{"id":"917b9682-4eb3-4045-9be8-89debc3533f2","arxiv_id":"2506.21298","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Adapter placement, architecture, and size strongly change generation quality and cost for MusicGen and Mustango on two non-Western genres, with late-layer, mid-sized (40M) adapters reported as the best tradeoff.","lead":"This paper tests different ways of fine-tuning two AI music models, MusicGen and Mustango, on Hindustani Classical and Turkish Makam music by adding small trainable adapter modules instead of retraining everything. It finds the best design depends on the musical style: convolutional adapters capture local ornament details, transformer adapters preserve long melodic structure, and around 40 million parameters is the best quality-versus-cost point.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 40M 'sweet spot' claim rests on single-run FAD/FD values with no error bars or significance tests, so the quantitative ranking may be noise; repeated-seed and bootstrap tests are needed.","rationale":"The reader's weakest assumption correctly identifies the missing statistical grounding of the central empirical claim. My stress-test converges on the same point: the 40M sweet spot and the architecture-by-genre rankings are load-bearing conclusions, and they are selected from single-point FAD/FD measurements with no uncertainty quantification. This is the most load-bearing concern because if measurement noise is comparable to the reported gaps, the headline trade-off—and the practical guidance that follows from it—does not survive. The concern is concrete and testable: repeated seeds and bootstrap intervals would either confirm or refute the ranking. I do not see an additional, more fundamental flaw. The internal inconsistency between the abstract's unconditional '40M' claim and the body's finding that Mustango peaks near 20M is real but secondary; it is a precision issue in the summary rather than an invalidation of the experimental direction. The failed middle-layer configurations are reported only qualitatively, but they are not the primary basis for the central claim. The code-release placeholder and missing baselines weaken reproducibility but do not by themselves falsify the argument. Therefore the correct verdict remains CONDITIONAL as the reader stated, with the condition being exactly the repeated, statistically grounded evaluation proposed here. No change to the reader's verdict is needed.","tokens_in":12749,"tokens_out":2807,"duration_ms":34946,"concrete_test":"Rerun the configurations that define the claimed optimum—at minimum 20M, 40M, and 70M for MusicGen-CNN/Transformer and Mustango-CNN/Transformer across both genres—with 3–5 independent training seeds, evaluating FAD and FD on the same fixed 400-prompt corpus. Report means with standard deviations or 95% bootstrap confidence intervals over prompts and seeds, and run a paired permutation test comparing each 40M configuration against its 20M and 70M neighbors. If the 40M advantage is not consistent across seeds or the intervals overlap substantially, the 'sweet spot' claim should be downgraded to a trend. As a secondary check, recompute scores with the canonical VGGish-based FAD to confirm the PANN-CNN14 backbone is not driving the ranking.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—'Across all configurations, we find that the 40M parameter scale offers the best FAD and FD scores' (§4.1)—rests on one FAD/FD value per configuration, computed on a fixed 400-clip prompt corpus with no repeated seeds, no error bars, and no significance test (§3.3.1). The gaps that select the 40M 'sweet spot' are small in FAD units (e.g., 10.0 for Hindustani_MG-C and 13.0 for Makam_MG-T), and the table of neighboring values is not shown, so it is impossible to tell whether the apparent optimum is larger than run-to-run or reference-set noise. The problem is compounded by using PANN-CNN14 features through the AudioLDM framework rather than the canonical VGGish-based FAD, which changes the metric's scale and sensitivity. If the dispersion of FAD across seeds or prompt subsets is on the order of a few points, the 40M ranking, and the genre-specific architecture ordering (CNN for Hindustani, Transformer for Makam), could be artifacts. This is not a disagreement with consensus; it is an internal validity issue: the quantitative basis for the headline trade-off is absent. The subjective evaluation (§3.3.2) does not rescue this because it only covers 40M models, uses two annotators, and reports no agreement statistic.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies adapter-based parameter-efficient fine-tuning (PEFT) of two music generation models, MusicGen (autoregressive transformer) and Mustango (latent diffusion), for two low-resource non-Western genres, Hindustani Classical and Turkish Makam. It compares three adapter architectures (linear/MLP, CNN, transformer) at parameter scales from roughly 2M to 70M, and two placement strategies (middle vs. late layers). The evaluation uses Fréchet Audio Distance (FAD) and Fréchet Distance (FD) computed once per configuration on a fixed 400-clip prompt corpus, a two-annotator arena-style subjective evaluation, and GPU-hour measurements. The central claims are that a 40M-parameter adapter scale offers the best quality/efficiency trade-off, that late-layer placement preserves generation quality while middle-layer placement breaks it, that CNN adapters suit Hindustani Classical while transformer adapters suit Turkish Makam, and that MusicGen is faster to train and produces cleaner audio while Mustango is more diverse and prompt-adherent but less stable.","tokens_in":12907,"tokens_out":3673,"duration_ms":43163,"significance":"If the empirical findings are reliable, the paper provides practical, cost-sensitive guidance for adapting large music models to underrepresented genres: roughly 20-45 GPU hours instead of full fine-tuning, with concrete architectural and placement recommendations. The release of datasets, models, and training code is a concrete reproducibility asset. However, the headline '40M sweet spot' and the architecture/placement conclusions rest on an evaluation instrument whose precision is not demonstrated: each FAD/FD value is a single measurement with no error bars, the subjective evaluation uses two annotators without agreement statistics, and middle-layer failures are reported without quantitative support. These gaps directly affect the load-bearing claims, so the paper's contribution is currently conditional on additional evidence rather than established.","major_comments":[{"comment":"The central '40M sweet spot' claim in Section 4.1 is based on a single FAD/FD value per configuration, computed on a fixed 400-clip prompt corpus with no repeated seeds, no bootstrap over prompt subsets, and no significance test. The differences that select 40M are small in FAD units (e.g., Hindustani_MG-C at 10.0, Makam_MG-T at 13.0, Mustango Hindustani MS-C at 6.4), and neighboring values are not shown, so it is impossible to rule out run-to-run or reference-set noise as the source of the ranking. Please provide confidence intervals or repeated-seed results, and report the full table of FAD/FD values across all sizes and architectures. Additionally, the metric is computed with PANN-CNN14 features through the AudioLDM framework rather than the canonical VGGish-based FAD; please justify this choice or report both versions, since it changes the metric's scale and sensitivity.","section":"Section 3.3.1 and Figure 2"},{"comment":"The placement contribution — late layers preserve quality, middle layers break generation — is reported only qualitatively for both MusicGen and Mustango. The text states that middle-layer placements led to 'complete breakdown' and 'significantly higher' FAD/FD values, but no numbers, table, or figure are provided for these failed configurations. This is load-bearing because the late-layer placement recommendation is a primary contribution. Please include the quantitative FAD/FD and/or listening outcomes for at least the middle-layer versus late-layer comparisons, with the same statistical treatment recommended above.","section":"Sections 2.4.1 and 2.4.2"},{"comment":"The subjective evaluation uses two annotators with no reported inter-annotator agreement, no per-annotator breakdown, and only 10 prompts per genre. The paper itself flags the annotator count as a resource constraint, but the resulting rankings (MG-C best for Hindustani, MG-T best for Turkish Makam) are then used as evidence for architecture conclusions. Please report an agreement statistic (e.g., Cohen's kappa) and the per-prompt or per-annotator results; without this, the subjective ranking is not distinguishable from chance-level variability.","section":"Section 3.3.2 and Figure 3"},{"comment":"The '40M sweet spot' is presented as an optimal trade-off 'for the given amount of data', but the experimental design does not vary the dataset size: each genre has one fixed training set size (208.58 hours for Hindustani, 157.01 hours for Turkish Makam). The claim that 40M is optimal for these data amounts is therefore an observation about two points, not a scaling-law statement. To support the general interpretation, the paper should either vary the training data size across at least two or three amounts or explicitly limit the claim to the specific dataset sizes studied and avoid the implication of a broader scaling relationship.","section":"Section 4.1 and Section 3.1"},{"comment":"The conjecture that Mustango's instability stems from adapting it without chord or beat conditioning is plausible but untested. It is presented as a hypothesis in the discussion, yet the conclusion section restates it as a finding. Please either test this hypothesis (e.g., by conditioning Mustango with chord and beat metadata during fine-tuning) or clearly mark it as speculation; as written, it risks being read as an empirical result.","section":"Section 4.2"}],"minor_comments":[{"comment":"The caption states 'across three adapter architectures at varying parameter scales' but does not describe what the colors, markers, or line styles represent; a legend or explicit caption should be added so the reader can map each curve to a specific adapter type and genre.","section":"Figure 2"},{"comment":"The sentence 'Mustango adapters (denoted MS- in the figure) require between 48 and 71 hours for MG-C and MG-T respectively' appears to contain a typo: it should refer to MS-C and MS-T, not MG-C and MG-T. Please also define the model names in the Figure 4 caption.","section":"Section 4.3"},{"comment":"The conclusion states 'Mustango showed high FAD scores' for Turkish Makam, but Section 4.1 reports Makam_MS-C FAD as 8.39, which is numerically low; please reconcile these statements or clarify whether the comparison is relative to MusicGen or to the Hindustani values.","section":"Conclusion"},{"comment":"For Mustango, the text says linear adapters are not used because 'MLPs require flattened 2D inputs', but a linear adapter could be applied pointwise to the channel dimension; this justification is not fully clear and would benefit from a one-sentence explanation of why pointwise MLPs are incompatible with the UNet's 3D latents.","section":"Section 2.3"},{"comment":"Reference [13] is titled 'Fréchet Audio Distance: A Reference-Free Metric for Evaluating Music Enhancement Algorithms'; FAD is computed against a reference distribution, so the word 'reference-free' in the title is potentially confusing, but the citation itself appears accurate. Please verify the exact title.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's core claims are plausible and the practical cost numbers are attractive, but the evaluation instrument is not yet rigorous enough to support the '40M sweet spot' and architecture/placement recommendations. The stress-test concern about single-run metrics is legitimate and should be addressed with error bars or repeated runs. The authors' own prior work [20,21] is used for motivation and for the evaluation framework; this is not improper, but the novelty of the present study over [21] could be made more explicit. The paper fits ACM MM scope. I recommend major revision, not rejection, because the missing statistical evidence is additive rather than contradictory to the claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a useful paper. It's the first systematic comparison I've seen of adapter architecture, placement, and size for PEFT on music generation models applied to two non-Western genres. The authors test linear, CNN, and transformer adapters on MusicGen and Mustango for Hindustani Classical and Turkish Makam, using a substantial Dunya-based dataset with careful train/test splits. The qualitative findings are credible and actionable: place adapters in late layers, use CNN adapters for ornament-heavy Hindustani music, transformer adapters for long-form Makam, and expect a mid-scale sweet spot. The GPU-hour reporting gives practitioners real budgets, and the divergence between FAD and human preference for Mustango is a genuinely interesting observation.\n\nThe soft spots are real and mostly quantitative. The headline claim that 40M parameters is the optimal scale rests on single-run FAD/FD values with no error bars or significance tests. The gaps between configurations are small—FAD 10.0 vs 13.0—so the ranking could be noise. Using PANN-CNN14 features through AudioLDM instead of the canonical VGGish-based FAD also makes the absolute numbers hard to compare with other work. The subjective evaluation has two annotators and no agreement measure; the paper says that's sufficient, but that's asserted, not demonstrated. There are no zero-shot, full fine-tuning, LoRA, or prefix-tuning baselines, so we don't know how adapters stack up against alternatives. Failed middle-layer placements are described but never quantified. And the code link is literally the string 'Github'—no repository, no commit hash. There's also an internal inconsistency: the abstract says 40M is optimal for both models, but Section 4.1 says Mustango peaks closer to 20M.\n\nNone of this is fatal to the paper's qualitative core. The systematic comparison is new, the dataset work is substantial, and the design guidance is plausible. But the quantitative sweet spot is not established by the evidence presented, and the missing artifacts matter for reproducibility.\n\nWho should read this? Practitioners working on PEFT for music generation, especially for under-represented genres. It's a good roadmap for what configurations to try first. It deserves a serious referee, but the revision should require repeated seeds with statistical tests, full reporting of failed placements, baseline comparisons, and an actual code release. My own verdict is conditional: the qualitative findings will likely survive, but the numeric ranking needs support.\n\nRecommendation: send to peer review, with a major-revision request. It's not desk-reject material.","headline":"Useful systematic adapter comparison for low-resource music PEFT, but the 40M sweet spot rests on single-run metrics.","tokens_in":13579,"tokens_out":2992,"would_cite":true,"duration_ms":27625,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Mid-sized 40M adapters at late layers give the best quality-per-cost for low-resource music generation.","keywords":["parameter-efficient fine-tuning","adapter-based learning","music generation","Hindustani Classical music","Turkish Makam music","MusicGen","Mustango","Fréchet Audio Distance"],"falsifier":"Re-run the size sweep (2M, 8M, 20M, 40M, 70M) for MusicGen-CNN and Mustango-CNN on both genres with at least five random seeds, computing FAD/FD with bootstrap confidence intervals over the 400-prompt corpus; if the 40M or 20M-40M confidence intervals overlap with the 20M and 70M intervals for either genre, the claimed sweet spot is not supported.","tokens_in":12392,"feed_emoji":"🎵","tokens_out":8277,"duration_ms":85321,"temperature":0.7,"pith_summary":"This paper asks how to adapt two large music-generation models, MusicGen and Mustango, to two underrepresented genres, Hindustani Classical and Turkish Makam, when training data and compute are limited. It claims that adapter-based fine-tuning has a reliable design sweet spot: adapters of about 40 million parameters, inserted at the end of late model blocks, produce the best balance of audio quality and training cost, while smaller adapters lack capacity and larger ones overfit or distort the base model. It further claims that the best adapter architecture is genre-dependent: convolution-based adapters capture local ornamentations and work best for Hindustani Classical, while transformer-based adapters preserve long-range structure and work best for Turkish Makam. The result matters because it gives practitioners a concrete recipe, around 20 to 45 GPU hours, for bringing non-Western genres into existing music-generation models without full fine-tuning.","feed_headline":"Mid-size 40M adapters hit the sweet spot for low-resource music AI","feed_subtitle":"Late-layer 40M adapters deliver best quality per GPU hour across MusicGen and Mustango for two underrepresented genres.","key_machinery":"The load-bearing mechanism is the adapter module: a small trainable block (a fully connected bottleneck, a CNN with dilated residual blocks and squeeze-and-excitation, or a transformer with multi-head self-attention) inserted into a frozen base model. The paper varies three design knobs, architecture, placement, and parameter scale from about 2M to 70M, and measures the effects with FAD and FD against a fixed 400-clip reference corpus. The 40M sweet spot emerges from two opposing forces: too few parameters cannot capture raga- or makam-specific structure, and too many parameters overfit the limited genre data and destabilize pre-trained representations; late-layer placement avoids corrupting the core musical features encoded in earlier layers, which middle-layer insertion destroys.","core_discovery":"The paper's central discovery is that adapter-based fine-tuning of music-generation models has a reliable design optimum: adapters of about 40 million parameters, placed at the end of late model blocks, give the best balance between audio quality and training cost, and this pattern holds for both base models studied. At 2M-8M parameters adapters lack the capacity to represent genre-specific melodic and rhythmic structure; at 70M they overfit the limited genre data and degrade fidelity, with the sharpest failure in MusicGen's linear adapters. Placement is equally decisive: inserting adapters in middle layers destroys generation entirely (beeping, static, no musical structure), while late-layer or block-end placement preserves the base model's core musical representations and lets the adapter add stylistic detail. Architecture preference is genre-dependent: convolution-based adapters win for Hindustani Classical, where ornamentations such as gamakas, meends, and murkis dominate, and transformer-based adapters win for Turkish Makam, where long modal structure and phrase development matter. The paper also finds that the diffusion-based Mustango produces more diverse and prompt-faithful audio but suffers from note instability, rhythm misalignment, and higher training cost, while the autoregressive MusicGen is more stable, faster to train, and rated higher subjectively despite being more repetitive.","pith_inferences":["A direct test of the capacity story: repeat the size sweep with 2x and 4x more training data; if the FAD/FD minimum moves above 40M, the sweet spot is set by data volume rather than by model architecture.","The placement result suggests that early and middle layers of these models carry genre-agnostic core musical representations; probing layer activations before and after adaptation could identify which layers are actually load-bearing.","Because the 40M ranking rests on single FAD/FD values, bootstrapping the 400-prompt reference set or repeating each configuration with several seeds would reveal whether the sweet spot is a real valley or measurement noise.","Feeding Mustango the chord and beat conditioning it was designed for, alongside the text-only prompts used here, may close its stability gap with MusicGen while keeping its diversity advantage."],"forward_implications":["A practitioner can adapt MusicGen to a new genre in about 20-22 GPU hours and Mustango in about 45 GPU hours at the 40M scale, instead of full fine-tuning over days.","The best configuration is genre-dependent: 40M CNN adapters for Hindustani Classical, 40M transformer adapters for Turkish Makam, and 40M CNN adapters for Mustango on both genres.","Middle-layer placement should be avoided entirely; placing adapters at block ends or late layers is a precondition for coherent generation.","Model choice involves a real trade-off: Mustango gives broader instrument, rhythm, and melody diversity with better prompt adherence, while MusicGen gives more stable, clearer, subjectively higher-quality audio with more repetition.","Linear adapters are the weakest option for MusicGen and are not usable with Mustango's continuous latent representation."],"supporting_citations":[{"why":"MusicGen: the autoregressive base model whose late-layer adapter placements and architecture variants are evaluated.","marker":"[5]"},{"why":"Mustango: the diffusion base model adapted for the same two genres, providing the contrast between autoregressive and diffusion behavior.","marker":"[22]"},{"why":"Houlsby et al. adapters: the PEFT methodology the paper builds on, establishing that frozen models can be adapted through small inserted modules.","marker":"[8]"},{"why":"Mehta et al. 2025: the prior PEFT music-generation study whose mixed results and unexplored configurations this paper systematically extends.","marker":"[21]"},{"why":"Dunya corpus: the data source for the Hindustani Classical and Turkish Makam recordings used for training and testing.","marker":"[26]"},{"why":"FAD: the reference-free Fréchet Audio Distance metric used to score generation quality against the test corpus.","marker":"[13]"},{"why":"AudioLDM framework: supplies the implementation used to compute FAD and FD scores.","marker":"[17]"},{"why":"PANN-CNN14: the feature-extraction backbone used to embed audio for FAD/FD computation.","marker":"[14]"}],"fun_headline_variants":["40M adapters: the size sweet spot for music AI fine-tuning","Late-layer adapters preserve music structure, size matters","Music genre dictates adapter design: conv vs transformer","Diffusion vs autoregressive: diversity vs stability in music AI","Fine-tune music AI with 40M late-layer adapters"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on the assumption that the FAD and FD scores can reliably rank adapter configurations: each configuration is scored once against a fixed 400-clip reference set with no error bars, repeated seeds, or statistical significance test, so the few-point gaps that separate the \"optimal\" 40M configuration from its neighbors could be measurement noise.","fun_headline_variants_meta":{"raw":{"variants":["40M adapters: the size sweet spot for music AI fine-tuning","Late-layer adapters preserve music structure, size matters","Music genre dictates adapter design: conv vs transformer","Diffusion vs autoregressive: diversity vs stability in music AI","Fine-tune music AI with 40M late-layer adapters"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000368,"raw_usage":{"total_tokens":2061,"prompt_tokens":1116,"completion_tokens":945,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":732,"completion_tokens_details":{"reasoning_tokens":861}},"tokens_in":732,"tokens_out":945,"duration_ms":10648,"temperature":1.0,"reasoning_tokens":861,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:28:39.637561+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the size sweep (2M, 8M, 20M, 40M, 70M) for MusicGen-CNN and Mustango-CNN on both genres with at least five random seeds, computing FAD/FD with bootstrap confidence intervals over the 400-prompt corpus; if the 40M or 20M-40M confidence intervals overlap with the 20M and 70M intervals for either genre, the claimed sweet spot is not supported.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"MusicGen: the autoregressive base model whose late-layer adapter placements and architecture variants are evaluated."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Mustango: the diffusion base model adapted for the same two genres, providing the contrast between autoregressive and diffusion behavior."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Dunya corpus: the data source for the Hindustani Classical and Turkish Makam recordings used for training and testing."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"FAD: the reference-free Fréchet Audio Distance metric used to score generation quality against the test corpus."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"AudioLDM framework: supplies the implementation used to compute FAD and FD scores."}],"review_version":1}