{"id":"366ff0e4-4112-4339-bebb-4ada8f2e5758","arxiv_id":"2502.05757","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"LingoNMF uses LLM feedback to set a frequency penalty inside nonnegative matrix factorization, improving most but not all tested heart and lung separation conditions.","lead":"This paper uses a large language model to guide a standard sound-separation math method as it splits stethoscope recordings into heart and lung parts. The authors say the method beats older approaches, but their own results show it fails on heart sounds in the manikin dataset, and the tuning procedure is not independent.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (7)'s penalty term is not minimized by updates (3)-(4), so LingoNMF's LLM feedback cannot be responsible for the claimed gains.","rationale":"The central claim is that LingoNMF, which integrates an LLM feedback loop into PL-NMF, consistently outperforms existing NMF methods for heart/lung separation. For that claim to hold, the LLM-driven penalty in Eq. (7) must influence the factorization. The paper, however, supplies no update rule for the penalty term: Eqs. (3) and (4) are the standard alpha-NMF multiplicative updates for D(Y||AX). Because f_est is defined as the argmax of the PSD (Eq. 8), its gradient with respect to X is zero almost everywhere, so the penalty term cannot contribute to gradient-based update rules. This means the algorithm as documented does not optimize its stated objective, and the reported SDR gains cannot be attributed to the proposed mechanism. An ablation with the LLM disabled would settle this. The reader's REJECT verdict is appropriate; my read does not change it. I agree with the reader's weakest assumption.","tokens_in":5838,"tokens_out":9935,"duration_ms":105354,"concrete_test":"Run the released code with the LLM feedback disabled: fix f=50 Hz and skip all LLM calls while keeping lambda_f, alpha, number of layers, and initialization identical to the reported setup. Compare the resulting heart/lung SDR values to Table 2; if they are unchanged within rounding, the penalty term has no effect and the proposed LLM mechanism is not responsible for the reported improvements.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.3 defines the LingoNMF objective in Eq. (7), D' = D(Y||AX) + lambda_f ||f_est - f||^2, where f_est is the vector of estimated fundamental frequencies computed as the argmax of the PSD of the current row estimates (Eq. 8). The implementation, however, uses the standard alpha-NMF multiplicative updates (3) and (4), which are derived for D(Y||AX) alone; no gradient, auxiliary function, or update rule for the penalty term is provided. This is not a minor omission: because f_est is an argmax, its derivative with respect to X is zero almost everywhere, so the penalty term cannot steer the X or A updates. Thus, as written, LingoNMF does not minimize the stated cost function. Any separation gains shown in Table 2 cannot be attributed to the LLM-driven fundamental-frequency penalty; they may stem from the hyperparameters selected on the test data in Section 3.2 or from PL-NMF's periodicity selection. Without a corrected update or a clear mechanism by which the penalty affects the factorization, the central claim that the LLM feedback improves cardiorespiratory sound separation is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LingoNMF, an extension of the periodic-layer NMF (PL-NMF) algorithm for separating heart and lung sounds, in which a LLaMA-2 language model analyzes features of the currently separated signals and suggests updated fundamental-frequency targets that enter a quadratic penalty term added to the α-NMF cost. The method is evaluated on two datasets (100 synthesized mixtures of real heart/lung sounds and 210 manikin recordings) and compared against standard NMF, α-NMF, and PL-NMF using SDR, SIR, and SAR metrics. The abstract claims that the approach 'consistently outperformed existing methods.' My reading of the manuscript is that this central claim is not supported: the modified cost function in Eq. (7) is not minimized by the stated multiplicative updates, the main hyperparameters were selected on the evaluation datasets, and Table 2 itself contains a clear counterexample to 'consistently outperforms.'","tokens_in":6157,"tokens_out":3431,"duration_ms":35177,"significance":"If the technical issues were resolved, the idea of using an LLM to adjust a penalty term in NMF for cardiorespiratory sound separation would be a novel and potentially interesting contribution, particularly because the authors release their code and use publicly available data. The paper also usefully extends PL-NMF with an adaptive fundamental-frequency feedback loop. However, as presented, the contribution is not established: the missing update derivation means the LLM feedback cannot be verified to affect the factorization, the hyperparameter tuning on test data undermines the comparison, and the reported results contradict the headline claim. These are load-bearing problems, not presentation issues, and they warrant rejection of the manuscript in its current form.","major_comments":[{"comment":"The LingoNMF objective in Eq. (7) includes the penalty term λ_f ||f_est − f||^2, where f_est is defined in Eq. (8) as the argument that maximizes the PSD of the current row estimates. However, the algorithm described in Section 2.3 and Algorithm 1 uses the standard α-NMF multiplicative updates (3) and (4), which are derived for D(Y||AX) alone. The paper gives no gradient, auxiliary function, or update rule for the penalty term. Because an argmax is piecewise constant in the factorization variables, the derivative of f_est with respect to the NMF factors is zero almost everywhere, so the penalty term cannot steer the updates as written. Consequently, the algorithm does not minimize the stated cost function, and the separation gains reported in Table 2 cannot be attributed to the LLM-driven fundamental-frequency penalty.","section":"2.3, Eq. (7) and Eq. (8)"},{"comment":"The hyperparameters α, the number of layers L, and λ_f were selected by scanning values and choosing the configuration that yielded the highest SNR on the evaluation datasets (Section 3.2; Table 1; Fig. 3). This is a form of test-set selection: the reported advantage of LingoNMF over the baselines may reflect that these parameters were tuned to the evaluation data, whereas no equivalent tuning is reported for standard NMF, α-NMF, or PL-NMF. The comparison is therefore not a fair test of the proposed LLM feedback mechanism.","section":"3.2, Table 1, Fig. 3"},{"comment":"The abstract's claim that LingoNMF 'consistently outperformed existing methods' is contradicted by Table 2. On Dataset Two heart sounds, LingoNMF achieves SDR 7.9 dB, while PL-NMF achieves 9.4 dB and α-NMF achieves 13.7 dB. This is a direct reversal of the claimed advantage, and it means the paper's own evaluation does not support the unqualified claim of consistent superiority. The claim should be restricted to the metrics and datasets where LingoNMF is actually better, or the experimental design and conclusions should be revised.","section":"Table 2"},{"comment":"No statistical uncertainty is reported for any of the SDR, SIR, or SAR values in Table 2. There are no standard deviations, confidence intervals, or numbers of independent runs. Without such information, it is impossible to judge whether small differences (e.g., lung SDR 8.9 vs. 8.4 dB in Dataset One) are significant, and the reverse result for heart sounds in Dataset Two cannot be assessed as a real failure versus noise. This is especially important because the paper's main claim depends on the quantitative comparisons in this table.","section":"3.3, Table 2"}],"minor_comments":[{"comment":"The α-divergence formula is garbled in the manuscript, with missing parentheses and unclear subscript/superscript ranges; it should be typeset carefully so the cost function is unambiguous.","section":"Eq. (2)"},{"comment":"The notation Y ← λ'.Y + λ. is not explained: it appears to be an affine normalization but the parameters λ' and λ. are not defined in the text, and the superscripts in the update equations are difficult to parse.","section":"Algorithm 1"},{"comment":"The sentence 'we applied the BSS_EVAL toolbox to measure performance' should specify the version and the exact global or local criteria used, since BSS_EVAL can report several variants of SDR/SIR/SAR.","section":"Section 2.4"},{"comment":"The claim that the LLM provides 'detailed insights for disease prediction' is presented as a contribution but is not evaluated quantitatively anywhere in the paper; if it remains a claimed benefit, it needs a separate assessment beyond a single illustrative example.","section":"Section 2.3 / Fig. 5"},{"comment":"References [3] and [4] (fever detection with infrared thermography and bipedal walking) appear unrelated to the topic of cardiorespiratory sound separation and should be removed or replaced with relevant prior work.","section":"References"}],"recommendation":"reject","confidential_remarks":"The manuscript has a publicly available code repository and uses a publicly available dataset, which are good practices. However, the technical flaw in the derivation of the update for Eq. (7) is fundamental: without a correct update or a clear mechanism by which the LLM's suggested frequencies affect the factorization, the proposed method is not actually optimizing the stated objective. The test-set hyperparameter selection and the contradictory result in Table 2 further undermine the central claim. I see no feasible patch within a normal minor revision; the authors would need to re-derive the optimization, re-implement the algorithm, and re-run the comparisons with an honest evaluation protocol. I therefore recommend rejection, while noting that a substantially revised version addressing these points could be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. The combination is genuinely new: an LLM in a feedback loop that adjusts fundamental-frequency targets inside a periodicity-regularized NMF for heart/lung separation. The authors also ship code and use a public manikin dataset. But the central claim—'consistently outperformed'—is undercut by their own Table 2, and the mechanism they describe for the LLM penalty does not actually enter the optimization as written.\n\nThe prior work is honestly cited; the novelty is real but narrow. The public code and data are a plus, and some numbers, like Dataset One heart SDR 35.5 vs 17.7 for PL-NMF, are striking.\n\nNow the soft spots, in order. First, and decisive: Eq. (7) defines D' = D(Y||AX) + λ_f ||f_est − f||^2, with f_est the argmax over the PSD of the current estimates. The updates used, (3) and (4), come from the α-divergence NMF for D alone. No update rule, gradient, or auxiliary function is given for the penalty. Because f_est is an argmax, its derivative w.r.t. X is zero almost everywhere, so the penalty term cannot affect those updates. So as written, LingoNMF is not minimizing D'. That means the reported gains cannot be attributed to the LLM penalty unless a corrected update is shown.\n\nSecond, the hyperparameter selection. Section 3.2 says α, number of layers, and λ_f were chosen by scanning for the highest SNR on the evaluation datasets. Comparing against baselines with that protocol gives an unfair advantage, and it breaks the claim that gains come from the LLM.\n\nThird, Table 2 itself. On Dataset Two heart sounds, LingoNMF SDR is 7.9 dB, versus 9.4 for PL-NMF and 13.7 for α-NMF. That is not consistent superiority. The abstract's 'consistently outperformed' needs to be repaired.\n\nNone of this makes the idea worthless. A referee could run the public code and check whether the penalty affects the factorization at all. If the mechanism is missing, the paper needs substantial revision, not just a wording fix. The right next step is a proper derivation or an ablation that isolates the LLM contribution, plus an evaluation protocol that does not tune on test data.\n\nBottom line: this is a paper for readers interested in LLM-driven adaptations of classic signal-processing methods. It deserves a serious referee because the code is checkable and the combination is new, but I would not cite it as a supporting result until the mechanism and evaluation are cleaned up.\n\nRecommendation: send to peer review, with a high bar: the authors must show a correct update or explicit ablation, and re-run the comparison without test-set tuning.","headline":"A new LLM-NMF combination that ships code and data, but the core penalty mechanism is not actually optimized and the evaluation is tuned on test data.","tokens_in":6630,"tokens_out":3873,"would_cite":false,"duration_ms":40308,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper presents LingoNMF, the first NMF method augmented by LLM feedback, and reports better heart and lung separation than standard NMF, alpha-NMF, and PL-NMF on two sound datasets.","keywords":["large language models","nonnegative matrix factorization","blind source separation","heart sound","lung sound","fundamental frequency penalty","alpha-divergence","digital stethoscope"],"falsifier":"Compute the gradient of the penalty term $\\lambda_f \\|\\mathbf{f}_{\\mathrm{est}}(\\mathbf{A},\\mathbf{X}) - \\mathbf{f}\\|^2$ with respect to $\\mathbf{A}$ and $\\mathbf{X}$ and check whether the unmodified multiplicative updates (3)-(4) form a descent step for it; if the penalty can increase, LingoNMF is not minimizing the stated objective. A paired experiment replacing the LLM's frequency choices with a grid search or random schedule on the same datasets would then settle whether the LLM feedback, rather than the penalty itself, produces the reported SDR gains.","tokens_in":5660,"feed_emoji":"🫀","tokens_out":7033,"duration_ms":71718,"temperature":0.7,"pith_summary":"This paper introduces LingoNMF, described as the first integration of a large language model with nonnegative matrix factorization (NMF) for separating heart and lung sounds. The central claim is that a LLaMA 2 model in a feedback loop improves separation by dynamically updating a fundamental-frequency penalty in the NMF cost function. The evidence comes from two datasets: 100 synthesized mixtures of real heart and lung sounds, and 210 clinical-manikin recordings of normal and abnormal sounds. A sympathetic reading of the paper takes the claimed separation gains as the contribution, with the LLM acting both as a tuner of the penalty term and as a provider of diagnostic text about the separated sounds.","feed_headline":"LLM feedback loop sharpens heart-lung sound separation","feed_subtitle":"First NMF algorithm that lets a language model tune a frequency penalty, beating prior NMF methods on two medical-sound datasets.","key_machinery":"The central object is the augmented cost function $D' = D(\\mathbf{Y}\\|\\mathbf{A}\\mathbf{X}) + \\lambda_f \\|\\mathbf{f}_{\\mathrm{est}} - \\mathbf{f}\\|^2$, where $\\mathbf{f}_{\\mathrm{est}}$ is the vector of fundamental frequencies estimated as the power-spectral-density argmax for each current separated estimate. The stated mechanism is the feedback loop: the LLM receives spectro-temporal features of the current estimates, adjusts $\\mathbf{f}$ toward values consistent with normal or abnormal cardiorespiratory sounds, and the factorization is re-run. This sits on top of PL-NMF, a multilayer, periodicity-aware version of $\\alpha$-NMF that already uses autocorrelation to separate heart and lung components by their different period ranges.","core_discovery":"The authors propose that a large language model can participate in NMF iteration as a feedback controller rather than only as a post-processing annotator. In LingoNMF, after each factorization update, features of the current separated estimates are sent to LLaMA 2, which returns updated values for the vector of source fundamental frequencies; these values enter the cost function as a penalty on the deviation between estimated and target fundamental frequencies. The paper reports that this loop raises source-to-distortion ratio on both datasets for lung sounds and, on the 100-mixture dataset, for heart sounds, with the largest gains appearing in the LingoNMF columns of its comparison table.","pith_inferences":["Not stated by the paper: on the manikin dataset the heart-sound SDR for LingoNMF is lower than for $\\alpha$-NMF, so the consistency claim is stronger for lung sounds and for the synthesized dataset; a per-condition comparison with error bars would clarify where the method actually helps.","An ablation that replaces the LLM with a simple numerical rule or a grid search for the fundamental-frequency vector would reveal whether the language model itself, rather than the penalty mechanism, drives the reported improvement.","Because $f_0$ is estimated as the PSD argmax, abnormalities that alter harmonic structure can shift the estimate; the penalty update may need a robustness check against such shifts before clinical use."],"forward_implications":["If the reported gains hold, LLM feedback can replace manual or grid-search tuning of NMF penalty parameters during the factorization itself.","The same loop could be applied to other periodic biomedical recordings, such as bowel or fetal sounds, wherever a fundamental-frequency prior is useful.","The reported optimal range of five to seven input mixtures would guide the design of stethoscope arrays for practical monitoring.","The algorithm separates sounds using only the mixture, so it can be deployed where pure reference sounds are unavailable.","The LLM's text output also yields a human-readable abnormality analysis, coupling separation with diagnostic suggestions in one pipeline."],"supporting_citations":[{"why":"Supplies the $\\alpha$-divergence cost and the multiplicative update rules (3)-(4) on which PL-NMF and LingoNMF are built.","marker":"[20]"},{"why":"Defines PL-NMF, the multilayer periodic $\\alpha$-NMF that LingoNMF extends with the LLM feedback loop.","marker":"[22]"},{"why":"Provides the LLaMA 2 model used in the feedback loop to interpret features and update the fundamental-frequency penalty parameters.","marker":"[23]"},{"why":"Defines the BSS_EVAL metrics (SDR, SIR, SAR) used to compare all methods.","marker":"[25]"},{"why":"Supplies the public clinical manikin dataset with normal and abnormal heart and lung sounds used in Dataset Two.","marker":"[26]"}],"fun_headline_variants":["LLM feedback loop tunes NMF to sharpen heart-lung sounds","First LLM-guided NMF separates cardiac and respiratory sounds","Language model optimizes NMF for better heart-lung separation","LLM-driven NMF improves medical sound separation over prior methods","NMF with LLM feedback loop sharpens stethoscope sound separation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that the standard alpha-NMF multiplicative update rules minimize the modified cost with the added fundamental-frequency penalty, even though no update rule or gradient for that penalty is derived.","fun_headline_variants_meta":{"raw":{"variants":["LLM feedback loop tunes NMF to sharpen heart-lung sounds","First LLM-guided NMF separates cardiac and respiratory sounds","Language model optimizes NMF for better heart-lung separation","LLM-driven NMF improves medical sound separation over prior methods","NMF with LLM feedback loop sharpens stethoscope sound separation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000532,"raw_usage":{"total_tokens":2482,"prompt_tokens":790,"completion_tokens":1692,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":406,"completion_tokens_details":{"reasoning_tokens":1604}},"tokens_in":406,"tokens_out":1692,"duration_ms":15059,"temperature":1.0,"reasoning_tokens":1604,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T18:05:32.946434+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the gradient of the penalty term $\\lambda_f \\|\\mathbf{f}_{\\mathrm{est}}(\\mathbf{A},\\mathbf{X}) - \\mathbf{f}\\|^2$ with respect to $\\mathbf{A}$ and $\\mathbf{X}$ and check whether the unmodified multiplicative updates (3)-(4) form a descent step for it; if the penalty can increase, LingoNMF is not minimizing the stated objective. A paired experiment replacing the LLM's frequency choices with a grid search or random schedule on the same datasets would then settle whether the LLM feedback, rather than the penalty itself, produces the reported SDR gains.","supporting_citations":[],"review_version":1}