{"id":"0f2ed28e-bde8-4c06-9fdd-1ed638241ae4","arxiv_id":"2607.26440","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Adding explicit note-event tokens to the decoder and masking pitch-invalid TAB positions during decoding improves guitar tablature accuracy over the Fretting Transformer on DadaGP and especially on the small François Leduc dataset.","lead":"The paper presents a T5-based system that turns MIDI guitar performances into tablature by adding explicit note-on/note-off tokens to the decoder output, plus a decoding rule that forbids string-fret choices that produce the wrong pitch. On the large DadaGP set it beats a retrained Fretting Transformer, and on a tiny 79-song set it avoids the baseline's collapse.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing ablation: gains attributed to note-event tokenization are confounded with regularization and architecture; the causal claim is not yet supported.","rationale":"The reader's weakest assumption concerns undefined alignment of 'corresponding' TAB tokens across different tokenizations. That is a real validity threat, and I partially agree with it. However, the single most load-bearing gap for the central claim is the missing ablation that isolates tokenization from the rest of the proposed recipe. The paper's title and the reader's strongest claim attribute the large accuracy gains to note-event tokenization, but the experimental comparison changes at least two factors simultaneously: decoder tokenization and training regularization (and possibly architecture/hyperparameters, since the baseline's configuration is not reported). The Leduc-only result is especially telling: a small training set, a baseline that collapses, and a regularized model that does not collapse—this is precisely the effect dropout and weight decay are expected to have. Without a control using the same T5 architecture and regularization but the FT-style target vocabulary, there is no evidence that tokenization, rather than regularization, drives the gains. I am not claiming the paper is wrong; I am claiming the causal attribution is underdetermined. The proposed concrete test would resolve it. Because the reader's verdict is already CONDITIONAL and my concern sharpens the conditions rather than changing the overall assessment, I recommend leaving the verdict unchanged.","tokens_in":8170,"tokens_out":10165,"duration_ms":105382,"concrete_test":"Run the proposed T5 encoder-decoder with the exact same configuration, regularization (dropout 0.1, weight decay 0.1), data splits, and 300-epoch schedule, but replace the target tokenization with the Fretting Transformer's vocabulary (TIME_SHIFT_TICKS and TAB<string,fret> only). Evaluate Pitch Accuracy and Tablature Accuracy on both DadaGP and Leduc test sets. If this ablated model performs at the level of the full model (roughly 90% on DadaGP and 98% on Leduc), tokenization is not the causal factor; if it drops toward the FT baseline levels (roughly 86% and 0.5%), the tokenization effect is confirmed. Report at least 3 seeds to bound noise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-2 specifies the proposed method's architecture (T5, d_model=128, dff=1024, 3+3 layers, 4 heads) and regularization (dropout 0.1, weight decay 0.1). The paper says only that the Fretting Transformer baseline was 'retrained' on the same data, not that it used the same architecture, regularization, or hyperparameters. The two conditions therefore differ in at least decoder tokenization and training regularization. This matters especially for the Leduc-only result (Table II): the baseline collapses to 0.49% tablature accuracy while the proposed model reaches 98.06%. That pattern is the canonical signature of overfitting controlled by dropout/weight decay. Consequently, the reported improvements—and the paper's headline attribution to explicit note-event tokenization—may be caused by regularization rather than by the token representation. A control model with identical T5 configuration and regularization but the FT-style TIME_SHIFT+TAB-only target tokenization is absent. Without it, the central claim that 'the tokenization is the cause' is underdetermined.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a sequence-to-sequence framework for MIDI-to-guitar-tablature transcription. The main technical changes are (i) a decoder target tokenization that interleaves explicit NOTE_ON_PITCH and NOTE_OFF_PITCH tokens with TAB<string,fret> and TIME_SHIFT_TICKS tokens, and (ii) a pitch-validity constrained decoding procedure that masks TAB candidates whose implied pitch does not match the source pitch, as defined in Eq. (2). The method is evaluated on DadaGP and the François Leduc dataset against a retrained Fretting Transformer baseline. The reported results show large gains in token, pitch, and tablature accuracy, especially when training directly on the small Leduc dataset, and the constrained decoding analysis is used to separate pitch-invalid errors from pitch-valid string–fret ambiguity. The authors also release the code.","tokens_in":8475,"tokens_out":5220,"duration_ms":107515,"significance":"If the causal claims are supported, the paper would make a useful contribution to symbolic guitar tablature transcription: it identifies target-side tokenization as a design factor, proposes a constrained decoding diagnostic that cleanly removes pitch-invalid outputs, and includes a two-scale evaluation. The explicit code-release commitment and the retrained baseline are strengths. However, the current experimental design does not isolate the proposed tokenization from changes in architecture and regularization, and the evaluation metrics are not defined precisely enough to support cross-method comparisons. The constrained-decoding analysis is potentially informative but needs a precise statement of what is masked versus copied. With additional controls and clarifications, the paper could be a solid contribution.","major_comments":[{"comment":"The central claim that explicit note-event tokenization causes the accuracy gains is underdetermined because the comparison is confounded. Section III-2 specifies the proposed T5 configuration (d_model=128, dff=1024, 3+3 layers, 4 heads) and regularization (dropout 0.1, weight decay 0.1), but the Fretting Transformer baseline is only described as 'retrained' on the same data. The two conditions therefore differ in at least two variables: decoder tokenization and training regularization/architecture. The Leduc-only result in Table II (0.49% vs 98.06% tablature accuracy) has the hallmark of an overfitting-controlled-by-regularization effect, not necessarily a tokenization effect. Please add a control model with the same T5 backbone and regularization but the FT-style TIME_SHIFT+TAB-only target tokenization, or otherwise factor out the regularization contribution.","section":"§III-2, §IV-C, Table II"},{"comment":"Token Accuracy is defined as token-level exact match against the 'ground-truth target sequence,' but the two compared systems have different target tokenizations: the Fretting Transformer outputs only TIME_SHIFT_TICKS and TAB tokens, while the proposed method also outputs NOTE_ON_PITCH and NOTE_OFF_PITCH. A predicted TAB token in one system will not sit at the same token index as the 'corresponding' ground-truth TAB token in the other. Pitch Accuracy and Tablature Accuracy are defined by matching 'the corresponding ground-truth TAB token,' but no alignment or parsing algorithm is given. Because of this, the reported token-accuracy gaps (e.g., 80.57% vs 91.13% on DadaGP) and the pitch/tab accuracy numbers are not verifiable. Please specify the exact alignment procedure—e.g., note-event-level alignment after parsing, sequence alignment, or position-wise matching—and provide pseudocode.","section":"§IV-B, Tables I–III"},{"comment":"The constrained decoding experiment is not fully specified. Section III-4 says that logits for TAB tokens outside T(p) are set to -inf, and that this 'preserves the original timing and note structure from the input.' Yet Table III reports Timing error and Pitch error of 0.00% under constrained decoding for both models. Masking TAB tokens alone cannot eliminate timing errors caused by incorrectly generated TIME_SHIFT_TICKS. Please state explicitly which tokens are generated by the model and which are copied or provided from the input/ground truth. If timing and note boundaries are fixed from the input, then constrained decoding is an oracle-augmented diagnostic upper bound rather than a fully autoregressive decoding strategy, and the Table III numbers should be labeled and interpreted accordingly. Also specify how the 'source pitch at each TAB prediction step' is determined for the FT bas","section":"§III-4, §IV-E, Table III"},{"comment":"All experiments appear to be single runs, and the Leduc test set contains only 8 songs. The Leduc-only comparison (FT 0.49% vs. Ours 98.06%) is therefore sensitive to seed and initialization. Please report mean and standard deviation over at least three seeds, or at minimum provide seed/trial details. In addition, the claim that the FT baseline was tuned and 'severely overfits even after hyperparameter tuning' needs details of the hyperparameter search; otherwise the comparison in the small-scale regime is not reproducible.","section":"§IV-C, Table II"}],"minor_comments":[{"comment":"Typo: 'In addtion' should be 'In addition'.","section":"Abstract/Introduction"},{"comment":"The Difficulty Score is referenced but not defined in the evaluation section; please provide the formula or a precise citation to [5].","section":"§IV-B"},{"comment":"The architecture notation 'Nenc = 3,N dec = 3' should be typeset consistently, e.g., N_enc = 3, N_dec = 3.","section":"§III-2"},{"comment":"The standard-tuning open-string pitches are denoted o=[40,45,50,55,59,64], but the relationship between this vector and tuning[s] in Eq. (2) could be stated more explicitly.","section":"Eq. (2)"},{"comment":"The rule-based post-processing baseline is reproduced 'according to the paper description' because source code is unavailable. Since this is a reproduction, please note potential implementation differences and their possible effect on the comparison.","section":"§IV-E"}],"recommendation":"major_revision","confidential_remarks":"The paper is in scope and the constrained-decoding diagnostic is a reasonable idea, but the central attribution to tokenization needs a proper ablation. The metric definitions and constrained-decoding protocol must be tightened before publication. I would encourage the editor to request a revision rather than reject, because the issues are addressable with additional experiments and clarifications."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hey,\n\nThe paper has two genuinely new pieces: the decoder tokenization that adds NOTE_ON/OFF_PITCH tokens alongside TAB tokens, and applying the pitch-validity mask during decoding rather than as post-hoc correction. The small-data result on Leduc is striking, and the constrained-decoding diagnostic is a nice way to separate pitch-invalid errors from string-fret ambiguity.\n\nBut I wouldn't take the headline claim at face value yet. The evaluation compares the proposed T5 model against a retrained Fretting Transformer, and the two conditions differ in at least tokenization and regularization. The paper specifies its own T5 configuration (d_model=128, dff=1024, 3+3 layers, dropout 0.1, weight decay 0.1) but doesn't say what hyperparameters the FT baseline used. On Leduc-only training, the baseline collapses to 0.49% tablature accuracy while the proposed model gets 98.06%—that pattern is exactly what you'd expect if the baseline is overfitting and the proposed model's dropout/weight decay is controlling it. Without a control model that uses the same T5 architecture and regularization but keeps the FT-style TIME_SHIFT+TAB-only tokenization, you can't attribute the gain to the token representation.\n\nThere's also a metrics problem. Token Accuracy is computed against each model's own target tokenization, so the FT's 80.57% and the proposed 91.13% on DadaGP aren't measuring the same thing. The definition of \"corresponding ground-truth TAB token\" for pitch and tablature accuracy is never spelled out—these are sequence-to-sequence outputs of different lengths and token types, so you need an explicit note-event alignment. Without that, the numbers might not be comparable.\n\nThe paper also gives single runs with no error bars, and the Leduc test set is 8 songs. That's small enough that one or two songs could shift the percentages noticeably.\n\nWhat's good: the constrained decoding is a clean diagnostic, and the error decomposition into timing/pitch/position is a sensible way to see where the remaining errors are. The paper is clearly written and the method section is reproducible in principle.\n\nBottom line: this deserves a serious referee, but it needs major revisions. I'd ask the authors to add the missing tokenization control, define the alignment procedure, report multiple runs, and release the code. If that comes together, the contribution would be solid for the MIR community.","headline":"New tokenization and constrained decoding for guitar tab, but the evaluation doesn't yet prove the tokenization is what helps.","tokens_in":8908,"tokens_out":2343,"would_cite":false,"duration_ms":22708,"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":"Adding explicit note-event tokens to the decoder output improves MIDI-to-guitar-tablature transcription accuracy, especially on small datasets, and enforcing pitch-valid string–fret choices during decoding removes the largest class of error","keywords":["guitar tablature transcription","MIDI-to-tab","note-event tokenization","constrained decoding","pitch validity","sequence-to-sequence","T5","small-data training"],"falsifier":"Take the Fretting Transformer's and the proposed model's outputs on the Leduc test set, align both to ground-truth note events by matching NOTE_ON boundaries (or by a note-level alignment that ignores tokenization), and recompute pitch and tablature accuracy. If the baseline's 0.49% accuracy jumps to near the proposed model's 98% once aligned note-event-aware, the reported advantage is an artifact of tokenization mismatch rather than better generation.","tokens_in":8115,"feed_emoji":"🎸","tokens_out":3932,"duration_ms":42142,"temperature":0.7,"pith_summary":"This paper argues that the way notes are tokenized on the decoder side matters more than previously acknowledged for MIDI-to-tablature transcription. By inserting NOTE_ON_PITCH and NOTE_OFF_PITCH tokens into the generated sequence alongside the usual string–fret TAB tokens, the model is forced to represent note boundaries and pitch explicitly, which improves token, pitch, and tablature accuracy over the Fretting Transformer baseline on DadaGP. The most striking gain appears on the small François Leduc dataset, where the baseline collapses to 0.49% tablature accuracy while the proposed method reaches 98.06% when trained on Leduc alone. The paper also introduces pitch-validity constrained decoding, which masks any string–fret pair that cannot physically produce the target pitch, showing that most of the baseline's errors are pitch-invalid and that the remaining challenge is choosing among pitch-valid positions.","feed_headline":"Note-event tokens raise MIDI-to-tab accuracy by up to 10 points","feed_subtitle":"Adding NOTE_ON/OFF tokens to decoder output beats Fretting Transformer and rescues small-dataset training.","key_machinery":"The central objects are the explicit note-event decoder tokenization and the pitch-valid set T(p) = {(s, f) | tuning[s] + f = p, s ∈ [1,6], f ∈ [0,24]}. The tokenization exposes note boundaries and pitch to the decoder during generation, giving the model a clear target structure. The constrained decoding mechanism zeroes out logits for any string–fret pair outside T(p) at each TAB prediction step, preventing pitch-invalid TAB tokens from being emitted. Together they isolate the residual difficulty of tablature transcription: among all physically correct positions for a given pitch, which one does the ground truth actually use.","core_discovery":"The authors claim that the decoder token representation itself is a primary cause of transcription quality, not just a vocabulary change. Their decoder outputs NOTE_ON_PITCH, TAB<string,fret>, NOTE_OFF_PITCH, and TIME_SHIFT_TICKS tokens per note event, so pitch and note boundaries are explicit during autoregressive generation. This separation lets the model learn note-event structure and string–fret assignment more easily, yielding a 10.56 percentage point token-accuracy gain and 5.32 point tablature-accuracy gain over the retrained Fretting Transformer on DadaGP. They further show that applying a pitch-validity constraint during decoding — setting logits of candidate string–fret pairs outsi","pith_inferences":["The pitch-validity constraint is domain-exact and cheap; it could be retrofitted to any existing MIDI-to-tab system at inference time, turning pitch consistency into a guarantee rather than a learned behavior.","The small-data result suggests that tokenization design may matter more than architecture or regularization when training data are scarce; a clean next experiment would hold the model architecture fixed and vary only the target tokenization.","A testable extension would condition the decoder on fretboard-position style or add a playability reward, since the residual 14–16% position error under perfect pitch constraints is the next bottleneck."],"forward_implications":["If the central claim is right, target-side tokenization is a first-order factor in tablature transcription: simply adding note-event tokens yields double-digit token-accuracy gains on DadaGP.","The constrained-decoding results show that a large share of the baseline's errors are pitch-invalid, and that pitch-validity is a separable, controllable axis of error.","The small-data result indicates that explicit note-event structure acts as a strong inductive bias, keeping the model trainable where the baseline overfits.","Under combined training, the proposed method remains competitive on Leduc while still improving DadaGP, suggesting the tokenization transfers across datasets of different scales.","The error decomposition implies that further progress depends on modeling guitarist string–fret preferences, since pitch-valid ambiguity persists after all pitch-invalid choices are removed."],"fun_headline_variants":["Explicit note-event tokens lift MIDI-to-tab accuracy","Pitch-validity masking improves tab accuracy, especially small data","NOTE_ON/OFF tokens beat transformer for guitar tab transcription","Small-data tab gains from explicit note-event decoding"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The comparison assumes that the accuracy metrics measure the same thing across models with different target tokenizations; specifically, the way predicted TAB tokens are aligned to ground-truth TAB tokens when one sequence contains NOTE_ON/OFF tokens and the other does not is not specified, so the reported gains could partly reflect alignment artifacts rather than true generation quality.","fun_headline_variants_meta":{"raw":{"variants":["Explicit note-event tokens lift MIDI-to-tab accuracy","Pitch-validity masking improves tab accuracy, especially small data","NOTE_ON/OFF tokens beat transformer for guitar tab transcription","Small-data tab gains from explicit note-event decoding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000238,"raw_usage":{"total_tokens":1364,"prompt_tokens":776,"completion_tokens":588,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":520,"completion_tokens_details":{"reasoning_tokens":520}},"tokens_in":520,"tokens_out":588,"duration_ms":6765,"temperature":1.0,"reasoning_tokens":520,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T15:44:10.625033+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the Fretting Transformer's and the proposed model's outputs on the Leduc test set, align both to ground-truth note events by matching NOTE_ON boundaries (or by a note-level alignment that ignores tokenization), and recompute pitch and tablature accuracy. If the baseline's 0.49% accuracy jumps to near the proposed model's 98% once aligned note-event-aware, the reported advantage is an artifact of tokenization mismatch rather than better generation.","supporting_citations":[],"review_version":1}