space Discriminability Enhancement Strategy adds a contrastive loss on intermediate music representations so that different samples' latents separate. The masked diffusion decoder then predicts all a
What would settle it
The decisive check is an order-permutation experiment: train the same attribute-decoder architecture with every possible fixed attribute order, or with the order that maximizes validation likelihood, and compare at equal parameters and steps; if any fixed order matches or beats the bidirectional diffusion decoder, the unordered-set premise is false. A cheaper signal is already in the paper's own appendix: H(Duration|Pitch) is about 1.95 while H(Pitch|Duration) is about 3.06, a directional asymmetry that predicts some orders should be easier to learn than others.
Extended reading notes
Core claim
The paper claims that the attributes defining a musical note—pitch, duration, velocity, instrument, beat, chord, tempo, type—are not a sequence with a fixed dependency order but a concurrent, unordered set. Existing generators decode these attributes one after another in a fixed order, which the paper argues is slower and lower-quality precisely because it imposes a temporal structure that the data do not have. Amadeus therefore splits generation into two levels: an autoregressive transformer predicts the next note as a latent vector, and a masked discrete diffusion model decodes that latent into all of the note's attributes bidirectionally, unmasking attributes in parallel rather than in a
Load-bearing premise
The core premise is that a musical note's attributes are an unordered, exchangeable set, so no fixed order of generating them is better than another; if there is actually a best order, or if the attributes constrain each other more strongly than the paper's statistics show, a carefully ordered autoregressive decoder could match or beat the proposed model.
Editorial extensions
If this is right
- If the unordered-attribute premise holds, autoregressive attribute-by-attribute decoding in symbolic music models is a needless bottleneck: parallel masked diffusion gives both higher reported fidelity and several-fold speed.
- Attribute values can be pinned during decoding without retraining, so a single unconditional model can produce music constrained by instrument, tempo, chord, or velocity.
- The number of diffusion steps becomes a quality-speed dial; fewer steps still roughly match baseline quality while running up to about 32 notes per second.
- Text-conditioned generation improves on all five fine-grained controls tested—tempo, key, time signature, instrument coverage, and mood coverage—as well as on prompt-audio alignment.
- The release of the AMD dataset (1.9 million pre-training samples and 320,000 annotated fine-tuning samples) lets others test how the architecture scales with data and parameters.
Reading between the lines
- Editorial extension: because the decoder treats attributes as an unordered set, the same recipe could transfer to any structured prediction where a small set of properties jointly define an entity, such as object-attribute prediction in images or knowledge-graph triples.
- Editorial extension: the paper's own mutual-information table shows directional asymmetries, so a hybrid that lets the model choose the decoding order per note, or that conditions the diffusion schedule on the most informative attribute, may squeeze out further gains.
- Editorial extension: step-adjustable decoding makes latency a deployment knob; this suggests real-time applications where steps are lowered for accompaniment and raised for offline quality, something the paper notes but does not benchmark as a user-facing control.
- Editorial extension: the contrastive latent-discriminability loss is not specific to music; the same idea could improve other latent-conditioned diffusion decoders where a compact vector must carry enough information to reconstruct structured outputs.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes Amadeus, a symbolic-music generation model that keeps autoregressive modelling at the note level but replaces sequential attribute-level decoding by a bidirectional masked discrete diffusion model over the attribute set. Two auxiliary components, MLSDES (a contrastive loss on intermediate note latents) and CIEM (cross-/self-attention conditioning of the diffusion decoder), are introduced. Experiments cover unconditional generation on LakhClean/SOD, text-conditioned generation on MidiCaps, and training-free attribute control, plus an ablation of the two modules and denoising steps. The authors report consistent gains over CPWord, MMT, NMT, REMI, Text2Midi, MuseCoco and T2M-inferalign, a 4x speed-up, and release the AMD dataset. The paper's central conceptual claim is that note attributes are concurrent/unordered, making fixed-order attribute autoregression both slower and lower-quality.
Significance. If established, the central claim would give a substantive modelling insight: intra-note attributes should be decoded as a set rather than a sequence, and it would justify replacing attribute-level AR decoding with parallel bidirectional diffusion. The proposed architecture is sensible, the speed advantage of parallel attribute decoding is credible, and the release of AMD (1.9M pre-training + 320k annotated fine-tuning samples) is a useful community resource. The paper also includes a useful ablation of CIEM/MLSDES and the denoising-step/speed trade-off. However, the main conceptual claim is not supported by the current experimental design: the comparison to baselines confounds the attribute decoder choice with architecture, auxiliary losses, and (for text conditioning) training data, and no control with a fixed-order attribute decoder is reported.
major comments (3)
- [§3.3, §4.6, Appendix B.2] The central 'unordered attributes' claim is never tested at the level that matters. Table 4 removes MLSDES/CIEM and varies denoising steps, but it never replaces the bidirectional DDM attribute decoder with a fixed-order autoregressive attribute decoder while keeping note generator, latent conditioning, data, and parameter count fixed. The anecdotal CPWord/NMT comparison in §1 is not a control because those systems differ in tokenization, architecture, and auxiliary modules. Moreover, Appendix B's own statistics go against the assumption: H(Duration|Pitch)≈1.95 vs H(Pitch|Duration)≈3.06 shows directional asymmetry, and NMI(Beat,Tempo)≈0.45 is substantial. Reported gains in Tables 1 and 3 could therefore come from diffusion/iterative refinement, CIEM/MLSDES, or training data rather than from the unordered-attribute insight.
- [§4, §4.7, Table 3] The text-conditioned comparison is confounded. §4 claims all comparisons use comparable parameter counts (170M) and datasets, but §4.7 states Amadeus was pretrained on the 1.9M-file AMD corpus and then fine-tuned, and Amadeus-M has 500M parameters. The baselines (Text2Midi, MuseCoco, T2M-inferalign) do not appear to receive comparable pretraining. The large gains in Table 3 (TBT 73.93 vs 39.32) may reflect pretraining data and model scale rather than the proposed decoder. Either report baselines with the same pretraining data or clearly restrict the comparison to matched training data and state what pretraining each baseline received.
- [Tables 1, 2, 3, 4] No error bars, confidence intervals, or significance tests are reported for any metric. The abstract and §4.3 use 'significantly outperforms', but many differences are small relative to likely evaluation noise (e.g., SC 0.95 vs 0.97 on LakhClean; CI 26.01 vs 22.42). Metrics are computed on a fixed set of 500 generated samples; repeated sampling or multiple seeds are needed to support the significance language. This matters because the paper's main contribution is a performance claim.
minor comments (5)
- [§4.4] Training-free attribute control is presented only qualitatively and with no baseline comparison. Please specify exactly how a user-specified attribute is injected into the reverse-diffusion decoding (which tokens are fixed, how confidence unmasking treats them) and compare against attribute control in CPWord/NMT or an attribute-masking baseline.
- [Eq. (7)] The formula for numtk is ambiguous: the floor notation and the modulo precedence should be made explicit (numtk = floor(numm/T)+1 if t < numm mod T, else floor(numm/T)). Also, t appears to range from 1 to T, but the condition 't < numm mod T' should be stated.
- [§E.1] The claim that lower pitch entropy and lower pitch-class entropy indicate higher quality is not self-evident; lower entropy can also mean reduced diversity. Please provide evidence or cite prior work that validates this direction for the generation-quality claim.
- [Abstract / Table 3] 'At least 4× speed-up' is not true for Amadeus-M (10.51 notes/s vs 4.02 notes/s ≈ 2.6×). If the speed claim applies only to the 170M model, please state this explicitly.
- [Appendix B.3] The MI analysis is used to conclude that dependencies are 'not strictly unidirectional', but the reported asymmetries are not quantified as a test of exchangeability. Consider reporting a permutation-based or model-based test of whether a fixed order is suboptimal.
Circularity Check
No significant circularity: benchmark comparisons are external and no parameter is fitted then relabeled as a prediction.
full rationale
The paper's derivation chain is not circular. The motivating claim that intra-note attributes form a concurrent, unordered set is supported—however debatably—by external observations about CPWord/NMT initial-token choices and by the Appendix B mutual-information statistics, not by Amadeus itself. The headline performance claims (Tables 1, 2, 3) are evaluated against external baselines (REMI, CPWord, MMT, NMT, Text2Midi, MuseCoco, T2M-inferalign) on external datasets (LakhClean, SOD, MidiCaps) using external metrics (MusPy, CLAP). No parameter is fitted to the evaluation metric and then reported as a prediction; the DDM is adopted from external work (Nie et al. 2025b) and MLSDES's contrastive loss from external work (Wang and He 2025), with no load-bearing self-citation chain. The absence of a same-architecture fixed-order autoregressive attribute decoder ablation means the specific attribution of the reported gains to bidirectionality is under-supported, but that is an experimental-validity limitation, not a circular reduction: the paper never defines 'bidirectional is better' into its equations, and the reported comparisons are independent of the unordered-attribute premise. The Appendix B asymmetry (H(Duration|Pitch) ≈ 1.95 vs. H(Pitch|Duration) ≈ 3.06) is explicitly acknowledged and does not make the analysis self-referential. Therefore the appropriate circularity score is 0.
Assumptions & free parameters
free parameters (5)
- contrastive loss weight lambda =
not reported
- contrastive temperature tau =
not reported
- denoising steps T =
8 (main), 4/2/1 (ablations)
- default tempo/time signature =
120 BPM, 4/4
- recovery schedule numtk =
Eq. (7)
assumptions (8)
- domain assumption Intra-note attributes are concurrent and unordered, so bidirectional modelling is preferable to fixed-order autoregression.
- standard math Masked diffusion forward/reverse process (Eqs. 3-4) correctly defines attribute generation.
- domain assumption Per-attribute tokens are independently masked in the forward process; the joint attribute distribution factorizes over K attributes.
- ad hoc to paper Greedy confidence-based unmasking under the uniform recovery schedule recovers the true joint attribute distribution.
- domain assumption Enhancing inter-sample distance in the note-latent space (MLSDES) improves attribute decoding quality.
- domain assumption The first note's latent z1 provides useful global musical context for decoding later notes.
- domain assumption The evaluation metrics (SC, PE, PCE, CLAP, TBT, CK, CTS, CI, CMtop3) capture generation quality and condition fidelity.
- domain assumption AMD pretraining data and the 80k self-crawled segments are high-quality and unbiased.
Cite this review
Pith. "Pith review of Amadeus: Autoregressive Model with Bidirectional Attribute Modelling for Symbolic Music." pith.science (2026). https://pith.science/paper/IE66IPWX
@misc{pith2026250820665,
author = {Pith},
title = {Pith review of: Amadeus: Autoregressive Model with Bidirectional Attribute Modelling for Symbolic Music},
year = {2026},
howpublished = {\url{https://pith.science/paper/IE66IPWX}},
note = {Machine review of arXiv:2508.20665}
}
abstract
Existing state-of-the-art symbolic music generation models predominantly adopt autoregressive or hierarchical autoregressive architectures, modelling symbolic music as a sequence of attribute tokens with unidirectional temporal dependencies, under the assumption of a fixed, strict dependency structure among these attributes. However, we observe that using different attributes as the initial token in these models leads to comparable performance. This suggests that the attributes of a musical note are, in essence, a concurrent and unordered set, rather than a temporally dependent sequence. Based on this insight, we introduce Amadeus, a novel symbolic music generation framework. Amadeus adopts a two-level architecture: an autoregressive model for note sequences and a bidirectional discrete diffusion model for attributes. To enhance performance, we propose Music Latent Space Discriminability Enhancement Strategy(MLSDES), incorporating contrastive learning constraints that amplify discriminability of intermediate music representations. The Conditional Information Enhancement Module (CIEM) simultaneously strengthens note latent vector representation via attention mechanisms, enabling more precise note decoding. We conduct extensive experiments on unconditional and text-conditioned generation tasks. Amadeus significantly outperforms SOTA models across multiple metrics while achieving at least 4$\times$ speed-up. Furthermore, we demonstrate training-free, fine-grained note attribute control feasibility using our model. To explore the upper performance bound of the Amadeus architecture, we compile the largest open-source symbolic music dataset to date, AMD (Amadeus MIDI Dataset), supporting both pre-training and fine-tuning.
Figures
Figures from the paper (2 more)
Reference graph
Works this paper leans on
-
[5]
Moonbeam: A MIDI Foun- dation Model Using Both Absolute and Relative Music At- tributes. arXiv:2505.15559. Hiller, L. A.; and Isaacson, L. M
-
[8]
MuseCoco: Generating Symbolic Music from Text. arXiv:2306.00110. Melechovsky, J.; Guo, Z.; Ghosal, D.; Majumder, N.; Her- remans, D.; and Poria, S
-
[9]
Mustango: Toward Con- trollable Text-to-Music Generation. In Proceedings of the 2024 Conference of the North American Chapter of the As- sociation for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), 8286–8309. Melechovsky, J.; Roy, A.; and Herremans, D
work page 2024
-
[10]
Mid- iCaps: A Large-scale MIDI Dataset with Text Captions. arXiv:2406.02255. Mogren, O
-
[13]
Byte Latent Transformer: Patches Scale Better Than Tokens. arXiv:2412.09871. Plasser, M.; Peter, S.; and Widmer, G
-
[14]
arXiv preprint arXiv:2404.06393
Mupt: A gener- ative symbolic music pretrained transformer. arXiv preprint arXiv:2404.06393. Radford, A.; Narasimhan, K.; Salimans, T.; and Sutskever, I
-
[15]
arXiv preprint arXiv:2505.12669
Text2midi- InferAlign: Improving Symbolic Music Generation with Inference-Time Alignment. arXiv preprint arXiv:2505.12669. Tian, S.; Zhang, C.; Yuan, W.; Tan, W.; and Zhu, W
-
[16]
arXiv preprint arXiv:2506.09027
Diffuse and Disperse: Im- age Generation with Representation Regularization. arXiv preprint arXiv:2506.09027. Wang, Y .; Wu, S.; Hu, J.; Du, X.; Peng, Y .; Huang, Y .; Fan, S.; Li, X.; Yu, F.; and Sun, M
Show all 19 references
-
[17]
arXiv:2502.18008
NotaGen: Advancing Musicality in Symbolic Music Generation with Large Lan- guage Model Training Paradigms. arXiv:2502.18008. Wu, S.; Wang, Y .; Li, X.; Yu, F.; and Sun, M
-
[18]
arXiv:2407.02277
MelodyT5: A Unified Score-to-Score Transformer for Sym- bolic Music Processing. arXiv:2407.02277. Wu, S.; and Yang, Y
-
[19]
In Cumming, J.; Lee, J
The Jazz Transformer on the Front Line: Exploring the Shortcomings of AI-composed Music through Quantitative Measures. In Cumming, J.; Lee, J. H.; McFee, B.; Schedl, M.; Devaney, J.; McKay, C.; Zangerle, E.; and de Reuse, T., eds., Proceedings of the 21th International Society...
2020
-
[2015]
In 16th International Society for Music Information Retrieval Conference (ISMIR)
Autoregressive hidden semi-markov model of symbolic music performance for score follow- ing. In 16th International Society for Music Information Retrieval Conference (ISMIR). Nie, S.; Zhu, F.; Du, C.; Pang, T.; Liu, Q.; Zeng, G.; Lin, M.; and Li, C. 2025a. Scaling up Masked Di...
-
[2016]
In Constructive Machine Learning Workshop (CML) at NIPS 2016,
C-RNN-GAN: A continuous recurrent neural network with adversarial training. In Constructive Machine Learning Workshop (CML) at NIPS 2016,
2016
-
[2018]
arXiv preprint arXiv:1809.04281
Music transformer. arXiv preprint arXiv:1809.04281. Huang, Y .-S.; and Yang, Y .-H
-
[2020]
arXiv preprint arXiv:2005.00341
Jukebox: A generative model for music. arXiv preprint arXiv:2005.00341. Dong, H.-W.; Chen, K.; Dubnov, S.; McAuley, J.; and Berg- Kirkpatrick, T
2005 arXiv
-
[2022]
arXiv:2205.05448
Symphony Generation with Permutation In- variant Language Model. arXiv:2205.05448. Lu, P.; Xu, X.; Kang, C.; Yu, B.; Xing, C.; Tan, X.; and Bian, J
-
[2023]
Cite arxiv:2301.11325Comment: Supplementary material at https://google-research.github.io/seanet/musiclm/examples and https://kaggle.com/datasets/googleai/musiccaps
MusicLM: Generating Music From Text. Cite arxiv:2301.11325Comment: Supplementary material at https://google-research.github.io/seanet/musiclm/examples and https://kaggle.com/datasets/googleai/musiccaps. Austin, J.; Johnson, D. D.; Ho, J.; Tarlow, D.; and Van Den Berg, R
- [2024]
-
[2025]
arXiv:2504.15071
Aria-MIDI: A Dataset of Piano MIDI Files for Symbolic Music Modeling. arXiv:2504.15071. Collins, T.; Laney, R.; Willis, A.; and Garthwaite, P. H
Reviewed August 5, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.