{"id":"e19bbdf7-b045-483f-b71e-65ba2bcd65ce","arxiv_id":"2505.07631","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"MixIT pre-training on unlabeled Free Music Archive audio improves MUSDB18 separation after fine-tuning, with uSDR gains of about 0.3 to 0.5 dB.","lead":"The authors test whether an unsupervised training method called MixIT can pre-train a music source separation model on unlabeled music, then fine-tune it on labeled data. They report consistent gains over training from scratch on the standard MUSDB18 test set.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Validation-guided channel selection confounds the MixIT pretraining comparison; an ablation with random channel choice is needed before attributing Table 1 gains to MixIT.","rationale":"The paper's central empirical claim is that MixIT pre-training on unlabeled music improves supervised music source separation. The most load-bearing condition for that claim is that the Table 1 comparison actually isolates the effect of MixIT. The channel-selection procedure in Section 4.3 violates that condition in a specific, testable way: the pretrained pipeline consumes MUSDB validation labels to pick 4 of 12 output heads, while the from-scratch models start with 4 heads and receive no such selection. Since the reader's weakest-assumption analysis identified the same confound, and since the paper includes no random-channel or selection-only ablation, the evidence is not yet sufficient to attribute the gains specifically to MixIT. This does not warrant rejection: the table results are consistent, and the frozen-separator experiment in Table 2 provides independent supporting evidence that the pretrained representation carries useful structure. But it does warrant a conditional verdict pending the proposed ablation. Other weaknesses, such as the absence of repeated-seed error bars and the untested causal hypothesis about stem ambiguity versus correlation, are real but secondary; the channel-selection confound directly threatens the main comparison.","tokens_in":10259,"tokens_out":9106,"duration_ms":100831,"concrete_test":"Recompute the medium model row using the FMA-large MixIT-pretrained weights, but choose the 4 fine-tuning output channels uniformly at random (repeat with 5 seeds) instead of by MUSDB validation alignment, keeping all other fine-tuning settings identical. If random selection reproduces M2's 9.46 cSDR / 9.30 uSDR within about 0.1 dB, the validation-guided selection is not the driver; if random selection falls back toward M1's 9.04 / 8.75, the Table 1 gain is confounded. As a second check, train a from-scratch 12-output medium model and apply the same validation-guided selection rule, to test whether the selection step alone, without MixIT pre-training, produces any improvement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The main-table comparison is intended to isolate MixIT pre-training, but Section 4.3 adds a supervised model-selection step to the pretrained pipeline that the from-scratch baseline does not receive. After MixIT pre-training with N=12 outputs, the authors use MUSDB validation ground truth to compute the optimal permutation of outputs against the four stems for each chunk, then keep the 4 output channels most frequently aligned with each stem. This gives the pretrained model a validation-label-driven channel-selection step (one of C(12,4)=495 possible choices) before fine-tuning. The from-scratch models in Table 1 have 4 output channels from initialization and receive no analogous selection step. Consequently, the reported gains, for example +0.42 dB cSDR and +0.55 dB uSDR for the medium model (M2 vs M1), could be partly due to selecting the most useful output heads on the validation set rather than to the MixIT objective itself. The paper provides no ablation with randomly chosen output channels, so the causal attribution that MixIT pre-training improves over training from scratch is not yet isolated. This is a concrete, testable confound rather than a generic reproducibility caveat.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper revisits the common belief that mixture-invariant training (MixIT) is unsuitable for music source separation due to high inter-source correlation. The authors pre-train a band-split TF-Locoformer (BS-Locoformer) on unlabeled Free Music Archive data with MixIT, using 12 output channels, then fine-tune on MUSDB18-HQ after selecting 4 output channels with MUSDB validation ground truth. They report that the MixIT-pretrained models outperform from-scratch training across small, medium, and large model sizes on cSDR and uSDR (Table 1), that fine-tuning only the encoder/decoder yields much better results than without pre-training (Table 2), and that larger pre-training corpora give modest gains (Table 3). The paper proposes that MixIT's difficulty in MSS is due to ill-posed stem definitions rather than source correlation.","tokens_in":10512,"tokens_out":5730,"duration_ms":48977,"significance":"If the reported gains are attributable to MixIT pre-training itself, the result is valuable: it demonstrates that large unlabeled in-the-wild music collections can substitute for expensive stem labels at the pre-training stage. The study has several strengths: it evaluates three model sizes, two metrics, per-stem results, four pre-training data scales, and includes a frozen-separator fine-tuning experiment that directly probes the quality of the MixIT-learned representation. The writing is clear and the central hypothesis is explicit. However, the comparison in Table 1 is confounded by a validation-guided output-channel selection step that the from-scratch baselines do not receive, and the paper does not directly test its correlation-versus-ill-posedness hypothesis; these issues prevent the central claim from being fully supported as stated.","major_comments":[{"comment":"Section 4.3, steps 1-4: The pre-trained pipeline selects 4 of 12 output channels using MUSDB validation ground truth, which gives the MixIT-pretrained model a supervised channel-selection step that the from-scratch models in Table 1 do not receive. Since this selection is one of C(12,4)=495 choices and is informed by the validation labels, the reported gains (e.g., +0.42 dB cSDR and +0.55 dB uSDR for M2 vs M1) may be partly due to this selection rather than to the MixIT objective. The authors should provide an ablation that fine-tunes the pre-trained model with randomly selected (or otherwise unsupervised-selected) output channels; without it, the causal claim that MixIT pre-training improves over scratch is not isolated.","section":"Section 4.3"},{"comment":"The paper frames the study as addressing whether MixIT is unsuitable for correlated sources, attributing its limited performance to ill-posed stem definitions rather than high inter-source correlation. However, no experiment varies the degree of inter-source correlation or the stem definition ambiguity; the argument in Section 3.2 is qualitative, and the pre-training results are equally consistent with both hypotheses. The claims should be reworded to focus on the demonstrated pre-training benefit, or supplemented with a direct test (e.g., measuring stem correlations in the training data and correlating them with MixIT behavior).","section":"Title, Abstract, Section 3.2"},{"comment":"All results are from a single training run with no repeated seeds, error bars, or statistical significance tests. Some differences are small (e.g., in Table 3, the Large and Medium pre-training differ by 0.01 dB in average uSDR), and the margins in Table 1 are modest relative to typical run-to-run variability. The authors should report results across at least two or three seeds for the main comparisons, or provide a clear justification for why a single run is sufficient.","section":"Tables 1-3"}],"minor_comments":[{"comment":"The claim that pre-training 'enables the model to achieve performance comparable to that of a larger model trained from scratch' is imprecise; for S2 vs M1 the average cSDR differs by 0.24 dB and for M2 vs L1 by 0.12 dB, which may not be 'comparable' depending on the reader's threshold.","section":"Section 4.4"},{"comment":"The loss function L is not defined at first use; the thresholded SNR loss is introduced later in Eq. (4). Consider defining it at the earlier point.","section":"Section 3.1, Eq. (2)"},{"comment":"Reference [24] is listed as 'Anonymous' and appears to be an unfinished citation; it should be completed or removed.","section":"References"},{"comment":"The manuscript promises to release code and model weights upon acceptance, but no code is available during review. Given the empirical nature of the claims, releasing code or providing a reproducibility appendix would strengthen the submission.","section":"Section 1"},{"comment":"The caption states that grey rows are not directly comparable, but the grey shading is not visible in the text; ensure the formatting is clear in the camera-ready version.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The main concern is the validation-based channel selection in Section 4.3; if an ablation with random channels still shows gains, the paper is likely publishable. The causal hypothesis about ill-posed stem definitions is not tested and should be softened. The paper is within scope for a signal-processing venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is the first paper I know that takes MixIT pretraining seriously for real music, with a modern architecture and in-the-wild FMA data, and it shows a consistent improvement on MUSDB. That is worth publishing after a fix. The frozen-separator result is the most persuasive bit: a MixIT-pretrained separator with only encoder/decoder fine-tuning lands at 7.27 dB uSDR, miles ahead of the random-init 3.68 dB. That suggests the pretraining genuinely learns something useful, not just an artifact of extra parameters.\n\nThe soft spot is the one the stress-test flags. After pretraining with 12 output channels, the authors pick the 4 channels to keep using MUSDB validation ground truth (Section 4.3). The from-scratch baseline has 4 channels from initialization and gets no equivalent supervised selection step. So Table 1's +0.4 dB cSDR / +0.5 dB uSDR is not a clean isolation of MixIT; part of it could be the validation-guided choice among 495 possible 4-channel subsets. This is a concrete, testable confound. A random-channel-selection ablation, or a from-scratch model trained with 12 outputs and then selected with the same rule, would settle it. Without that, the causal claim \"MixIT pre-training improves over training from scratch\" is overstated as written.\n\nMinor points: no repeated-seed runs or error bars, which is common for this field but still worth saying; and the paper never actually tests the correlation-versus-ambiguity hypothesis it opens with. The title asks a question the experiments don't answer. Also code and weights are only promised after acceptance, so no immediate reproducibility check.\n\nThe citation pattern is fine, and the data-scale experiment is a nice addition even if the trend is modest.\n\nWho is this for? People working on self-supervised audio or low-resource MSS. It deserves a serious referee. My recommendation: send it to review with a required ablation on channel selection, and ask the authors to soften the causal language so the claims match the evidence.","headline":"A useful, mostly solid empirical result with one real confound: validation-guided channel selection may be carrying part of the reported gains, so the title question remains open.","tokens_in":10961,"tokens_out":3014,"would_cite":true,"duration_ms":31463,"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":"MixIT pre-training on unlabeled music improves supervised music source separation across model sizes.","keywords":["music source separation","MixIT","unsupervised pre-training","mixture invariant training","unlabeled music data","MUSDB18","Free Music Archive","band-split TF-Locoformer"],"falsifier":"Repeat the fine-tuning pipeline on a MixIT-pretrained model whose four output channels are chosen at random instead of by validation ground truth; if the improvement over from-scratch training disappears, the central claim is not supported.","tokens_in":10090,"feed_emoji":"🎵","tokens_out":11738,"duration_ms":92016,"temperature":0.7,"pith_summary":"Music source separation normally needs isolated stems, which are costly to obtain, and the standard labeled dataset MUSDB18 has only about 150 tracks. This paper asks whether MixIT, an unsupervised training method that separates mixtures of mixtures without labels, can pre-train a separation model on unlabeled in-the-wild music so that later supervised fine-tuning starts from a better place. The authors claim it can: across three model sizes, fine-tuning a MixIT-pretrained band-split TF-Locoformer on MUSDB18 beats training from scratch, with average uSDR gains above 0.5 dB for the medium and large models. They also argue that the common belief that MixIT fails on music because musical sources are too correlated is untested, and that the real difficulty is the ill-posed definition of stems, an ambiguity that supervised fine-tuning can resolve. If this is right, unlabeled music data can substitute for expensive stem labels in the pre-training stage.","feed_headline":"MixIT pre-training on unlabeled music beats training from scratch","feed_subtitle":"Fine-tuned MixIT-pretrained models outperform from-scratch models on MUSDB18, cutting reliance on expensive stem labels.","key_machinery":"The central mechanism is mixture-invariant training (MixIT), a loss that trains a separator on a mixture of two unlabeled mixtures: the model must assign each output to one of the two input mixtures so that remixing reconstructs them, and the minimum over binary assignment matrices is the training loss. An efficient assignment solver (argmin over real matrices followed by projection to binary) is used instead of exhaustive search. The paper wraps this objective around the band-split TF-Locoformer, an encoder-separator-decoder separation model, first pre-training on unlabeled FMA music with 12 output channels and then selecting 4 channels for supervised fine-tuning on MUSDB18. The MixIT loss is what forces the unlabeled pre-training to learn source structure without any stem labels; the subsequent fine-tuning resolves the stem-definition ambiguity that MixIT cannot know.","core_discovery":"The paper's central claim is that MixIT-based pre-training on large-scale unlabeled music improves supervised music source separation, and that the gains are attributable to MixIT learning source structure rather than to the assumption that it should fail because of high inter-source correlation. Concretely, after pre-training band-split TF-Locoformer models on the Free Music Archive with MixIT and fine-tuning on MUSDB18-HQ, every model size beats its from-scratch counterpart: for the medium model average cSDR rises from 9.04 to 9.46 dB and average uSDR from 8.75 to 9.30 dB, and for the large model average uSDR rises from 9.21 to 9.72 dB. Supporting evidence comes from freezing the separator and fine-tuning only the encoder and decoder, where the pre-trained model still separates far better than a randomly initialized one, and from data-scale experiments where larger unlabeled corpora give further improvements. The authors interpret these results as showing that MixIT's earlier poor showing in music reflects the ill-posed nature of stem definitions (for example, the 'other' stem is a group of instruments), not an intrinsic failure on correlated sources.","pith_inferences":["If ill-posed stem definitions are the real obstacle, MixIT pre-training should transfer more cleanly to separation tasks with semantically clear single-instrument stems; evaluating on a per-instrument labeled corpus would test that implication directly.","A random-channel-selection ablation would isolate whether MixIT pre-training itself, rather than the supervised selection of 4 of 12 output channels on validation ground truth, produces the reported gains.","The small gains from a tenfold increase in pre-training data suggest model capacity is the current bottleneck; repeating the data-scale curve with a larger separator should show steeper returns."],"forward_implications":["If the claim is correct, unlabeled in-the-wild music can replace expensive stem labels during pre-training, lowering the cost of building music source separation systems.","Pre-training effectively increases model capacity: the MixIT-pretrained small model outperforms the from-scratch medium model, and the pretrained medium model reaches the level of the from-scratch large model on average uSDR.","Freezing the separator and fine-tuning only the encoder and decoder still yields strong separation, showing the MixIT pre-trained separator has learned source structure rather than only task-agnostic features.","Scaling unlabeled pre-training data improves fine-tuned performance across all FMA subset sizes, although the gains become modest beyond the large subset."],"supporting_citations":[{"why":"Defines mixture-invariant training and the mixture-reconstruction loss that the paper adapts for unsupervised pre-training.","marker":"[14]"},{"why":"Supplies the Free Music Archive unlabeled corpus used for MixIT pre-training.","marker":"[23]"},{"why":"Provides MUSDB18-HQ, the labeled fine-tuning dataset and test set for evaluation.","marker":"[6]"},{"why":"Introduces the band-split TF-Locoformer architecture that is pre-trained and fine-tuned.","marker":"[25]"},{"why":"Contributes the efficient binary assignment solver used to compute the MixIT loss.","marker":"[20]"},{"why":"Provides the source-activity detection for silence removal and the band-split encoder design the model builds on.","marker":"[4]"},{"why":"Pac-HuBERT, the unsupervised pre-training baseline whose reported improvements the paper compares with its MixIT results.","marker":"[7]"}],"fun_headline_variants":["MixIT pre-training rescues music separation from correlated sources","Unlabeled music pre-training with MixIT beats from-scratch training","MixIT isn't useless for music: pre-training wins","Surprise: MixIT pre-training improves music source separation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison depends on the premise that MixIT pre-training, not the extra step of choosing which output channels to keep using ground-truth validation data, causes the improvement; an ablation with randomly chosen channels would test whether that premise holds.","fun_headline_variants_meta":{"raw":{"variants":["MixIT pre-training rescues music separation from correlated sources","Unlabeled music pre-training with MixIT beats from-scratch training","MixIT isn't useless for music: pre-training wins","Surprise: MixIT pre-training improves music source separation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000218,"raw_usage":{"total_tokens":1482,"prompt_tokens":1027,"completion_tokens":455,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":643,"completion_tokens_details":{"reasoning_tokens":385}},"tokens_in":643,"tokens_out":455,"duration_ms":4227,"temperature":1.0,"reasoning_tokens":385,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:11:09.286505+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the fine-tuning pipeline on a MixIT-pretrained model whose four output channels are chosen at random instead of by validation ground truth; if the improvement over from-scratch training disappears, the central claim is not supported.","supporting_citations":[{"cited_title":"Unsupervised sound separation using mixture invariant training,","cited_arxiv_id":null,"evidence_quote":"Defines mixture-invariant training and the mixture-reconstruction loss that the paper adapts for unsupervised pre-training."},{"cited_title":"Task- aware unified source separation,","cited_arxiv_id":null,"evidence_quote":"Introduces the band-split TF-Locoformer architecture that is pre-trained and fine-tuned."},{"cited_title":"Sparse, efficient, and semantic mixture invariant training: Taming in- the-wild unsupervised sound separation,","cited_arxiv_id":null,"evidence_quote":"Contributes the efficient binary assignment solver used to compute the MixIT loss."},{"cited_title":"Music source separation with band-split rnn,","cited_arxiv_id":null,"evidence_quote":"Provides the source-activity detection for silence removal and the band-split encoder design the model builds on."},{"cited_title":"Pac-hubert: Self-supervised music source separation via primitive auditory clustering and hidden-unit bert,","cited_arxiv_id":null,"evidence_quote":"Pac-HuBERT, the unsupervised pre-training baseline whose reported improvements the paper compares with its MixIT results."}],"review_version":1}