acceptodds
Under review as a conference paper at ICLR 2027

Hyperbolic Music Attribution

Abstract

Music production commonly works with separate components such as drum parts, bass lines, stems and submixes. As generative models are developed to produce such components, their training data may also include individual stems or mixtures of stems rather than only complete recordings. This makes provenance a structured problem: can exposure to a component during adaptation be detected, and can that effect be localised to the component rather than extending to the song it belongs to? This distinction matters for auditing generative music systems, including applications in content identification and copyright analysis. Because stems, submixes, and full recordings are related by containment, provenance also has a hierarchical structure, motivating representations that can capture these relations explicitly. We first study training-data membership under controlled partial exposure. A latent-diffusion text-to-music model is fine-tuned with a low-rank adapter on selected stems and mixtures, and we compare trained mixtures, untrained mixtures sharing constituent stems with trained mixtures, and clips containing no exposed stems. This allows us to examine whether exposure can be detected at the component level rather than only at the level of the complete song. We then ask whether the structure among these musical parts is better represented in hyperbolic than Euclidean space. We compare hyperbolic and Euclidean projection heads of identical size over a frozen audio encoder, trained on known stem-containment relations and evaluated on unseen songs. To the best of our knowledge, this is one of the first studies to jointly examine component-level training exposure and hyperbolic representations of the hierarchical relation among stems, submixes, and complete recordings. Together, these experiments provide a controlled framework for studying partial training exposure in structured data, where the relevant unit of provenance may be a component, a composition of components, or a larger hierarchy containing them.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
training data provenance, hyperbolic representation learning, hierarchical representation learning, music generation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_A87UOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper focuses on detecting whether individual components of a song were used in training of a music generation model. The authors experiment on multiple DoRA fine-tuning runs of a Stable Audio 3 Small diffusion model. Importantly, the paper distinguishes exact-mixture membership from constituent exposure using candidate-matched denoising scores. They find loss differences consistent with constituent exposure in mixtures that were not themselves used for fine-tuning. Additionally, the paper explores hyperbolic representations of musical components. Hyperbolic representations achieve better containment prediction than Euclidean representations under certain evaluation settings.

Critical strengths
  • The paper investigates an interesting problem of detecting the training exposure of music components. The membership inference/data contamination literature is quite saturated, so focusing on innovative and novel aspects in this branch of research is important. The problem statement is the strongest aspect of the paper in my view.

  • The authors evaluate on multiple datasets and include comparisons between several types of training. (However, see weaknesses). Additionally, the paper includes several additional experiments, ablations, and analyses of its limitations

Critical weaknesses
  • The membership-inference evaluation is limited to fine-tuned models. raising some concerns about the generality and soundness of the results. I am aware of the challenging nature of running membership inference problems (open checkpoints with public data splits are hard to find, and training from scratch is costly). Still, the paper would be strengthened by experimentally exploring a full model training pipeline. (see for example: Dubiński et al., Towards More Realistic Membership Inference Attacks on Large Diffusion Models (WACV 2024))

  • The benefit of hyperbolic representations is limited.

Questions for authors

The comparison between hyperbolic and Euclidean representations may not be fully sound. In Section 4.3, the authors state that 1) some hyperparameters were selected based on the test set, and 2) the tuning was performed only for the hyperbolic model; the Euclidean baseline used the same settings as the hyperbolic model. The difference in results between hyperbolic and Euclidean is small. Would the hyperbolic approach still outperform the Euclidean approach if both were tuned independently (without using test set results)?

I maintain my positive score after the rebuttal.

1 comment

Thank you for your review and suggestions.

  1. We conducted a bidirectional small-model study comprising twelve fits of a 9.68-million-parameter denoiser from a shared random initialization, with frozen pretrained audio and text encoders. We trained on MoisesDB-4 and then fine-tuned on Billboard, and repeated the experiment in the reverse direction. Listening to eight fixed generation examples spanning both directions and training phases revealed no coherent musical output. Although denoising loss improved, the inadequate generation quality limits interpretation of the exposure contrasts. Limited training data may have contributed, but this experiment does not establish the cause. We therefore treat the study as inconclusive.
  2. We independently tuned the hyperbolic and Euclidean models using equal validation search budgets on corrected artist/recording-group splits, then trained fresh final models across five folds and three seeds. Hyperbolic achieved higher native 11-stem MoisesDB containment AUC than Euclidean (0.766 versus 0.707), while the MoisesDB-4 difference was small (0.872 versus 0.868). Our audio-only adaptation of Vendrov et al.’s order embeddings achieved AUCs of 0.765 and 0.909, respectively, matching hyperbolic on the native task and exceeding it on MoisesDB-4. Count-matched superset retrieval scores (hyperbolic/Euclidean/order) were 70.6/68.1/58.0% on native MoisesDB and 90.7/89.7/85.7% on MoisesDB-4. A retrieval trial succeeds when a true containing mixture ranks above every same-window, same-size non-containing candidate, this evaluates ranking rather than thresholded containment decisions.
Reviewer_U2nWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper has two parts. The first part fine-tunes Stable Audio 3 Small with a low-rank adapter on stems and mixtures, and uses denoising scores to test whether training exposure can be detected on untrained mixtures that share constituent stems with trained mixtures, versus clips with no exposed stems. The second part compares hyperbolic and Euclidean projection heads over a frozen CLAP encoder on stem-containment relations. Section 5.3 tests whether the containment representations help exposure discrimination via reconstruction.

Critical strengths
  1. Separating exact-mixture membership from constituent exposure in the first half is well motivated for multitrack music. Although the pooled AUC between partial exposure and unexposed clips is modest, the candidate-matched paired test compares the exact same mixture under an adapter where a constituent stem was exposed versus an adapter where no stem was exposed. This controls for clip difficulty and yields positive differences in 74% to 82% of pairs.

  2. The authors are honest about negative results and their configuration search history. They clearly report the chance-level reconstruction test in Section 5.3 and the tuning stages in Section 4.3 and Appendix A.

Critical weaknesses
  1. The head comparison does not isolate the effect of geometry. The four-stage development process used test results to choose settings. The Euclidean head simply inherited settings tuned for the hyperbolic head, including the sibling hinge weight and margin. Also, the final setup uses only one seed per fold. Because of this, we cannot tell if the gain comes from hyperbolic geometry or from tuning in its favor.

  2. The evidence for a general hyperbolic advantage in stem containment is weak. In Table 2, Euclidean direction accuracy is actually higher at the trained aperture scale, and the order only flips after validation scaling. Validation scaling is fine, but it means the win depends on the threshold choice. Scaling also pushes the maximum cone angle above 90 degrees, so it only tests pairwise angles and does not give a transitive containment order. On aperture-free superset retrieval, the two heads perform almost the same. Since song stems form a subset lattice rather than a tree, tree-based intuition does not show why hyperbolic space is needed here.

  3. The proposed link between containment learning and attribution is not supported by the experiments. In Section 5.3, using the containment heads for reconstruction-based exposure scoring does not beat raw CLAP. While the separate denoising tests in Section 5.1 do show training exposure, the hyperbolic head does not help exposure detection, and the paper does not attribute generated outputs to training tracks. This makes the title broader than what the paper actually shows.

  4. The exposure experiments have unresolved data leakage and dependence issues. Section 4.1 notes artist name variants and possible duplicate recordings across folds, so an unseen track may have a duplicate in training. Also, Section 4.3 gives no confidence intervals for the denoising exposure tests because mixtures share songs and stems, while Appendix E uses candidate-level bootstrap intervals that ignore song-level grouping.

  5. The source and licensing of the Billboard audio are not clear. The public McGill Billboard release only shares annotations and features due to copyright limits, whereas this paper uses Billboard audio processed with Demucs. The paper does not explain how the recordings were obtained, what license allows their use, or how its track counts map to the official release.

Questions for authors
  1. If both heads are tuned independently on validation data only with the same search budget, does the final hyperbolic setup still win across multiple seeds on clean splits? Testing a simple order-embedding baseline would also help show whether hyperbolic space is really needed.

  2. Can you add song-clustered confidence intervals for the partial-versus-none denoising tests, and recompute the Appendix E intervals at the song level? Also, how much do the exposure results change if you fix the artist aliases and cross-fold duplicates?

  3. How did you obtain the Billboard audio recordings, and what license or legal basis covers using them for Demucs separation and adapter fine-tuning? Please also explain how your recording entries map to the official McGill release.

1 comment
  1. We independently tuned the hyperbolic and Euclidean models using the same validation search budget on corrected artist/recording group splits, then trained fresh final models across five folds and three seeds. On full 11-stem MoisesDB, AUC was 0.766 for hyperbolic and 0.707 for Euclidean; the four-stem difference was smaller (0.872 versus 0.868). We also added an audio-only adaptation of the order embeddings of Vendrov et al. (2015), using the same features, splits, and seeds. Its AUC was 0.765 on full 11-stem MoisesDB and 0.909 on the four-stem setting, matching hyperbolic performance on the former and exceeding it on the latter. Count-matched superset retrieval scores (hyperbolic/Euclidean/order) were 70.6/68.1/58.0% on native MoisesDB and 90.7/89.7/85.7% on MoisesDB-4. A trial succeeds when a true containing mixture ranks above every non-containing candidate from the same recording window with the same number of stems. This ranking does not depend on post-training aperture calibration. These results support task-specific benefits rather than general hyperbolic superiority. On full 11-stem MoisesDB, direction scores were 8.1/8.5% at the trained aperture and 61.0/49.2% after validation calibration (hyperbolic/Euclidean). Calibration does not establish global transitivity, and this comparison does not isolate the effect of curvature. Historical test inspection remains disclosed. All geometry results average three seed-specific metrics, each computed by pooling the five test folds. Both MoisesDB versions use original recorded stems; MoisesDB-4 groups them into four families rather than using source separation. These experiments therefore do not establish how separation errors affect performance.

  2. We trained new adapters using the corrected artist and recording groups. We compared the same individual stem under two conditions: it appeared within other fine-tuning mixtures, or it was excluded entirely while other stems from the same song and recording window remained eligible for fine-tuning. The stem was never included as a standalone training example in either condition. Partial-versus-none AUC was 0.532 [0.511, 0.553] on Billboard. On MoisesDB-4, it was 0.616 [0.591, 0.642] after one quarter of training and 0.722 [0.680, 0.761] at the end, on native 11-stem MoisesDB, it was 0.653 and 0.730 [0.696, 0.767], respectively. Brackets indicate 95% confidence intervals obtained by resampling whole recording groups, keeping clips from the same group together. These intervals describe uncertainty conditional on the fitted models, they do not capture variation from retraining with different random seeds. These follow-ups use newly constructed exposure pools, so their AUCs do not isolate the effect of correcting the splits.

  3. we were able to acquire a license to work with McGill–Billboard audio data, but it is not possible to reveal specifics about the licensing details without breaking anonymity. For the camera-ready version of the manuscript, we can and will be clearer.

Vendrov, I., Kiros, R., Fidler, S., & Urtasun, R. (2015). Order-embeddings of images and language. arXiv preprint arXiv:1511.06361.

Reviewer_pmf6OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

In this paper the authors study training exposure and containment representation in the context of multi-stem musical data. In the case of training exposure, they fine-tune a reference diffusion model (via a PEFT adapter) over datasets MtrainM_{\textnormal{train}} obtained by mixing the stems in subsets of the tracks, keeping some partially mixed audios held-out. Given a candidate track cc composed by a number of stems (e.g., bass and drums), one can identify 4 cases: the track is in MtrainM_{\textnormal{train}} ('exact'), at least one stem in cc appears in a track in MtrainM_{\textnormal{train}} ('partial'), no stem in cc appears in a track in MtrainM_{\textnormal{train}} but some other stem (e.g., a vocal) from the same song (on the same window of cc) appears ('none') and the song cc belongs to does not appear in MtrainM_{\textnormal{train}} ('unseen'). The authors compute a "denoising score" considering the difference between the denoising loss on noisy candidates over the reference model and a finetuned version. They compare scores in 4 cases: (i) 'exact' vs 'unseen' candidate (ii) 'exact' vs 'none' (iii) 'partial' vs 'none' and (iv) 'exact' vs 'partial', computing the 'denoising score' with a finetuned model where the candidate has the specific role (e.g., when doing 'exact' vs 'unseen', one compares a finetune for which the candidate is 'exact' and one for which it is 'unseen'). The different results are combined in an AUC score. The main rationale is to study if one identifies better a training example, given a candidate, if the model has seen it verbatim than if the model has seen partial version of it. The second study trains an embedder for studying containment representation (estimate if u is a submix of v). They study both a hyperbolic and an euclidean representation of model containment. The authors then try to see if they can use the containment model to improve the first exposure task, with negative results.

Critical strengths
  • Both studies are interesting in the contest of multi-stem audio, given that main focus of membership inference is on full mixtures, which fail to cover many cases in which sub-mixtures of songs could be used to train generative models. Studying hyperbolic representations for containment estimation is also relatively novel, and interesting on its own, but it's intended applicative usage is not really explained in the paper. Also the two contributions, while sharing the theme of handling multi-stem audio data, they seem more like two independent contributions that don't share a common use-case.
  • The experimental sections looks extensive and well-posed.
Critical weaknesses
  • Main critical issue: the paper can be defined as AI slop. It is extremely incomprehensible, even if apparently it doesn't contain orthographic errors. This strongly signals either unsupervised use of AI to create the manuscript / insufficient effort from the authors to make the AI outputs comprehensible to a human reader. The worst part is the fact that many phrases have implicit assumptions and undefined terms, that are very difficult to get. The paper becomes readable only if read through an LLM. I used GPT6 Astra to be able to decypher many incomprehensible parts. A list of some of these:

    • Long list of em dashes suggests authors have not edited AI generated content (e.g., "bass–drums", Line 040; "source–output", Line 998; "Distance–set-difference" and "radius–count", Line 1126; "title–year", Line 263; and many other). Interestingly enough other occurrences of composed words contain standard dashes, for example "MoisesDB-4" at Line 259 and "four-family" at Line 260.
    • Second paragraph: Authors use “exposure label” without defining it. When they list what values the "“exposure labels” may have, the labels have to be between quotation marks or in italics to distinguish them from the rest of the text. Is giving these labels a practice already common or the paper suggests it? A "trained mixture" or "untrained mixture" is bad wording: mixtures are audios, they cannot be trained. A model is trained. This bad phrasing appears in different places in the paper, including the Figure 1.
    • Figure 1: Different from what in A? Using GPT6 to decipher this, I get that the training data is different since “bass” is not used for training in A individually, but only in the submixture “bass + drums”. The authors don't make this clear.
    • Line 054: "Exact membership and constituent exposure are different labels": Those are not labels, are properties. One should say something like: “Exposure to a clip’s components does not necessarily imply that the clip itself was used as a training example.”
    • Line 084: "differences between member and non-member populations can provide apparent membership signals": This is not clear at all. What do the authors mean here?
    • Line 092: "We distinguish exact-mixture membership from constituent exposure by comparing denoising scores for the same candidate across adapters, with the pretrained model as a reference.". At this point in the text this is the first mention of "adapters", GPT having context of all the paper suggests that later they are introduced, but they should have been introduced here since it's the first mention.
    • Line 105: "These approaches relate generated audio to training material, whereas our denoising probe evaluates candidate exposure during adaptation." Again the adaptation part is left very implict. What does one do there?
    • Line 161: "An exploratory crossover compares raw CLAP with the learned heads using reconstruction gain: the decrease in source–generation embedding distance from the pretrained to the fine-tuned generator (App. E)." This phrase reads extremely bad after the colon.
    • Line 174: "Only exact candidates are members." Members of what?
    • Line 176: "one stem is withheld entirely from a tenth of the training songs, and a balanced quarter of the remaining rendered nodes is held out at each mixture size, leaving every size eligible for training. Most partial candidates contain only exposed stems; they include untrained full mixes.": The recipe for creating the datates is very poorly described, with the last sentence without much meaning.
    • Paragraph "Paired exposure contrasts" (Line 207): the list of the cases is extremely dry and their meaning only drafted, i had to use colors to distinguish the cases. Also the last phrase "An increment beyond full constituent exposure requires restricting partial candidates to those whose every stem was exposed. Interpretation therefore depends on exposed-stem fraction" doesn't have much meaning.
    • Line 230: ss is used in this definition while being introduced at line 247. And so on...
  • Unclear goal and objective: The authors do not clearly explain what what usage / application the training exposure and the stem containment techniques and analysis have. They should make some use-cases simple such an artist that thinks a bass track has been used to train a generative model and wants to discover that. So given that the model could have trained on sub-mixtures with that bass track, the artist has to rely on a (hypotetical) method akin to A to understand that. And then the B method would be useful to find super-mixtures of the bass stem that could be analysed by A? This is not really clear, like what connects B and A? From the text it really seems that they are two separate contributions. As of now, B seems to be useful only for retrieval, independent of a generative model and not linked to A. The third idea of trying to see if method B improves similarity between generated outputs by the reference and the fine-tunes in A I think it’s ill-posed, since containment should not be used to evaluate similarity. This is also reflected in the failed experiment the authors had, so I would not add that as the link between A and B.

  • Insufficient citations. The authors should cite:

    • More extended literature in accompaniment generation, which is highly relevant for the multi-source analysis in the paper [1, 2, 3]
    • Other works that build encoders from multi-stem data such as [4, 5, 6]. The authors should discuss the difference between their hyperbolic model withe these other works in the Related work section.

References:

  1. Parker, J. D., Spijkervet, J., Kosta, K., Yesiler, F., Kuznetsov, B., Wang, J. C., ... & Le, D. (2024, April). Stemgen: A music generation model that listens. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (pp. 1116-1120). IEEE.
  2. Mariani, G., Tallini, I., Postolache, E., Mancusi, M., Cosmo, L., & Rodola, E. (2023). Multi-Source Diffusion Models for Simultaneous Music Generation and Separation. In The Twelfth International Conference on Learning Representations.
  3. Nistal, J., Pasini, M., Aouameur, C., Grachten, M., & Lattner, S. (2024). Diff-a-riff: Musical accompaniment co-creation via latent diffusion models. arXiv preprint arXiv:2406.08384.
  4. Ciranni, Ruben, et al. "Cocola: Coherence-oriented contrastive learning of musical audio representations." ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2025.
  5. Alain Riou, et al. ‘Stem-JEPA: A Joint-Embedding Predictive Architecture for Musical Stem Compatibility Estimation’. Proceedings of the 25th International Society for Music Information Retrieval Conference [San Francisco, California, USA and Online], 2024, pp. 625–33
  6. Garoufis, Christos, Athanasia Zlatintsi, and Petros Maragos. "Multi-source contrastive learning from musical audio." SMC (2023).
Questions for authors

Continuation of unclear parts:

  • Line 046: "making this ambiguity concrete", what ambiguity?
  • Line 046: "Rendering subsets of a song’s stems", “Rendering” is not clear what it means. It could both mean generating or simply mixing. I assume here is mixing but it’s not clear.
  • Line 055: An embedding of which model?
  • Line 098: "replication detection depends on the similarity measure.", Phrase is not clear, authors should specify that a similarity mesure is used for the comparison in first place.
  • Line 099: Authors should link "loss-based membership" to the previous subsection
  • Line 183: A score in diffusion modeling is a very specific object (namely ∇xlog⁡p(x)\nabla_x \log p(x)). Please say here and in next paragraphs that score here is used in the common sense.
1 comment

Thank you for the detailed comments. We now define adapters at first mention and describe the four exposure categories, mixture construction and fine-tuning pools in the Methods and Experimental Setup. In the text, we describe mixtures as included in or excluded from fine-tuning. The Methods define the notation and the denoising loss difference and explain how we compare the same clip across adapters.

The Introduction now uses an artist's bass recording to motivate the exposure question, and the Discussion states the model access the controlled test requires. We present containment retrieval as a separate task with a concrete use, finding mixtures that contain a given recorded part. The exploratory reconstruction test found no consistent benefit for exposure detection.

In Related Work, the new ``Learning from musical stems'' paragraph discusses all six suggested references. It explains how work on accompaniment generation motivates the exposure question and distinguishes our containment test from missing-stem retrieval, musical compatibility and coherence.

Comments on the paper

1 comment

Thank you for the constructive reviews. We made many revisions throughout the manuscript to clarify what our experiments measure, improve complicated sentence, and reran the main evaluations using corrected artist and recording group splits. The revision distinguishes exact-mixture membership from constituent exposure and presents stem containment retrieval as a separate task.

In the revised exposure experiment, we score the same recorded stem under two conditions. It either appears only within other fine-tuning mixtures or is excluded entirely. In both conditions, other stems from the same recording window remain eligible for fine-tuning. The denoising scores distinguish these conditions more clearly on MoisesDB (final AUCs of 0.730 and 0.722 in the 11-stem and four-family settings) than on Billboard (0.532). We clarify the model access required for this controlled comparison and limit the claim to exposure during fine-tuning.

For containment, we independently tuned the hyperbolic and Euclidean models with equal validation budgets and added an audio-only order-embedding baseline. The hyperbolic model achieved the highest count-matched superset retrieval scores in both MoisesDB settings, while order embeddings achieved the highest containment AUC in the four-family setting. The exploratory reconstruction experiment found no consistent improvement in exposure detection from the learned containment representations.

Once the discussion phase ends and title changes are permitted, we plan to retitle the paper From Stems to Songs: Training Exposure and Hierarchical Containment in Music to reflect the two questions studied in the revised manuscript.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.