{"id":"0c12ef03-549a-41ac-8924-29a2d321d6a4","arxiv_id":"2607.06274","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":4,"one_line_summary":"A source-position-dependent implicit neural representation for the directional weighting in a physics-constrained kernel enables sound field estimation that generalizes to unseen source positions without per-snapshot fine-tuning.","lead":"This paper makes a neural network learn how sound bounces around a room, so it can predict the sound field from new source positions without retraining. A smart generalist might read it because it could improve real-time spatial audio and acoustic imaging by making sound field estimation generalize across source locations.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"The SB-NK baseline lacks validation-based early stopping (fixed 300 epochs), creating an asymmetric advantage for LB-NK that confounds the central comparison supporting the paper's claim.","rationale":"The reader correctly identified the narrow experimental scope (single simulated room, no real data) as a weakness, and also noted the SB-NK early-stopping issue. However, I consider the baseline fairness issue to be the more load-bearing concern because it directly undermines the primary experimental evidence for the paper's central claim. The generalization-to-real-world concern is a scope limitation that the authors themselves acknowledge ('Future work is to generalize this approach to multiple rooms'), whereas the baseline confound is not acknowledged and affects the validity of the headline result. The reader's verdict of CONDITIONAL is appropriate, but the specific reason should foreground the baseline comparison fairness rather than (or in addition to) the generalization scope. The mathematical formulation is sound: the Helmholtz constraint is preserved by construction through the Herglotz kernel (§4.1), and the positive semi-definiteness of the kernel is guaranteed by the Softplus output (§4.2). The architecture and training procedure are clearly described. The concern is purely empirical: the evidence supporting the claim is confounded. If the concrete test above shows the gap persists under fair comparison, the paper's contribution is solid. If not, the contribution narrows to 'multi-source training enables a validation set,' which is less novel.","tokens_in":8516,"tokens_out":1856,"duration_ms":115700,"concrete_test":"Re-run the SB-NK baseline with a validation-based stopping criterion. For each test source, hold out a subset of microphones (e.g., 20% of M) from the measurement vector s, optimize SB-NK on the remaining 80%, and use the held-out microphones' reconstruction error for early stopping. If SB-NK with this fair early-stopping protocol achieves NMSE within 1 dB of LB-NK at the representative frequencies (150, 225, 300, 450, 750 Hz), the central claim that multi-source training is the key driver of improvement weakens substantially. If the gap persists, the claim is strengthened.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim rests on LB-NK outperforming SB-NK (Figs. 4–5). However, the experimental design introduces a confound: LB-NK benefits from validation-based early stopping (§5.1: 'The early-stopping criterion for LB-NK is based on the validation loss L'), while SB-NK is denied any validation mechanism ('Since SB-NK does not have the validation dataset, the number of epochs... is fixed at 300'). This asymmetry means LB-NK's advantage could stem partly from early stopping rather than from the multi-source training paradigm itself. If 300 epochs causes SB-NK to overfit (plausible given that the paper's own motivation is that single-snapshot optimization 'can lead to strong overfitting'), then the comparison does not isolate the contribution of source-position-dependent training. SB-NK could be given a fair early-stopping mechanism by holding out a subset of the M microphones as a validation set, or by using the known ground-truth ATF at a few evaluation points for model selection. Without this control, the reported performance gap conflates two factors: (1) the benefit of learning across sources, and (2) the benefit of having a validation set. The paper does not discuss or address this confound.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"This paper proposes a learning-based physics-constrained neural kernel (LB-NK) for sound field estimation. The key idea is to represent the directional weighting function in the Herglotz-wave-function-based kernel as an implicit neural representation (INR) that depends on the source position, trained on a set of acoustic transfer functions (ATFs) from multiple sources in a fixed room. This allows the kernel to generalize to unseen source positions without per-snapshot fine-tuning, addressing the overfitting problem of the snapshot-based neural kernel (SB-NK) of Ribeiro et al. [18]. The Helmholtz constraint is preserved by construction through the kernel formulation. Experiments on synthetic image-source-method data in a single shoebox room show that LB-NK outperforms SB-NK and a uniform-weighting baseline in normalized mean square error (NMSE) across frequencies.","tokens_in":9238,"tokens_out":1060,"duration_ms":165350,"significance":"The paper makes a reasonable contribution to physics-informed sound field estimation by extending the neural kernel framework from single-snapshot optimization to multi-source training. The mathematical framework is clean: the Herglotz wave function kernel (Eq. 4) preserves the Helmholtz constraint by construction, and the RKHS estimation (Eq. 3) is standard. The source-position-dependent INR is a natural and well-motivated extension. The Softplus output guarantees positive semi-definiteness of the kernel. The directional weighting visualizations in Fig. 6 provide qualitative evidence that the learned weights align with physical source and image-source directions. However, the experimental evaluation is limited to a single simulated room, and the comparison protocol between LB-NK and SB-NK has a methodological asymmetry that needs to be addressed.","major_comments":[{"comment":"§5.1, experimental design: LB-NK benefits from validation-based early stopping, while SB-NK is denied any validation mechanism and uses a fixed 300-epoch budget. This asymmetry confounds the central comparison. The paper's own motivation states that single-snapshot optimization 'can lead to strong overfitting'; if 300 epochs causes SB-NK to overfit, the reported performance gap may partly reflect the absence of early stopping rather than the benefit of multi-source training alone. A fairer comparison would provide SB-NK with a validation mechanism, e.g., by holding out a subset of the M microphones as a validation set for model selection. Without this control, the claim that LB-NK outperforms SB-NK due to source-position-dependent training is not fully isolated from the benefit of having a validation set.","section":null},{"comment":"§5.1–5.2, scope of evaluation: The entire experimental evaluation uses a single simulated shoebox room (T60=200 ms) with image-source-method data. The claim that the learned INR 'generalizes to unseen source positions' (Abstract, §4.1) is tested only within this narrow setting. Whether the learned directional weighting generalizes to real measured ATFs, non-shoebox geometries, or different reverberation times is unaddressed. The conclusion mentions 'future work to generalize to multiple rooms,' but the current scope is quite limited for a method whose central selling point is generalization. At minimum, the paper should discuss this limitation explicitly and clarify that the generalization claim is currently restricted to unseen sources within the same simulated room.","section":null}],"minor_comments":[{"comment":"§4.2, Eq. (6): The RFF bandwidth B is stated to be 'scaled proportionally to k,' but the proportionality constant is not given. Please specify.","section":null},{"comment":"§5.1: The regularization parameter λ in Eq. (3) is not stated. Was it the same for LB-NK, SB-NK, and Uniform? Please specify the value and whether it was tuned.","section":null},{"comment":"§5.1: The number of Lebedev quadrature points D is not stated. Please state the value used.","section":null},{"comment":"Fig. 4: The frequency axis labels are difficult to read. Consider enlarging or providing a table of key values.","section":null},{"comment":"Fig. 5: The boxplot frequency labels are small. Consider enlarging or annotating directly on the figure.","section":null},{"comment":"§5.2: The claim that LB-NK's tighter inter-quartile ranges indicate a 'structural regularizer' effect would be strengthened by reporting the number of test sources and whether the differences are statistically significant.","section":null},{"comment":"§3.2: The directed component (von Mises–Fisher superposition) from [18] is described, but it is unclear whether LB-NK also uses this directed component or only the INR. Please clarify the full LB-NK architecture used.","section":null}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about the asymmetric early-stopping protocol is well-founded and is the primary reason for the major_revision recommendation. This is a fixable experimental design issue, not a fundamental flaw in the method. The single-room evaluation is also a concern but may be acceptable for a short conference-style paper if the limitation is acknowledged. The paper is otherwise technically sound in its mathematical framework."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for the careful reading and constructive feedback. The referee raises two major points: (1) a methodological asymmetry in the comparison between LB-NK and SB-NK regarding validation-based early stopping, and (2) the limited experimental scope (single simulated shoebox room). We agree that both points warrant attention and will revise the manuscript accordingly. For the first point, we will add an additional SB-NK baseline that uses a held-out microphone subset for validation-based early stopping, to isolate the benefit of multi-source training from the benefit of having a validation set. For the second point, we will explicitly acknowledge the scope limitation of the generalization claim and clarify that it is currently restricted to unseen sources within the same simulated room.","responses":[{"response":"The referee is correct that the current comparison has a methodological asymmetry: LB-NK uses validation-based early stopping while SB-NK uses a fixed 300-epoch budget without any validation mechanism. We agree that this confounds the comparison and that a fairer protocol is needed to isolate the benefit of multi-source training from the benefit of having a validation set. We will address this in the revision by adding an additional SB-NK baseline that uses a held-out subset of microphones for validation-based early stopping. Specifically, we will hold out a subset of the M microphones as a validation set, optimize SB-NK on the remaining microphones, and use the validation loss for model selection. This will allow us to disentangle the contribution of the source-position-dependent INR from the contribution of early stopping. We note that even with this control, SB-NK still optimizes the directional weighting for a single snapshot without any mechanism to share information across source positions, so the core architectural advantage of LB-NK remains. However, we agree that the current presentation does not fully isolate this advantage, and the revised experiments will make the comparison fairer.","revision_made":"yes","referee_comment":"LB-NK benefits from validation-based early stopping, while SB-NK is denied any validation mechanism and uses a fixed 300-epoch budget. This asymmetry confounds the central comparison. A fairer comparison would provide SB-NK with a validation mechanism, e.g., by holding out a subset of the M microphones as a validation set for model selection."},{"response":"The referee is correct that the experimental evaluation is limited to a single simulated shoebox room with T60=200 ms, and that the generalization claim is currently tested only within this narrow setting. We agree that the abstract and the body of the paper should be more precise about the scope of the generalization claim. In the revision, we will explicitly state that the current evaluation demonstrates generalization to unseen source positions within the same simulated acoustic environment, and that generalization across different rooms, geometries, and reverberation times is left as future work. We will add this clarification to the abstract, Section 5.1, and the conclusion. Regarding additional experiments with real measured data or different room configurations: while we agree these would strengthen the paper, we are constrained by the availability of suitable multi-source ATF datasets with dense ground-truth evaluation points. We will, however, add a discussion of this limitation and outline the specific challenges and planned directions for extending the evaluation to real and diverse environments.","revision_made":"partial","referee_comment":"The entire experimental evaluation uses a single simulated shoebox room (T60=200 ms) with image-source-method data. The claim that the learned INR 'generalizes to unseen source positions' is tested only within this narrow setting. Whether the learned directional weighting generalizes to real measured ATFs, non-shoebox geometries, or different reverberation times is unaddressed. At minimum, the paper should discuss this limitation explicitly and clarify that the generalization claim is currently restricted to unseen sources within the same simulated room."}],"tokens_in":8354,"tokens_out":834,"duration_ms":74634,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"The main thing to know: this paper takes the physics-constrained neural kernel from Ribeiro et al. [18] and makes the directional weighting function source-position-dependent, trained across multiple sources instead of optimized per-snapshot. That is a legitimate and useful extension. The Helmholtz constraint is preserved by construction through the Herglotz wave function kernel, so the estimate still satisfies the wave equation regardless of what the INR learns. The closed-form inference (Eq. 3) means once trained, reconstruction is a linear operation — no per-source optimization at test time. That is the practical payoff and it is real. The directional weighting visualizations in Fig. 6 are genuinely informative: LB-NK concentrates weight on the true source and first-order image sources, while SB-NK spreads energy isotropically. That is the kind of evidence that makes the central claim believable. The math is clean and the architecture is reasonable — RFF embeddings for direction and source position, small MLP, Softplus for non-negativity. No invented machinery, no unnecessary complexity. Now the soft spots. The stress-test concern about the SB-NK baseline lacking validation-based early stopping is valid but not as damaging as it sounds. The paper itself motivates the whole approach by arguing that single-snapshot optimization overfits — so SB-NK overfitting at 300 epochs is partly the point. Still, the comparison would be cleaner if SB-NK had a fair stopping criterion, and the authors should address this. It is a confound, not a fatal flaw. The bigger issue is experimental scope. One simulated shoebox room, T60=200ms, image-source-method RIRs only. No real measurements, no varied geometries, no code or data released. Several parameters (λ, RFF bandwidth constant, quadrature order) are unspecified. The generalization claim rests entirely on held-out source positions within the same room. That is a meaningful test of within-room generalization, but it does not tell us whether the learned weighting transfers to real reverberant fields or different rooms. The authors acknowledge this — future work mentions multiple rooms. For a conference paper this scope is acceptable; for a journal submission it would need expansion. Who benefits: researchers in sound field estimation and physics-informed machine learning who want to understand how to adapt kernel methods across source positions. The paper is clearly written and the formulation is reproducible in principle, though the missing parameter values would require contacting the authors. This deserves a serious referee. The core contribution is sound, the math checks out, and the practical motivation is genuine. The experimental gaps are the kind that a revision cycle can address.","headline":"Clean extension of physics-constrained neural kernel to multi-source training; validation is narrow but the core idea is sound","tokens_in":9458,"tokens_out":604,"would_cite":true,"duration_ms":120751,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Neural kernel learns room acoustics from many sources, not one","keywords":["sound field estimation","kernel regression","implicit neural representation","physics-constrained neural kernel","directional weighting","Herglotz wave function","acoustic transfer function","room acoustics"],"falsifier":"If the source-position-dependent INR, trained on synthetic transfer functions from one room, produces reconstruction errors no better than the snapshot-based method when evaluated on measured acoustic transfer functions or on rooms with different geometries and reverberation times, the central claim of generalization would not hold.","tokens_in":8555,"feed_emoji":"","tokens_out":924,"duration_ms":113099,"temperature":0.7,"pith_summary":"The paper addresses the problem of estimating the spatial distribution of a sound field from a small set of microphone measurements. Prior physics-constrained neural kernel methods adapt a directional weighting function to a single snapshot of microphone data, which overfits and fails to generalize. The authors propose making the directional weighting function an implicit neural representation that takes both the propagation direction and the source position as inputs. By training this network on acoustic transfer functions from many source positions in a room, the kernel learns common directional patterns—such as direct sound and early reflection directions—and produces an appropriate weighting function for unseen source positions without requiring per-measurement optimization. Experiments on simulated room acoustics show lower reconstruction error and tighter error distributions compared to the snapshot-based approach, with the learned weighting function aligning to true source and image-source directions.","feed_headline":"Neural kernel learns room acoustics from many sources, not one","feed_subtitle":"A source-position-aware neural weighting function trained on multi-source data cuts sound field reconstruction error and overfits less than逐","key_machinery":"Herglotz wave function kernel, implicit neural representation (INR), Random Fourier Features, Lebedev quadrature, image source method","core_discovery":"The central mechanism is a source-position-dependent implicit neural representation (INR) for the directional weighting function inside a Herglotz-wave kernel. The Herglotz wave function expresses the acoustic field as a weighted integral of plane waves over the unit sphere; the weighting function determines which propagation directions the kernel emphasizes. By parameterizing this weighting as a neural network that conditions on both the propagation direction and the source position, and by training the network on transfer functions from multiple sources in a fixed room, the model learns a mapping from source position to directional pattern. At inference time, given a new source position, a","pith_inferences":["The alignment of learned weighting functions with image-source directions suggests the INR may be implicitly encoding a compressed representation of the room's image-source geometry, raising the question of whether explicit geometric priors could improve sample efficiency.","Training on 80 source positions in a single simulated room with moderate reverberation (T60=200ms) may not capture the diversity of real acoustic environments; generalization to measured data with diffuse late reverberation, non-convex geometries, or higher reverberation times remains the critical untested boundary.","The bandwidth scaling of the Random Fourier Features proportional to the wavenumber implies that higher frequencies require higher-capacity embeddings, which could create a computational bottleneck for broadband or high-frequency applications."],"forward_implications":["If the source-position-dependent INR generalizes beyond the tested shoebox room, it could enable pre-computed acoustic models for specific venues that produce accurate sound field estimates from sparse microphones without per-event calibration.","The multi-source training paradigm could be extended to multiple rooms, potentially learning a universal mapping from room geometry and source position to directional weighting, as the authors note in their conclusion.","The learned weighting functions visually align with image-source directions, suggesting the INR implicitly discovers the room's reflection structure, which could be extracted for room geometry inference or acoustic analysis.","Since inference reduces to a forward pass through the INR plus a linear kernel operation, the approach is compatible with real-time spatial audio applications."],"fun_headline_variants":["Sound field kernel conditions on source position to generalize","Source-aware neural directional weighting reduces acoustic overfitting","Physics-constrained kernel maps source position to directional pattern","Multi-source training lets neural kernel generalize to unseen positions","Herglotz wave kernel learns directional weighting per source position"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The claim of generalization rests on tests within a single simulated shoebox room with moderate reverberation; whether the learned directional weighting function transfers to real measured data, different room shapes, or more reverberant environments is not established.","fun_headline_variants_meta":{"raw":{"variants":["Sound field kernel conditions on source position to generalize","Source-aware neural directional weighting reduces acoustic overfitting","Physics-constrained kernel maps source position to directional pattern","Multi-source training lets neural kernel generalize to unseen positions","Herglotz wave kernel learns directional weighting per source position"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":614,"prompt_tokens":538,"completion_tokens":76,"prompt_tokens_details":null},"tokens_in":538,"tokens_out":76,"duration_ms":13193,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-08T11:09:49.337042+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If the source-position-dependent INR, trained on synthetic transfer functions from one room, produces reconstruction errors no better than the snapshot-based method when evaluated on measured acoustic transfer functions or on rooms with different geometries and reverberation times, the central claim of generalization would not hold.","supporting_citations":[],"review_version":1}