{"id":"6e49590b-9be2-477d-b1e9-7a5aafd4a560","arxiv_id":"2501.05068","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"D3RM uses discrete diffusion with neighborhood attention and an asymmetric train/inference masking schedule to improve piano transcription F1 over a feed-forward baseline and DiffRoll on MAESTRO.","lead":"D3RM adds a discrete denoising diffusion module to a pretrained piano transcription model, refining piano rolls with neighborhood attention. On the MAESTRO dataset it reports higher F1 than its feed-forward baseline and previous diffusion-based transcription, but without significance tests and with only modest gains.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"AS-sampling gain rests on an explicit but unproven manifold equivalence in §III-D; without invariance of pθ to substitution-vs-mask corruption, the +0.34 F1 is an unexplained heuristic.","rationale":"I agree with the reader's weakest-assumption identification. The empirical tables are internally consistent and D3RM beats the baseline across nearly every reported gamma_T row, so I do not see grounds to reject the headline comparison. The sharpest unresolved point is the AS-sampling justification: the paper states the manifold equivalence as an assumption, does not prove it, and the corruption-count matching is loose (the state-change probability differs by beta under Eq. (1)). The concrete experiment above would test whether the decoder is actually invariant to the corruption type. The reader's CONDITIONAL verdict remains appropriate; I would not move it. In addition, the paper should report multiple seeds or uncertainty for the small AS ablation, but that is secondary to the missing manifold validation.","tokens_in":7860,"tokens_out":19930,"duration_ms":210268,"concrete_test":"Using the released code, corrupt a fixed batch of MAESTRO test piano rolls at several timesteps tau in two ways while keeping the paper's 'transitioned-token count' matched: (a) with the training Q_tau (substitution + mask) and (b) with the AS mask-only posterior with mask probability gamma_tau + K beta_tau. Feed both to the trained decoder and compare p_theta(y0_hat | y_tau, x) on the corrupted positions (KL divergence, or downstream F1 after a full reverse pass). If the outputs are statistically equivalent, the manifold assumption gains empirical support; if they differ, AS sampling is operating off the training manifold. A second check: retrain the decoder with the same gamma_T=0.9 schedule but beta=0 (mask-only forward) and test AS sampling; if the AS gain disappears, the benefit is caused by the train/inference mismatch rather than by manifold-consistent refinement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing weakness is Section III-D's AS-sampling step. The paper explicitly 'simply assume[s] that the model identifies diffusion manifolds as samples with the same total portion of corrupted tokens' and then replaces the learned posterior with a masking-only posterior q~ using mask probability gamma_tau + K beta_tau. This equivalence is never derived or tested. Under Eq. (1), the probability that a token's state actually changes is gamma + (K-1)beta (the diagonal is alpha+beta, so beta of the 'replacement' mass is a self-loop), whereas AS corrupts with gamma + K beta = 1 - alpha. So the two processes do not even agree on the expected number of changed tokens under the natural reading; at best they agree on the number of tokens that have undergone a transition operation. No invariance of p_theta(y0_hat | y_tau, x) to substitution-vs-mask corruption is shown. If the model is not invariant, AS sampling does not sample from the trained reverse process, and the reported +0.34 F1 (97.21 -> 97.55) is an unexplained heuristic rather than a validated refinement mechanism. This matters because the asymmetric train/inference transition is the paper's main methodological novelty, even though the headline baseline comparison may survive.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes D3RM, a piano transcription system that couples a pretrained HPPNet-style acoustic encoder with a discrete denoising diffusion decoder built from Neighborhood Attention layers. The target label is a multi-state (5 note states plus mask) piano-roll representation, and the model is trained with a standard discrete diffusion loss plus an auxiliary denoising objective. The main methodological novelty is 'absorbing state sampling' (AS sampling): during inference, the transition posterior is replaced by a masking-only posterior with total corruption probability γτ + Kβτ, based on an explicit assumption in Section III-D that the model identifies diffusion manifolds as samples with the same total portion of corrupted tokens. Experiments on MAESTRO V3 report note F1 of 97.57 (with offsets 90.44), outperforming the baseline HPPNet-based model (96.45) and DiffRoll (78.1). Ablations show AS sampling improves F1 from 97.21 to 97.55, and that training with partial substitution (γT = 0.9) outperforms training with a pure absorbing state (γT = 1.0).","tokens_in":8092,"tokens_out":2884,"duration_ms":29475,"significance":"If the results are reliable, the paper makes a useful empirical contribution: it shows that a lightweight discrete diffusion refinement decoder can improve over a strong feedforward acoustic model for piano transcription, and it provides a concrete architecture (NA-based denoising, cross-attention conditioning, multi-state labels) with a modest parameter count. The code is released, which supports reproducibility. However, the significance is tempered by two factors. First, the headline gain of the proposed AS-sampling mechanism rests on an unproven manifold-invariance assumption, so the paper's central methodological novelty is not yet validated as a principled inference procedure. Second, all reported F1 differences are small and come from single runs, with no uncertainty estimates; the AS-sampling gain of +0.34 F1 and the γT selection (0.9 vs 0.4 or 0.6) are within a range that could plausibly be noise. The baseline comparison is internally consistent and the empirical framework is standard, but the lack of statistical grounding weakens the confidence one can place in the specific claims about the proposed components.","major_comments":[{"comment":"The manifold equivalence assumption is load-bearing for the AS-sampling contribution and is neither derived nor tested. The paper states, 'we first simply assume that the model identifies diffusion manifolds as samples with the same total portion of corrupted tokens,' and then replaces q(y_{τ−1}|yτ,y0) with a masking-only posterior using probability γτ + Kβτ. As the skeptical note observes, under Eq. (1) the probability that a non-mask token actually changes state is γ + (K−1)β (the diagonal is α+β), not γ + Kβ; the two processes agree only on the number of tokens that undergo a transition operation, not on the expected number of changed tokens. No invariance of pθ(y0|yτ,x) to substitution-versus-mask corruption is shown. This matters because the claimed +0.34 F1 improvement (97.21 to 97.55) is attributed to this substitution. I request either a formal argument or a concrete empirical test: for example, feed the same partially corrupted sample through the model with the two corruption schemes at matched corruption levels and compare the predicted y0 distributions, or validate AS sampling against an alternative inference that uses the true learned posterior. Without this, AS sampling is an unexplained heuristic that happens to help on one benchmark.","section":"III-D"},{"comment":"All main results are reported from what appear to be single training runs, with no variance, confidence intervals, or statistical significance tests. The differences among the ablation conditions are small: AS sampling changes F1 by 0.34 (97.21 vs 97.55), encoder initialization by 0.09 (97.55 vs 97.46), and γT variations across 0.4/0.6/0.8/0.9 span only 0.07 (97.51–97.58). Without repeated runs or a significance test, it is impossible to know whether the AS-sampling improvement or the γT ranking is a reliable effect or run-to-run noise. The paper should provide mean±std over at least three seeds for the main configuration and for the key ablations, or otherwise justify that the differences exceed the noise floor.","section":"V-A / Table I"},{"comment":"The choice of γT = 0.9 as the main configuration is not justified by any stated validation procedure. The ablation table shows γT = 0.4 gives note F1 97.58 and F1-with-offsets 90.37, while γT = 0.9 gives 97.55 and 90.28; the differences are tiny and the best numbers actually favor γT = 0.4. Since the paper linearly increases γτ to a selected γT but does not describe how the grid {0.4,0.6,0.8,0.9,1.0} was chosen or post-selected, the reader cannot tell whether the headline configuration was chosen based on held-out data or on the test set. Please specify the validation protocol used to select γT and report test-set metrics for the selected configuration.","section":"IV-B and V-B / Table II"}],"minor_comments":[{"comment":"The phrase 'Building on these foundings' should be 'Building on these findings'.","section":"I"},{"comment":"'hierachically' is a typo for 'hierarchically'.","section":"II-D"},{"comment":"The column header 'END. COND.' should be 'ENC. COND.' for consistency with the caption's 'encoder conditioning method'.","section":"Table II"},{"comment":"The caption begins with stray text, '4MIDI-Unprocessed 12 R1 2006 01-08 ORIG MID–', which appears to be a leftover from a dataset path; it should be removed.","section":"Fig. 3 caption"},{"comment":"The notation is confusing because the single-step transition matrix is written as Qτ and the cumulative product is also written as Qτ (for example, 'The cumulative products Qτ = Qτ . . . Q1'). The two should be distinguished, e.g., using Q‾τ or Q(τ) for the cumulative product.","section":"III-C"},{"comment":"The first term DKL[q(yT|y0)||p(yT)] is written with a subscript 'T' but the equation is then summed over τ = 1 to T; please check that the prior term is intended to be the Kullback-Leibler divergence for the final timestep and clarify the indexing.","section":"III-C, Eq. (4)"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the empirical comparison is internally consistent, but the main methodological novelty (AS sampling) is justified by an explicit, unproven manifold-invariance assumption. The authors themselves state the assumption, so the gap is acknowledged rather than hidden. I do not see evidence of circularity or target-fitting: the F1 numbers come from measured comparisons. The main risk is that the headline AS-sampling improvement could be a fluke of a single run; the requested seed-variance analysis is essential. If the authors cannot provide a principle or test for the manifold assumption, they should reframe AS sampling as a heuristic and de-emphasize the claim that it is a 'novel strategy' with a theoretical basis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is the first paper to apply discrete multinomial diffusion to piano transcription, and the empirical core holds up: D3RM beats its own feed-forward baseline clearly (97.57 vs 96.45 note F1) and crushes DiffRoll (78.1). The model is small (2M params), the code is public, and the ablations are honest. That is real, if incremental, progress for the transcription subfield.\n\nThe genuinely new piece is the 'absorbing state sampling' (AS) trick: train with a mixture of substitution and masking transitions, then at inference replace the posterior with a masking-only transition using total corruption gamma_T + K beta_T. The paper justifies this with an explicit 'simply assume' that the model sees two corruption processes as the same diffusion manifold. That is the soft spot, and the stress-test is right that it is not just unproven but likely false under a straightforward reading. In Eq. (1), the probability a token actually changes state is gamma + (K-1)beta, not gamma + K beta; the latter includes the beta self-loop mass. So the two processes do not even match on the expected number of corrupted tokens. Without an invariance argument for p_theta, AS sampling is a heuristic that happens to give +0.34 F1 on one dataset. The paper says 'simply assume' rather than 'we prove', so it is honest, but the claim that it 'places the sample on the manifold' is unsupported and probably wrong.\n\nThat said, the main result is not load-bearing on AS. Even without it, D3RM gets 97.21, still above the baseline and above DiffRoll. So the paper's central claim survives. The other weaknesses are minor: no error bars, and the gamma_T sweep (0.4-1.0) is reported on the test set, so there is a mild selection-on-test issue. The ablation differences among 0.4/0.6/0.9 are within noise, so claiming any one is optimal is overreach.\n\nThis is a solid engineering paper with a weak theoretical justification for one trick. It deserves a serious referee, and a good referee will ask the authors to either prove the manifold equivalence under a defined metric (e.g., expected fraction of changed tokens) or reframe AS as a heuristic with a matching experimental defense (e.g., multiple seeds, error bars). I would send it out, not desk-reject.","headline":"First discrete diffusion for piano transcription with a solid baseline gain, but the AS-sampling trick's manifold equivalence is unproven and likely wrong; the main result survives anyway.","tokens_in":8627,"tokens_out":4764,"would_cite":false,"duration_ms":40811,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A discrete diffusion refinement decoder raises piano transcription F1 to 97.57, beating feed-forward baselines and prior diffusion models.","keywords":["automatic music transcription","piano transcription","discrete diffusion model","denoising refinement","Neighborhood Attention","absorbing state sampling","MAESTRO dataset","music information retrieval"],"falsifier":"Run D3RM on a second piano dataset (e.g., MAPS) and compare inference with AS sampling against ordinary posterior sampling; if ordinary sampling matches or beats AS sampling, or if the empirical fraction of masked tokens at each timestep under AS sampling deviates from the training-time distribution, then the manifold-equivalence assumption is not holding.","tokens_in":7625,"feed_emoji":"🎹","tokens_out":6440,"duration_ms":53493,"temperature":0.7,"pith_summary":"The paper tries to show that discrete denoising diffusion can act as a refinement module for piano transcription, not just a generator. The authors attach a lightweight Neighborhood Attention denoising decoder to a pretrained acoustic model, and let it iteratively correct a multi-state piano roll. On MAESTRO, the resulting D3RM model reaches 97.57 note F1 (90.44 with offsets), beating its feed-forward baseline (96.45/87.60) and the prior continuous diffusion model DiffRoll (78.1). The paper also introduces 'absorbing state sampling,' an inference-only change to the diffusion posterior that improves refinement further (97.21 to 97.55). If correct, the result shows that diffusion-based iterative refinement is a practical way to make note predictions depend on each other.","feed_headline":"Discrete diffusion refines piano transcription to 97.6% F1","feed_subtitle":"A lightweight denoising decoder with neighborhood attention beats feed-forward baselines on MAESTRO.","key_machinery":"The load-bearing mechanism is the discrete diffusion transition matrix and the asymmetric posterior used at sampling. The forward transition $Q_\\tau$ mixes a uniform replacement probability $\\beta_\\tau$ with a masking probability $\\gamma_\\tau$ and an absorbing [MASK] state; the training loss is the variational lower bound with the auxiliary denoising objective from VQ-diffusion. During refinement, the model first predicts the clean sample $\\hat{y}_0$, then applies the posterior transition. For 'absorbing state sampling,' the posterior $q(y_{\\tau-1}|y_\\tau,\\hat{y}_0)$ is replaced by a masking-only transition $\\tilde{q}$ with total corrupted-token probability $\\gamma_\\tau + K\\beta_\\tau$, which the authors argue places the sample on the same diffusion manifold as training while avoiding token-replacement disruption. Neighborhood Attention layers with window size 3 and dilation pattern [1,2,4,8,1,2,4,8] provide the locality bias that lets each piano-roll token attend to neighboring tokens, enabling iterative refinement.","core_discovery":"The central claim is that a discrete diffusion model, trained to denoise a multi-label piano roll with a transition matrix that both replaces and masks tokens, can serve as an effective refinement decoder for automatic piano transcription. The target is a $T \\times 88 \\times 6$ tensor with five note states plus a mask, and the decoder uses a pitchwise bidirectional LSTM plus stacked Neighborhood Attention (NA) layers with AdaLN conditioning, cross-attending to penultimate features of the pretrained HPPNet-style acoustic encoder. At inference, the authors replace the posterior $q(y_{\\tau-1}|y_\\tau,y_0)$ with a mask-only posterior $\\tilde{q}$ in which the total corruption probability is $\\gamma_\\tau + K\\beta_\\tau$, based on the assumption that diffusion manifolds are identified by the total fraction of corrupted tokens. With this asymmetric train/inference schedule, D3RM obtains 97.57/90.44 note F1 (with/without offsets) on MAESTRO, versus 96.45/87.60 for the baseline and 78.1 for DiffRoll.","pith_inferences":["The 'manifold equals total corrupted-token fraction' assumption, if it holds, suggests a general recipe: any discrete diffusion discriminative decoder can be trained with replace+mask and sampled with mask-only transitions; the trick may transfer to other dense prediction tasks such as sound event detection or instrument separation.","A principled derivation of the optimal inference-time transition, rather than the simple mask-only posterior, could yield further gains; one testable variant is to choose the posterior that minimizes the KL to the training-time marginal at each timestep.","Because $\\gamma_T$ values of 0.4–0.9 all performed within 0.07 F1, the refinement benefit appears robust to the exact corruption schedule; a follow-up could test whether the same holds on datasets with more polyphony or pedal complexity (e.g., MAPS).","The authors' own future direction of note-level token representations might let the same refinement framework handle note onsets/offsets as structured events rather than frame-wise states, potentially improving offset F1 more than the 2.84-point current gain."],"forward_implications":["Piano transcription can be improved by a lightweight diffusion refinement decoder rather than a larger feed-forward model: D3RM uses 2.0M parameters and beats models with up to 86.8M parameters (DiffRoll) and 19.7M (PAR).","Learning state-to-state transitions during training is what gives refinement power: setting $\\gamma_T=1.0$ (absorbing-only training) drops F1 from 97.55 to 96.82.","Applying the mask-only 'absorbing state sampling' at inference consistently helps (F1 97.21 → 97.55; with offsets 89.10 → 90.28), so the refinement benefit is not just from the architecture.","Computing the acoustic encoder features once and reusing them across diffusion timesteps makes the refinement loop fast, needing only one forward pass of the encoder."],"supporting_citations":[{"why":"Supplies the multi-state note label representation (onset, sustain, re-onset, offset, off) that D3RM uses as its discrete target.","marker":"[14]"},{"why":"Provides the reparameterized training objective with auxiliary denoising loss and the training schedule D3RM follows.","marker":"[20]"},{"why":"Defines the absorbing-state diffusion posterior that D3RM's 'AS sampling' adopts for inference.","marker":"[19]"},{"why":"Introduces Neighborhood Attention, the sliding-window attention mechanism used in the denoising decoder.","marker":"[7]"},{"why":"DiffRoll, the only prior diffusion-based transcription model, serves as the main comparison baseline.","marker":"[6]"},{"why":"HPPNet, the feed-forward acoustic model whose architecture and penultimate features are adapted as the conditioning encoder.","marker":"[10]"},{"why":"MAESTRO V3 dataset used for training and evaluation.","marker":"[22]"}],"fun_headline_variants":["D3RM: Discrete diffusion sharpens piano roll transcription to 97.6 F1","Neighborhood attention denoising lifts piano transcription F1 to 97.6","Discrete diffusion refinement beats baselines on piano transcription","Piano transcription hits 97.6 F1 with discrete diffusion refinement","D3RM: Discrete denoising diffusion refines piano transcription"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's improvement hinges on the assumption that diffusion manifolds are exactly the sets of samples with the same total number of corrupted tokens, which lets the authors swap in a mask-only posterior during inference; if wrong, the sampling is heuristic and the F1 gains may not generalize.","fun_headline_variants_meta":{"raw":{"variants":["D3RM: Discrete diffusion sharpens piano roll transcription to 97.6 F1","Neighborhood attention denoising lifts piano transcription F1 to 97.6","Discrete diffusion refinement beats baselines on piano transcription","Piano transcription hits 97.6 F1 with discrete diffusion refinement","D3RM: Discrete denoising diffusion refines piano transcription"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000712,"raw_usage":{"total_tokens":3197,"prompt_tokens":935,"completion_tokens":2262,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":551,"completion_tokens_details":{"reasoning_tokens":2165}},"tokens_in":551,"tokens_out":2262,"duration_ms":14564,"temperature":1.0,"reasoning_tokens":2165,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:19:43.119173+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run D3RM on a second piano dataset (e.g., MAPS) and compare inference with AS sampling against ordinary posterior sampling; if ordinary sampling matches or beats AS sampling, or if the empirical fraction of masked tokens at each timestep under AS sampling deviates from the training-time distribution, then the manifold-equivalence assumption is not holding.","supporting_citations":[{"cited_title":"Polyphonic piano transcription using autoregressive multi-state note model,","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-state note label representation (onset, sustain, re-onset, offset, off) that D3RM uses as its discrete target."},{"cited_title":"Neighborhood attention transformer,","cited_arxiv_id":null,"evidence_quote":"Introduces Neighborhood Attention, the sliding-window attention mechanism used in the denoising decoder."},{"cited_title":"Diffroll: Diffusion-based gen- erative music transcription with unsupervised pretraining capability,","cited_arxiv_id":null,"evidence_quote":"DiffRoll, the only prior diffusion-based transcription model, serves as the main comparison baseline."},{"cited_title":"Hppnet: Modeling the harmonic struc- ture and pitch invariance in piano transcription,","cited_arxiv_id":null,"evidence_quote":"HPPNet, the feed-forward acoustic model whose architecture and penultimate features are adapted as the conditioning encoder."},{"cited_title":"Enabling factorized piano music modeling and generation with the MAESTRO dataset,","cited_arxiv_id":null,"evidence_quote":"MAESTRO V3 dataset used for training and evaluation."}],"review_version":1}