{"id":"1337f8cd-feb0-44ff-b0be-441b131d987d","arxiv_id":"2506.17818","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"CultureMERT adapts the MERT-95M music model to Greek, Turkish and Indian traditions via two-stage continual pre-training, improving non-Western auto-tagging by 4.9% on average while keeping Western performance intact.","lead":"A large music model that learned mostly from Western songs was retrained on Greek, Turkish and Indian music, and it became better at tagging those traditions without forgetting Western music. The paper shows this can be done with modest compute and that merging separately trained culture models works almost as well.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 2 contradicts the SOTA claim: CultureMERT's Carnatic AP (43.1±0.22) and CultureMERT-TA's AP (43.3±0.13) are below the cited prior SOTA of 43.9, so 'surpasses SOTA on all non-Western tasks' is unsupported as stated.","rationale":"The reader identifies the frozen EnCodec tokenizer as the weakest assumption, and the paper itself acknowledges that limitation in Section 6. That is a plausible risk, but it is not the most load-bearing concern about the central claim: the empirical results already show large improvements over MERT-v1 across non-Western tasks, and the CQT loss in Equation 4 provides an additional non-discrete training signal that can partly compensate for EnCodec limitations. By contrast, the SOTA claim is directly checkable from the paper's own Table 2, and the reported Carnatic AP values contradict the unqualified statement that CultureMERT and CultureMERT-TA 'surpass previous state-of-the-art results on all non-Western music tagging tasks.' This is an internal inconsistency, not merely an outside-consensus risk, and it affects a headline contribution. The issue is correctable with a more precise claim or a corrected evaluation, so the CONDITIONAL verdict remains appropriate; no movement to REJECT is warranted because the core adaptation result—consistent gains over the base model with minimal Western forgetting—still stands. I disagree with the reader's choice of weakest assumption because the EnCodec concern, while real, is secondary to a discrepancy that the manuscript itself contains.","tokens_in":15661,"tokens_out":6356,"duration_ms":65420,"concrete_test":"","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim that the adapted models 'surpass previous state-of-the-art results on all evaluated non-Western music tagging tasks' appears in the Abstract, in Contribution 4, and in Section 5. This claim is not supported by the paper's own Table 2 for average precision on Carnatic music. The table lists previous SOTA Carnatic AP as 43.9 [46], while CultureMERT obtains 43.1 with standard deviation 0.22 and CultureMERT-TA obtains 43.3 with standard deviation 0.13. Thus, on one of the four non-Western benchmarks and one of the two reported metrics, neither released model beats the cited prior state of the art. Because the SOTA comparison is a headline contribution, this discrepancy is load-bearing: either the claim must be restricted to ROC-AUC (where CultureMERT does exceed the cited SOTA of 77.0 for Carnatic), or to a majority or average of metrics, or the Carnatic evaluation setup differs from [46] in an undocumented way. The separate '4.9% average improvement over MERT-v1' claim is not affected by this issue, but it is not equivalent to the SOTA claim.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces CultureMERT-95M, a continual pre-training (CPT) adaptation of the MERT-v1-95M music foundation model to Greek, Turkish, and Indian music traditions, using 650 hours of multi-cultural data. The method is a two-stage CPT recipe with learning-rate re-warming/re-decaying and a Western replay fraction in Stage 1. The authors also construct CultureMERT-TA by merging single-culture adapted models with task arithmetic. Evaluation on six auto-tagging benchmarks (four non-Western, two Western) reports consistent gains over MERT-v1 on non-Western tasks, an average relative improvement of 4.9%, minimal forgetting on Western benchmarks, and a claim of surpassing prior state-of-the-art on all non-Western tasks. The paper also analyzes cross-cultural transfer and token-level similarity. The authors release both model checkpoints.","tokens_in":15917,"tokens_out":5927,"duration_ms":56511,"significance":"If the empirical results survive scrutiny, this is a useful contribution to the under-studied problem of cross-cultural music representation learning. The paper provides a concrete, computationally cheap adaptation recipe, a systematic comparison against the base MERT-v1 model on five seeds, and a comparison of continual pre-training against weight-space task arithmetic. The public release of the adapted models is a practical asset for the community. However, the headline claim of surpassing prior state-of-the-art on all non-Western tasks is contradicted by the paper's own Table 2 on Carnatic average precision, and several hyperparameters are selected using the same evaluation benchmarks that later appear in the main table. These issues currently weaken the central narrative, though they are fixable with a careful rewriting of the claims and an explicit separation of model selection from evaluation.","major_comments":[{"comment":"The claim in the Abstract, Contribution 4, and Section 5 that the adapted models 'surpass previous state-of-the-art results on all evaluated non-Western music tagging tasks' is contradicted by the Carnatic AP row of Table 2. The cited prior SOTA is 43.9 [46], while CultureMERT obtains 43.1±0.22 and CultureMERT-TA obtains 43.3±0.13. With standard errors of roughly 0.1–0.2 points, these are not ties within noise; on this metric both models are below the cited SOTA. The claim must be restricted to ROC-AUC, or to a defined aggregate, or the evaluation protocol against [46] must be documented and justified if a difference is claimed.","section":"§5, Table 2"},{"comment":"The final models are selected using the same benchmarks on which they are subsequently evaluated. The task-arithmetic scaling factor λ=0.2 is chosen from Figure 4, which plots ROC-AUC across all six test tasks, and the two-stage CPT recipe (stage budgets, Western replay fraction, warm-up ratios) is selected using Turkish-makam and MTAT in Table 1, both of which later appear in Table 2. For these datasets, the comparison against MERT-v1 is in-sample. Please introduce a validation split for recipe selection, or explicitly report which results are model-selection results and which are independent hold-out evaluations.","section":"§5.3, Figure 4 and §4.3, Table 1"},{"comment":"The '4.9% average improvement' is not defined precisely. It appears to be the relative change in the overall all-dataset average in Table 2 (66.1 to 69.3), not an average of per-dataset relative improvements (which is about 5.4% across all twelve cells and about 8% across the non-Western cells). Please state the exact aggregation formula, and report the non-Western and Western improvements separately so that the reader can see how the headline number relates to the 'diverse non-Western music auto-tagging tasks' mentioned in the Abstract.","section":"Abstract and §5"},{"comment":"The paper's own limitation that the frozen EnCodec tokenizer 'may be suboptimal for encoding culturally diverse musical languages' directly concerns the training targets in Eq. (2), because all continual pre-training optimizes prediction of these discrete tokens. The empirical gains mitigate this concern, but the manuscript would be stronger with a diagnostic showing that the tokenizer provides meaningful targets for the four non-Western traditions (for example, codebook utilization, token entropy, or reconstruction error on held-out non-Western audio). As written, the central adaptation mechanism rests on an assumption that is acknowledged but not tested.","section":"§6 and §3.1, Eq. (2)"}],"minor_comments":[{"comment":"The notation c∈C_k is confusing: C is defined as the number of codewords (C=1024) and K as the number of codebooks. Please write c∈{1,...,C} for a codebook index and clarify that the sum in Eq. (3) runs over codewords in the k-th codebook.","section":"§3.1, Eq. (3)"},{"comment":"The column header 'Turkish-makam MTA T' contains a spacing typo; it should read 'Turkish-makam MTAT'.","section":"Table 1"},{"comment":"The phrase 'we strike to balance plasticity' should read 'we strike a balance between plasticity'.","section":"§3.2"},{"comment":"The flame and snowflake symbols (🔥/❄) used to denote trainable and frozen components are not defined in the caption; please define them or use standard textual labels.","section":"Figure 1"},{"comment":"The evaluation is said to follow the MARBLE protocol 'under constrained settings', but the constraints are not specified. Please state which MARBLE elements are adopted and which are modified, since these details affect comparability with prior results.","section":"§4.2"},{"comment":"The paper states that 30-second segments are extracted from the training split of each non-Western dataset, but it does not explicitly confirm that the evaluation split is disjoint from the continual pre-training segments. This should be stated for reproducibility.","section":"§2"}],"recommendation":"major_revision","confidential_remarks":"The core comparison against MERT-v1 is credible and the released models are a useful contribution. My main concern is the unsupported SOTA claim on Carnatic AP and the selection-on-test issue; both are fixable in revision. I do not see a fatal flaw in the central training comparison."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuinely useful empirical study of continual pre-training and task arithmetic for adapting a music foundation model to Greek, Turkish, and Indian traditions, with released weights. The core comparison to MERT-v1 is clean and credible. But the abstract and Contribution 4 claim that the adapted models 'surpass previous state-of-the-art on all evaluated non-Western music tagging tasks,' and Table 2 doesn't support that: on Carnatic average precision, the prior SOTA is 43.9 while CultureMERT gets 43.1±0.22 and CultureMERT-TA gets 43.3±0.13. So one of the four tasks and one of the two metrics fails the claim. That is load-bearing because it appears in the abstract. The fix is easy: restrict the claim to ROC-AUC, where CultureMERT does exceed the cited SOTA, or say 'on average across metrics.' The 4.9% average improvement over MERT-v1 is unaffected and holds up.\n\nWhat's actually new: the two-stage CPT recipe (stable adaptation with limited GPU), the comparison against task arithmetic as a training-free alternative, and the token-level similarity analysis as a predictor of transfer. The paper ships the adapted weights, which is real value. Evaluation uses five seeds and reports standard deviations, which is more careful than most of this literature. The forgetting numbers on Western benchmarks are small, and the claim of minimal forgetting is consistent with Table 2.\n\nSoft spots, in rough order of weight: first, the SOTA overclaim above. Second, the main hyperparameters—the TA scaling factor lambda and the CPT stage budgets—are selected on the same evaluation benchmarks. That's a selection-on-the-test-set concern, though standard practice in this kind of paper; it means the reported numbers are tuned numbers, and the reader should know that. Third, the non-Western SOTA baselines mostly come from the authors' own prior work [46]. That's not a flaw by itself, but independent confirmation of the comparison would be good. Fourth, the frozen EnCodec tokenizer's suitability for non-Western audio is a real limitation, and the authors acknowledge it; I'd like to see at least a diagnostic of codebook usage on the non-Western data.\n\nWho this is for: MIR researchers working on cross-cultural representation, and anyone doing continual pre-training on small budgets. The paper deserves a serious referee; the overclaim should be fixed before publication, but the underlying adaptation recipe and released models are worth engaging with. I'd accept for review.","headline":"Useful, honest adaptation study whose 'SOTA on all tasks' claim is not supported by its own Table 2.","tokens_in":16457,"tokens_out":2272,"would_cite":true,"duration_ms":21735,"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":"Continual pre-training on 650 hours of Greek, Turkish, and Indian music adapts a Western-trained music foundation model to non-Western traditions, improving auto-tagging by 4.9% on average with almost no forgetting on Western benchmarks.","keywords":["music foundation models","continual pre-training","cross-cultural music representation","task arithmetic","model merging","MERT","non-Western music","auto-tagging"],"falsifier":"Take an unseen non-Western tradition absent from the 650-hour mix (for example gamelan or Arabic maqam) and compare CultureMERT against MERT-v1 on its auto-tagging benchmark: if there is no improvement, the model adapts only to the four trained cultures, not to world music generally. Alternatively, retrain or fine-tune EnCodec on the same non-Western audio and repeat the two-stage continual pre-training: if the gains disappear or shrink sharply, the frozen tokenizer was the limiting component.","tokens_in":15450,"feed_emoji":"🎵","tokens_out":10812,"duration_ms":90351,"temperature":0.7,"pith_summary":"This paper tries to establish that a music foundation model trained mostly on Western music can be adapted to other musical traditions without retraining from scratch and without losing its Western abilities. The authors continually pre-train MERT-v1-95M on a 650-hour mix of Greek, Turkish, and Indian classical and traditional music using a two-stage schedule with learning-rate re-warming. The resulting CultureMERT beats the original model on every non-Western auto-tagging task they test, with an average 4.9% improvement in ranking metrics (ROC-AUC and average precision), and loses only about 0.05% on Western benchmarks. They also show that merging single-culture adapted models with task arithmetic performs about as well, offering a training-free alternative.","feed_headline":"Music AI gains 4.9% on non-Western music after 650-hour retraining","feed_subtitle":"Two-stage retraining lifts a Western-trained model's non-Western tagging with almost no forgetting.","key_machinery":"The engine is the MERT-95M encoder (a HuBERT-style CNN plus 12-layer transformer) trained with a masked-language-modeling objective whose targets come from two frozen teachers: EnCodec's eight 1024-codeword residual vector-quantization codebooks for acoustic tokens, and constant-Q transform spectrogram reconstruction for pitch and harmonic content. The paper's contribution is the two-stage continual pre-training recipe around that engine: Stage 1 freezes the transformer and trains only the CNN feature extractor and codeword embeddings on 100 hours (20% Western replay) with a warm-up to $\\eta_{\\max}=5\\times10^{-4}$; Stage 2 unfreezes everything and trains on the full 650 hours with a gentler schedule peaking at $\\eta_{\\max}=5\\times10^{-5}$. This staging is what lets a small effective batch size adapt a large model without the stability gap or catastrophic forgetting. Task arithmetic supplies the alternative mechanism: each single-culture adapted model defines a task vector $\\tau_i=\\theta_i-\\theta_{\\mathrm{pre}}$, and merging sums scaled task vectors $\\lambda\\tau_i$ onto the base weights.","core_discovery":"The core claim is that staged continual pre-training, not architectural change, is enough to make a Western-biased music encoder useful across cultures. Using the same masked pseudo-label objective as MERT—predicting EnCodec acoustic tokens and reconstructing constant-Q transform (CQT) spectrograms—the authors first adapt the low-level feature extractor and codeword embeddings on 100 hours of multi-cultural audio with 20% Western replay, then unfreeze the transformer and train on all 650 hours. On Turkish makam, Hindustani, Carnatic, and Greek Lyra auto-tagging, the adapted model consistently outperforms the base MERT-v1 and also exceeds the previous state of the art on the non-Western tasks, while Western tagging performance stays essentially flat. A companion result is that task arithmetic—adding task vectors from separately adapted single-culture models back to the base with scaling $\\lambda=0.2$—matches the multi-culturally trained model on non-Western tasks and does even better on Western benchmarks.","pith_inferences":["If the token-similarity result holds, the same EnCodec token statistics could be used prospectively to choose which cultures to add to a pre-training mix before training, not just to explain transfer afterward.","The method's ceiling is likely set by the frozen Western-trained EnCodec tokenizer; fine-tuning or replacing that tokenizer on non-Western audio is a natural test of whether the 4.9% gain is a floor or a starting point.","The two-stage recipe should transfer to other music foundation models with masked-token objectives, including larger ones, but the hyperparameters and scaling factor would need re-tuning; the paper only demonstrates one model scale.","Because evaluation is limited to auto-tagging with a frozen probe, the gains may not carry over to generation, retrieval, or fine-tuned downstream tasks; that is an open extension, not something the paper establishes."],"forward_implications":["Non-Western auto-tagging improves on average 4.9% in ROC-AUC and average precision over MERT-v1, while Western benchmarks drop only 0.05%.","A 50-hour low-resource tradition (Lyra) still benefits from adaptation, and only the multi-cultural models beat the base model on Lyra.","Cross-cultural transfer is asymmetric: Carnatic adaptation transfers most consistently, and Turkish-makam and Carnatic models transfer strongly to each other.","Task arithmetic with scaling factor 0.2 gives a training-free route to a multi-cultural model that also slightly improves Western benchmarks over the base model.","Token distribution similarity between cultures predicts positive transfer, giving a data-selection signal for future continual pre-training mixes."],"supporting_citations":[{"why":"Supplies the base MERT-v1-95M model, its HuBERT-style architecture, and the masked pseudo-label pre-training objective that CultureMERT continually re-trains with.","marker":"[1]"},{"why":"Provides the learning-rate re-warming and re-decaying recipe that the two-stage continual pre-training strategy is built on.","marker":"[14]"},{"why":"Defines task arithmetic and task vectors, the weight-space merging method used to build CultureMERT-TA.","marker":"[37]"},{"why":"Defines the four non-Western auto-tagging tasks, the data splits, and the prior state-of-the-art numbers that the adapted models are compared against.","marker":"[46]"},{"why":"Supplies the frozen EnCodec neural codec whose eight residual vector-quantization codebooks generate the acoustic pseudo-label targets.","marker":"[48]"},{"why":"Provides the MagnaTagATune Western benchmark used to measure forgetting.","marker":"[39]"},{"why":"Provides the FMA-medium Western benchmark used to measure forgetting.","marker":"[40]"},{"why":"Supplies the Greek Lyra corpus, the smallest and lowest-resource non-Western culture in the study.","marker":"[41]"},{"why":"Supplies the Turkish-makam corpus, one of the three main non-Western traditions in the training mix.","marker":"[43]"},{"why":"Supplies the Hindustani and Carnatic Indian art-music corpora used for pre-training and evaluation.","marker":"[45]"}],"fun_headline_variants":["650-hour retraining lifts AI's non-Western music tagging by 4.9%","Continual pretraining improves non-Western music AI by 4.9%","Music AI gains 4.9% on non-Western tunes, forgets little","Task arithmetic matches multi-culture training for music AI","Staged retraining: 4.9% better on non-Western music, minimal forgetting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the frozen audio tokenizer (EnCodec), trained mostly on Western music, still turns Greek, Turkish, and Indian audio into discrete pseudo-labels useful enough for the model to learn from; if those codebook symbols miss microtonal pitches or unfamiliar rhythms, the adaptation signal is weak or biased.","fun_headline_variants_meta":{"raw":{"variants":["650-hour retraining lifts AI's non-Western music tagging by 4.9%","Continual pretraining improves non-Western music AI by 4.9%","Music AI gains 4.9% on non-Western tunes, forgets little","Task arithmetic matches multi-culture training for music AI","Staged retraining: 4.9% better on non-Western music, minimal forgetting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001308,"raw_usage":{"total_tokens":5362,"prompt_tokens":1006,"completion_tokens":4356,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":622,"completion_tokens_details":{"reasoning_tokens":4250}},"tokens_in":622,"tokens_out":4356,"duration_ms":26863,"temperature":1.0,"reasoning_tokens":4250,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:00:43.645322+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take an unseen non-Western tradition absent from the 650-hour mix (for example gamelan or Arabic maqam) and compare CultureMERT against MERT-v1 on its auto-tagging benchmark: if there is no improvement, the model adapts only to the four trained cultures, not to world music generally. Alternatively, retrain or fine-tune EnCodec on the same non-Western audio and repeat the two-stage continual pre-training: if the gains disappear or shrink sharply, the frozen tokenizer was the limiting component.","supporting_citations":[{"cited_title":"CultureMERT: Continual Pre-Training for Cross-Cultural Music Representation Learning","cited_arxiv_id":"2506.17818","evidence_quote":"Supplies the base MERT-v1-95M model, its HuBERT-style architecture, and the masked pseudo-label pre-training objective that CultureMERT continually re-trains with."},{"cited_title":"We also grate- fully acknowledge the Music Technology Group (MTG) at Universitat Pompeu Fabra for providing access to datasets used in this study","cited_arxiv_id":null,"evidence_quote":"Provides the learning-rate re-warming and re-decaying recipe that the two-stage continual pre-training strategy is built on."},{"cited_title":"A practitioner’s guide to real-world continual multimodal pretraining,","cited_arxiv_id":null,"evidence_quote":"Defines task arithmetic and task vectors, the weight-space merging method used to build CultureMERT-TA."},{"cited_title":"Patching open-vocabulary models by interpolating weights,","cited_arxiv_id":null,"evidence_quote":"Supplies the frozen EnCodec neural codec whose eight residual vector-quantization codebooks generate the acoustic pseudo-label targets."},{"cited_title":"Breaking language barriers: Cross-lingual continual pre-training at scale,","cited_arxiv_id":null,"evidence_quote":"Provides the MagnaTagATune Western benchmark used to measure forgetting."},{"cited_title":"Continual pre-training mitigates forgetting in language and vision,","cited_arxiv_id":null,"evidence_quote":"Provides the FMA-medium Western benchmark used to measure forgetting."},{"cited_title":"Predicting positive trans- fer for improved low-resource speech recognition us- ing acoustic pseudo-tokens,","cited_arxiv_id":null,"evidence_quote":"Supplies the Turkish-makam corpus, one of the three main non-Western traditions in the training mix."},{"cited_title":"Model soups: averag- ing weights of multiple fine-tuned models improves ac- curacy without increasing inference time,","cited_arxiv_id":null,"evidence_quote":"Supplies the Hindustani and Carnatic Indian art-music corpora used for pre-training and evaluation."}],"review_version":1}