{"id":"c17e5c1c-25b3-4d82-9f83-524de6840fef","arxiv_id":"2505.03314","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A diffusion U-Net for piano-roll music generation that combines Mamba and a learnable wavelet transform improves quality and chord controllability over Polyffusion and its own ablations.","lead":"This paper proposes a diffusion model for generating symbolic music in piano-roll form, adding Mamba and a learnable wavelet transform to the denoising U-Net. It reports that the model follows target chord progressions and scores above the Polyffusion baseline on objective and listening tests.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Chord F1 may be an evaluation loop: the same rule-based extractor [24] builds the conditioning labels and scores the generated music, with no independent validation of extractor or chord VAE.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing concern: the chord conditioning labels and the Chord F1 evaluation are produced by the same unvalidated rule-based extractor, creating a potential evaluation loop. This is the single most load-bearing issue because the controllability half of the central claim—'obtains the highest chord control accuracy of 0.541'—depends entirely on this metric being a faithful measure of musical chord adherence. The other concerns raised by the reader (no significance tests, narrow baseline set, unclear data split) are real but secondary: they affect the precision and generalizability of the comparison, whereas the chord-pipeline problem affects whether the controllability claim measures what it purports to measure. The paper does provide some independent support for the quality half of the claim through the listening test, which is why a full rejection is not warranted. A conditional verdict is appropriate: the claim can be salvaged if independent chord evaluation and VAE reconstruction checks confirm the reported numbers, but it should not be accepted as established while the evaluation loop remains unaddressed.","tokens_in":8420,"tokens_out":5733,"duration_ms":57858,"concrete_test":"Take a random subset of, say, 100 of the 879 chord-conditioned generations plus their paired ground-truth segments. Have two or more musician annotators label beat-wise chords, or use a second independently validated chord recognizer, while blinded to model identity and to the conditioning labels. Compute Chord F1 against these independent labels and compare the resulting ranking of Polyffusion/Proffusion-M/Proffusion-W/Proffusion-WM and the absolute F1 values. Also decode the 100 pretrained-VAE chord latents back to 36-D chord sequences and report the exact-match rate against the original extractor labels. If the ranking changes or the VAE reconstruction is poor, the controllability claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing assumption is that the reported Chord F1 measures real controllability. In Section III-B, beat-wise chords are extracted from ground-truth excerpts using the rule-based approach of [24] and encoded by a pretrained chord VAE [24]; in Section V, the same rule-based approach is used to extract chords from generated pianorolls and compare them with those conditioning labels. If this extractor is noisy or systematically biased, a model can earn a high F1 by fitting the extractor's heuristics without producing chord progressions a musician would recognize. The paper provides no validation of the extractor's accuracy on POP909, no reconstruction-accuracy check for the chord VAE, and no human or independent evaluation of whether generated samples follow the specified chords. The reported peak of 0.541 is therefore not interpretable as 'chord control accuracy' without additional evidence; it is agreement with one unvalidated labeler. This concern is load-bearing because the controllability half of the central claim rests almost entirely on this metric, while the listening test only rates general harmony, not progression adherence. A secondary issue compounds it: the test split is never defined, and the '879 chords from the POP909 test set' is inconsistent with the described 90/10 song-level split, leaving the exact evaluation population unclear.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents Proffusion-WM, a diffusion model that generates symbolic music as two-channel pianorolls. The denoising U-Net is augmented with a hybrid Transformer-Mamba block and learnable discrete wavelet transform blocks in the skip connections. Chord conditioning is implemented by extracting beat-wise chords from the training data with a rule-based tool, encoding each 8-bar chord sequence with a pretrained chord VAE, and injecting the resulting latent via cross-attention; classifier-free guidance is used at inference. The method is evaluated on POP909 against Polyffusion and two ablations (with and without wavelet, with and without Mamba) using overlapping-area distribution metrics, Chord F1, and a 30-participant listening test. The authors report that Proffusion-WM achieves the best average OA, the highest Chord F1 (0.541), and the highest ratings on all subjective metrics.","tokens_in":8660,"tokens_out":6674,"duration_ms":60320,"significance":"If the empirical claims are sustained, the paper makes a useful architectural contribution: it is the first application of learnable wavelet transforms to symbolic-music diffusion, it combines Mamba with attention inside a U-Net, and it evaluates the design with both ablations and a blind listening test. The code is released, and the consistent ordering of Proffusion-WM above all baselines on every subjective metric and on most objective metrics is encouraging. The main limitation is that the headline controllability result depends on a single unvalidated chord-extraction pipeline used on both sides of the evaluation, and the objective and subjective comparisons lack inferential statistics; these issues currently prevent the claims from being fully established.","major_comments":[{"comment":"The chord-controllability claim is supported by Chord F1, but the same rule-based chord extractor [24] is used both to create the conditioning labels in Section III-B and to extract chords from the generated pianorolls in Section V. There is no validation of the extractor's accuracy on POP909, no reconstruction-fidelity check for the pretrained chord VAE, and no human rating of whether generated samples follow the specified chord progressions. Under these conditions, a model that learns the extractor's heuristics can achieve a high F1 without producing progressions a musician would recognize, and the reported peak of 0.541 is not interpretable as \"chord control accuracy\" without additional evidence. This issue is load-bearing because the controllability half of the central claim rests almost entirely on this metric; the listening-test Harmony rating is not a substitute because it does not measure progression adherence.","section":"Section III-B and Section V"},{"comment":"The evaluation population is not defined consistently. Section IV-A states that POP909 is split at song level into 90% training and 10% validation, with no mention of a test set; Section V then says chords are extracted from the \"POP909 test set\" and that 879 chords are used to generate 879 samples. It is unclear whether the validation set is being reused as the test set, and how the count of 879 arises from a 10% song-level split. This ambiguity affects whether the reported comparisons are genuinely on held-out data and must be resolved for reproducibility.","section":"Section IV-A and Section V"},{"comment":"All objective results are single point estimates. No error bars, confidence intervals, or significance tests are reported for OA or Chord F1, and no information is given about how many generation runs or inference seeds were used. Consequently, the statement in the abstract that the method \"significantly outperforms\" the baseline is not supported by statistical evidence; differences such as 0.929 vs 0.939 in average OA may lie within sampling variability. Reporting bootstrap intervals over multiple generated sets, or at least a significance test, is necessary before making a comparative claim.","section":"Table I"},{"comment":"The listening test used 30 participants and only 12 generated pieces (three per model), and it reports means with standard errors but no inferential statistics. Because the same pieces are rated by multiple participants and the samples are not independent, the claim that Proffusion-WM outperforms all models on all subjective metrics requires at least a repeated-measures comparison or effect-size reporting. As it stands, the subjective ranking cannot be distinguished from rating noise.","section":"Table II and Section IV-D"}],"minor_comments":[{"comment":"The two channels of the (2,128,128) pianoroll are never defined; the paper should state whether they represent onset and sustain (as in the Polyffusion baseline) or some other pair of attributes.","section":"Section IV-A"},{"comment":"The definition of \\hat{V}_{\\lfloor N/2\\rfloor} is unclear (\"a vector with a center position value of two\"), and the summation index k is not fully specified; please clarify how the wavelet loss is computed in practice and how it prevents the wavelet convolution from degrading into a general convolution.","section":"Equation (15)"},{"comment":"The sentence \"the chords of the generated music pieces align well with those extract from the ground truth\" contains a typo ('extract' should be 'extracted'); additionally, the text says \"879 ground truth segments\" but it should be explicit that these are 8-bar excerpts, not whole songs.","section":"Section V"},{"comment":"Several hyperparameters needed for reproducibility are missing, including the number of training steps or epochs, model size, and the expansion factor r in the learnable wavelet node; please report these values.","section":"Section IV-B"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads like a workshop-paper version, and if it is intended as a full journal paper the evaluation needs to be materially strengthened. The chord-extractor circularity is the main obstacle; the authors should be asked to add an independent chord-adherence check, such as human annotation or a second validated chord estimator, and to provide proper statistical reporting for both objective and subjective metrics."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a competent architecture paper for controllable symbolic music generation. The Transformer-Mamba + learnable wavelet U-Net is a reasonable new combination, and the ablations are consistent. But the headline chord-control metric is suspicious because the same rule-based chord extractor is used to build the conditioning labels and to grade the outputs, and the test split is never properly defined. Worth reading, worth reviewing, but the evaluation needs serious tightening.\n\nWhat's new: the hybrid Transformer-Mamba block inside a diffusion U-Net, the learnable wavelet transform block in the skip connections, and the chord-conditioning pipeline via a pretrained chord VAE + classifier-free guidance. The self-supervised wavelet loss is a nice touch to stop the wavelet filters from degrading into ordinary convolutions. The paper reports consistent ablation ordering: removing either Mamba or the wavelet block drops most objective and subjective scores, and Proffusion-WM beats Polyffusion by a large margin in the listening test. Those are real results, if the evaluation is trustworthy.\n\nWhere it gets soft. The Chord F1 metric is potentially circular. Section III-B uses the rule-based extractor [24] to build chord conditions from the ground truth; Section V uses the exact same extractor to evaluate chords in generated samples. The paper gives no evidence that this extractor is accurate on POP909, and no check of the chord VAE's reconstruction quality. A model that learns to mimic the extractor's heuristics could score well without producing musically recognizable progressions. The listening test asks about \"harmony\" generally, not about adherence to the specified chords, so it doesn't rescue the controllability claim. This is load-bearing because the abstract's \"controllability\" claim rests almost entirely on that F1 number.\n\nAlso, the data split is murky. The paper says 90/10 train/validation, then talks about a \"test set\" with 879 samples. 10% of 64K would be around 6400, so 879 does not add up. Either the split is different or the test set is a subsample; either way the reader can't tell.\n\nThe comparison set is also thin: one external baseline (Polyffusion) plus two ablations. No error bars or significance tests on OA or F1, and the listening test (30 participants, 12 samples) has no inferential statistics. The mean differences are large, so they're probably real, but \"probably\" shouldn't be doing that much work.\n\nBottom line: this is a serious paper with a plausible architecture and consistent ablations. It deserves peer review, but the chord-control metric needs to be fixed with an independent labeler or a human chord-adherence test, and the split needs clarifying. If that happens, it's a solid contribution. I don't work in this subfield, so I wouldn't cite it soon, but I'd point a student working on symbolic music generation toward it.","headline":"A useful architecture paper with a potentially circular chord-control metric and an undefined test split; worth reviewing but the evaluation needs fixing.","tokens_in":9191,"tokens_out":3857,"would_cite":false,"duration_ms":34884,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A pianoroll diffusion model with a hybrid Transformer-Mamba U-Net and learnable wavelet transform claims to produce higher-quality, chord-controlled symbolic music than the Polyffusion baseline.","keywords":["symbolic music generation","diffusion models","pianoroll","Mamba","learnable wavelet transform","chord controllability","classifier-free guidance","U-Net"],"falsifier":"Take the released model, condition it on chords labeled by a second independent chord-annotation tool on POP909, and score the generated MIDI with that tool; if chord F1 drops to baseline levels or human listeners cannot reliably identify the target chord progressions, the controllability claim is not about musical chords but about matching that extractor's heuristics.","tokens_in":8222,"feed_emoji":"🎹","tokens_out":5387,"duration_ms":45596,"temperature":0.7,"pith_summary":"The paper proposes a diffusion model for symbolic music that treats MIDI as image-like pianorolls and conditions generation on chord progressions. Its central claim is that a U-Net denoiser built from hybrid Transformer-Mamba blocks and a learnable discrete wavelet transform yields better musical quality and chord controllability than the Polyffusion baseline. Evidence offered includes overlapping-area distribution scores, a chord F1 of 0.541, and listening-test ratings where the full model leads on every subjective metric. If the claim is right, the same architectural recipe can strengthen controllable symbolic generation without abandoning the pianoroll representation.","feed_headline":"Mamba-wavelet diffusion model beats chord-control baseline","feed_subtitle":"A pianoroll U-Net with learnable wavelets lifts chord F1 to 0.541 and top quality scores over Polyffusion.","key_machinery":"The load-bearing mechanism is the denoising U-Net: each encoder and decoder stage uses a Transformer-Mamba block in which features first pass through self-attention and then through a two-branch Mamba selection mechanism, while skip connections run through a Learnable Wavelet Node built from learnable 2D wavelet convolution, depth-wise filtering, and inverse wavelet transform. A self-supervised wavelet loss keeps the learned filters from degenerating into ordinary convolution. Chords are extracted by rule-based methods, encoded by a pretrained chord VAE, injected through cross-attention, and the model is trained with classifier-free guidance so that sampling can steer generation toward a target progression.","core_discovery":"The core discovery is that combining a Mamba-based selective state-space path with Transformer attention inside a U-Net, and routing skip connections through a learnable wavelet transform, improves both the statistical fidelity and chord-following behavior of diffusion-generated 8-bar pianorolls. The author would state it as: Proffusion-WM, our pianoroll diffusion model with a hybrid Transformer-Mamba block and learnable discrete wavelet transform, outperforms Polyffusion and both ablations on averaged overlapping area, achieves the highest chord control accuracy of 0.541, and receives the highest mean listening ratings across humanness, richness, rhythm, harmony, and overall preference.","pith_inferences":["Because the same rule-based chord extractor builds the conditioning labels and scores the output, the controllability margin may partly reflect the model learning the extractor's heuristics; an independent annotation would separate musical chord-following from extractor mimicry.","The Transformer-Mamba plus learnable-wavelet U-Net is a generic denoiser, so the recipe could transfer to other image-like generative tasks, such as spectrogram synthesis or MIDI velocity rendering, where sharp high-frequency edges matter.","A testable extension is to ablate the self-supervised wavelet loss separately from the wavelet transform itself; if the loss alone drives the gains, the transform's role is regularization rather than frequency decomposition.","The linear-complexity Mamba path may allow the model to scale to longer excerpts than 8 bars, and conditioning on whole-song chord progressions is a natural next experiment."],"forward_implications":["The hybrid Transformer-Mamba block can improve global feature modeling in pianoroll diffusion without the quadratic cost of pure self-attention, making longer musical contexts more feasible.","The learnable wavelet transform in skip connections helps the denoiser preserve sharp onset and offset boundaries, which shows up in higher overlapping-area scores for pitch range and inter-onset interval.","Chord-conditioned sampling with classifier-free guidance gives users a practical control signal: a target chord progression can steer the generated 8-bar pianoroll, with chord F1 rising from 0.494 for Polyffusion to 0.541 for Proffusion-WM.","The full model's listening-test ratings suggest that the objective OA and chord F1 gains translate into perceived humanness, harmony, rhythm, and richness."],"supporting_citations":[{"why":"Provides the Polyffusion baseline that the paper must beat, a pianoroll diffusion model with internal and external controls.","marker":"[15]"},{"why":"Supplies the rule-based chord extractor and pretrained chord VAE used to create conditioning labels and score chord F1.","marker":"[24]"},{"why":"POP909 is the dataset that yields the 64K 8-bar excerpts used for training and validation.","marker":"[25]"},{"why":"Classifier-free guidance is the sampling mechanism that mixes conditional and unconditional scores to steer generation toward target chords.","marker":"[18]"},{"why":"Mamba's selective state-space model is the basis of the hybrid Transformer-Mamba block, contributing linear-complexity sequence modeling.","marker":"[8]"},{"why":"Transformer attention is used both inside the hybrid block and for cross-attention conditioning on chord latents.","marker":"[7]"},{"why":"Defines the Wavelet Transform Block construction that the paper adapts into its Learnable Wavelet Node.","marker":"[23]"},{"why":"Provides the learnable discrete wavelet transform formulation whose high-frequency capture motivates the skip-connection design.","marker":"[22]"}],"fun_headline_variants":["Mamba-wavelet diffusion tops chord control at 0.541","Proffusion-WM: Mamba-wavelet diffusion beats Polyffusion","Diffusion with Mamba and learnable wavelet boosts music control","Hybrid Mamba-wavelet diffusion improves pianoroll generation","Learnable wavelet Mamba diffusion bests chord-control baseline"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The chord-control result depends on the rule-based chord extractor being accurate enough that both the training labels and the evaluation scores reflect real harmonic content.","fun_headline_variants_meta":{"raw":{"variants":["Mamba-wavelet diffusion tops chord control at 0.541","Proffusion-WM: Mamba-wavelet diffusion beats Polyffusion","Diffusion with Mamba and learnable wavelet boosts music control","Hybrid Mamba-wavelet diffusion improves pianoroll generation","Learnable wavelet Mamba diffusion bests chord-control baseline"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000611,"raw_usage":{"total_tokens":2785,"prompt_tokens":830,"completion_tokens":1955,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":446,"completion_tokens_details":{"reasoning_tokens":1866}},"tokens_in":446,"tokens_out":1955,"duration_ms":14078,"temperature":1.0,"reasoning_tokens":1866,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:55:06.445093+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the released model, condition it on chords labeled by a second independent chord-annotation tool on POP909, and score the generated MIDI with that tool; if chord F1 drops to baseline levels or human listeners cannot reliably identify the target chord progressions, the controllability claim is not about musical chords but about matching that extractor's heuristics.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Polyffusion baseline that the paper must beat, a pianoroll diffusion model with internal and external controls."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the rule-based chord extractor and pretrained chord VAE used to create conditioning labels and score chord F1."},{"cited_title":"Wang et al., ‘Pop909: A pop-song dataset for music arrangement generation’, in Proceedings of the 21nd International Society for Music Information Retrieval Conference , 2020","cited_arxiv_id":null,"evidence_quote":"POP909 is the dataset that yields the 64K 8-bar excerpts used for training and validation."},{"cited_title":"Ho and T","cited_arxiv_id":null,"evidence_quote":"Classifier-free guidance is the sampling mechanism that mixes conditional and unconditional scores to steer generation toward target chords."},{"cited_title":"Gu and T","cited_arxiv_id":null,"evidence_quote":"Mamba's selective state-space model is the basis of the hybrid Transformer-Mamba block, contributing linear-complexity sequence modeling."},{"cited_title":"Vaswani et al., ‘Attention is all you need’, in Advances in Neural Information Processing Systems , 2017","cited_arxiv_id":null,"evidence_quote":"Transformer attention is used both inside the hybrid block and for cross-attention conditioning on chord latents."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Wavelet Transform Block construction that the paper adapts into its Learnable Wavelet Node."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the learnable discrete wavelet transform formulation whose high-frequency capture motivates the skip-connection design."}],"review_version":1}