{"id":"1a5f5ef7-0019-47d5-bfc9-8c4a0ca1a654","arxiv_id":"2505.13617","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A direction-aware neural field trained with an intensity-vector loss predicts first-order Ambisonic room impulse responses and adapts to new rooms from one or few examples.","lead":"A new neural network model learns to estimate room impulse responses in first-order Ambisonic surround-sound format, including the direction sound arrives from, using a small number of measurements. It adds a direction-aware loss based on sound intensity vectors and shows the model can adapt to a new room with as few as one example.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The few-shot early-stopping holdout may be the same 200 points as the test set, making Warm-Start and LoRA few-shot comparisons optimistic; this is the central unresolved issue.","rationale":"The reader's weakest assumption is exactly the concern I would raise: the early-stopping holdout and the test set are both described as 200 points, and the paper never states they are disjoint. This is not an internal inconsistency, but it is an unresolved ambiguity that is load-bearing for the few-shot adaptation claims, which are a central part of the paper. The IV-loss claim is supported by the single-room experiments and is not affected by this issue. Because the concern is concrete, testable, and was already identified by the reader, I agree with the conditional verdict. No additional objection rises to the same level of importance.","tokens_in":8403,"tokens_out":3976,"duration_ms":38575,"concrete_test":"Check the released code or ask the authors for the exact split indices used in Section 4.3. Verify whether the 200-point early-stopping holdout is the same set as the 200-point test set. If they overlap, re-run Table 1 for Warm-Start and LoRA with a disjoint holdout (for example, reserve 200 new points from the same room, or use a validation split from the training data) and report updated T60, C50, EDT, and DoA for the 1- and 80-training-example cases. If any entry changes by more than 20% relative, the few-shot equivalence claim should be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 4.3, the paper states 'Each environment C′ has 200 points in its test set' and then says 'we implemented early stopping on a 200 sample holdout set for those cases' (referring to Warm-Start models with fewer than 80 training examples). The paper nowhere states that this holdout is disjoint from the test set. If the same 200 points are used for both early stopping and final evaluation, the training epoch is selected using the test data, so the reported few-shot errors for Warm-Start, and any comparison against it, are optimistically biased. This directly affects the central claim that rank-3 LoRA with less than 1% of parameters achieves near-equivalent performance to Warm-Start in the low-data regime: if Warm-Start used the test set for early stopping while LoRA did not, the equivalence comparison is unfair. The few-shot adaptation results in Table 1 are a major contribution, so this ambiguity is load-bearing. It does not undermine the IV-loss result in Table 2 and Figure 3, which uses a separate per-room train/test split of 800/200 without early stopping.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DANF, a direction-aware neural field that predicts first-order Ambisonic room impulse responses as a function of source position, listener position, listener orientation, and a room-geometry context encoded through bounce points. The main contributions are: (i) the first neural-field model for Ambisonic-format RIRs, (ii) an intensity-vector loss based on the cosine distance between averaged intensity vectors of the predicted and reference RIRs, and (iii) an empirical study of few-shot adaptation to novel rooms using warm-start fine-tuning, LoRA with rank 1 and 3, cold-start training, and zero-shot prediction. Experiments are carried out on Matterport3D rooms rendered with SoundSpaces. The reported results suggest that the intensity-vector loss substantially improves direction-of-arrival estimation, that a moderate loss weight can be beneficial, and that rank-3 LoRA with far fewer parameters is competitive with full warm-start fine-tuning in low-data regimes.","tokens_in":8617,"tokens_out":4754,"duration_ms":43567,"significance":"If the claims hold, the paper makes a useful step beyond monaural and binaural neural acoustic fields by modeling directional sound-field information in Ambisonic form. The intensity-vector loss is a simple and sensible inductive bias, and the DoA improvements in Table 2 are consistent across all ten rooms at λ=10. The few-shot adaptation study, especially the LoRA comparison, is practically relevant for deploying neural acoustic fields in new environments. The paper does not release code or trained models, and all results are based on simulated data, so reproducibility rests entirely on the manuscript details. The explicit strengths are the architectural adaptation of INRAS to Ambisonic output, the novel IV loss, and the low-rank adaptation comparison with parameter counts.","major_comments":[{"comment":"The early-stopping protocol for low-data warm-start fine-tuning is ambiguous and potentially test-set-contaminating. Section 4.3 states that each target environment C' has 200 points in its test set and that the Warm-Start models with fewer than 80 training examples were early-stopped on a 200-sample holdout set, but the paper nowhere states that this holdout is disjoint from the test set. If the same 200 points are used both for early stopping and for final evaluation, then the training epoch is selected using test data, and the Warm-Start errors (and any comparison against them, including the LoRA equivalence claim in Table 1) are optimistically biased. Please state explicitly whether the holdout is disjoint from the test set; if it is not, rerun the affected experiments with a separate validation split and update Table 1 and the associated claims.","section":"§4.3 (Table 1)"},{"comment":"The claim that the intensity-vector loss at λ=10 improves 'both room acoustic metrics as well as directional performance' is not supported by the per-room results in Table 2. Averaging over the ten rooms, T60 error drops from 1.59% to 0.99%, but C50 error increases from 2.58 dB to 2.80 dB and EDT error increases from 56.8 ms to 59.6 ms; only 5 of 10 rooms improve in C50 and 6 of 10 improve in EDT. The direction-of-arrival improvement is clear and consistent, but the claim of simultaneous acoustic-metric gains at λ=10 should either be removed or qualified, unless Figure 3, which is described as showing the same single-room setup, uses a different aggregation or different data. Please reconcile Table 2 with the text and Figure 3.","section":"§5.1 (Table 2)"},{"comment":"The λ-sweep results are presented as point curves without error bars, repeated runs, or significance tests, despite the sweep being used to draw the load-bearing conclusion that λ=10 is a sweet spot. With only ten rooms and a single training run per configuration, the apparent improvements at λ=10 in T60, C50, and EDT could be driven by one or two rooms, and the per-room Table 2 already shows non-uniform behavior. Please provide per-room variance or significance information, or at minimum state the number of runs per condition, to support the 'improves both room acoustic metrics' conclusion.","section":"§5.1 (Fig. 3)"},{"comment":"The few-shot comparisons in Table 1 are based on single runs and do not report the pretraining data, the number of rooms used for pretraining, the architecture hyperparameters, or the variability across training examples. Since the 1-example and 80-example regimes are highly sensitive to initialization and example selection, the equivalence between Warm-Start and LoRA(3) should be supported by multiple fine-tuning seeds (or bootstrap over training subsets). Without this, the 'near equivalent performance' claim in Section 5.2 is not yet quantitatively established.","section":"§4.3, §5.2 (Table 1)"}],"minor_comments":[{"comment":"The definition of the spectral convergence term L_sc appears to contain a typo: the numerator should be |||H_hat| - |H|||_2, with a hat on the predicted magnitude, not |||H| - |H|||_2.","section":"Eq. (5)"},{"comment":"The x-axis label reads 'log10 10', which appears to be a formatting artifact; it should likely read 'log10 λ'.","section":"Fig. 3 caption"},{"comment":"There is a typo in 'matricies' (should be 'matrices'), and the author list formatting in the title page shows an unusual space in 'Franc ¸ois'.","section":"§4.3"},{"comment":"The dataset description says 100 unique rooms are available from five buildings, but the single-environment experiments use only 10 rooms with no explanation of how those rooms were selected; please state the selection criterion and list or characterize the room volumes.","section":"§4.1"},{"comment":"Equation (1) uses W, X, Y, Z for the STFT components but the preceding definition uses lowercase w, x, y, z for the FOA channels; please add a sentence clarifying the correspondence.","section":"§2"}],"recommendation":"major_revision","confidential_remarks":"The early-stopping ambiguity in Section 4.3 is the most serious issue: it directly affects whether the central few-shot claims in Table 1 are valid, and the paper must either confirm that the holdout is a separate 200-point set or rerun the affected experiments. The discrepancy between the text's claim of improved C50/EDT at λ=10 and the averages implied by Table 2 is also concerning and needs to be resolved before publication. I would not recommend rejection because the DoA improvement from the IV loss is consistent across rooms and the proposed loss is a reasonable contribution, but the presentation and experimental reporting need substantial revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this paper is worth engaging. It is the first neural-field model I have seen that targets first-order Ambisonic RIRs, and the intensity-vector loss actually does what it claims: across ten rooms, adding it at moderate weight improves DoA error consistently, and at λ=10 it also nudges T60/C50/EDT in the right direction. The architecture is a straightforward multi-channel extension of INRAS, but the application to Ambisonics is not just cosmetic — it lets the model capture per-point directionality that monaural/binaural NFs miss. The LoRA adaptation idea is sensible, and the few-shot framing is timely.\n\nSoft spots, in order of importance. First, the early-stopping protocol in Section 4.3 is ambiguous in exactly the way that matters. The paper says each target room has 200 test points, then says Warm-Start models used a 200-sample holdout for early stopping when training data is below 80. It never states that holdout is disjoint from the test set. If it is the same 200 points, then the training epoch is selected on the test data, and the Warm-Start few-shot numbers are optimistic — which would bias the comparison against LoRA. This needs a one-sentence clarification, but it is load-bearing for the adaptation claims. Second, all experiments are single-run with no error bars or significance tests; given the per-room variance visible in Table 2, a couple of runs with confidence intervals would materially change how much weight we put on \"near-equivalent.\" Third, Figure 3's λ sweep is evaluated on the test set to pick λ; that is model selection on the test set, mild but worth stating. Fourth, there are no baselines beyond Cold/Warm/LoRA — for instance, a simple nearest-neighbor interpolation or a geometric-acoustics fit would help calibrate how much of the gain comes from the NF.\n\nThe IV loss result, in contrast, looks solid: it is a clean idea, the per-room table shows consistent DoA improvement, and the failure mode at large λ (over-focusing on direct sound) is described honestly. No red flags in the math or citations.\n\nFor whom: spatial audio people working on room acoustics, neural rendering, and few-shot adaptation. It deserves a serious referee; with the holdout question resolved and some statistical rigor, it would be a solid conference paper. I would ask for a revision rather than reject.","headline":"A genuinely new application of neural fields to Ambisonic RIRs with a useful direction-aware loss, but the few-shot adaptation results rest on an early-stopping protocol that may leak the test set.","tokens_in":9176,"tokens_out":2406,"would_cite":true,"duration_ms":22138,"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":"The paper proposes a direction-aware neural field (DANF) that learns spatially continuous first-order Ambisonic room impulse responses from finite measurements, adds an intensity-vector loss to capture direction of arrival, and shows that…","keywords":["spatial audio","neural acoustic fields","Ambisonics","room impulse responses","direction-of-arrival estimation","intensity vector","few-shot adaptation","low-rank adaptation"],"falsifier":"Inspect the data splits: if the early-stopping holdout and the 200-point test set share indices, the reported warm-start and LoRA errors were selected on test data and are optimistic; re-running with a disjoint holdout and test set would settle whether the few-shot gains remain.","tokens_in":8196,"feed_emoji":"🔊","tokens_out":7255,"duration_ms":64238,"temperature":0.7,"pith_summary":"Sound in a room reaches a listener from many directions, but earlier neural-field models of room impulse responses treated sound as monaural or binaural only. This paper proposes DANF, a neural field that outputs first-order Ambisonic impulse responses, four channels that encode the pressure wave and its directional gradient, so the model can represent where sound comes from as well as what arrives. It adds an intensity-vector loss that compares the direction of energy flow in predicted and true responses; with a moderate weight this improves direction-of-arrival error without hurting standard acoustic measures, while an overly large weight sacrifices late reverberation. The paper further claims that a model pre-trained across many rooms can adapt to a new room from very few measurements, and that low-rank adaptation using under 1% of the parameters comes close to warm-start fine-tuning. If these claims hold, directional acoustic scene capture becomes practical in data-scarce settings such as virtual reality and telepresence.","feed_headline":"Direction-aware neural field maps room sound direction from few samples","feed_subtitle":"Intensity-vector loss improves direction and acoustics; low-rank tuning matches full retraining with under 1% of parameters.","key_machinery":"The central object is the first-order Ambisonic impulse response $h\\in\\mathbb{R}^{4\\times T}$, whose four channels $w,x,y,z$ capture the omnidirectional pressure and the three Cartesian pressure-gradient components. The direction-aware machinery is the intensity vector $IV(U)(m,f)$ computed from the STFT of these channels; the paper averages it over time and frequency into a single 3D vector and applies a cosine-distance loss $L_{IV}$ between prediction and target. That loss supplies the training signal that ties the network's output to direction of arrival. The architecture follows prior neural acoustic fields: bounce points and source/listener relative positions are encoded into spatial features, multiplied with sinusoidal time encodings to form a spatio-temporal embedding, then decoded with an MLP conditioned on listener orientation. Low-rank adaptation is the few-shot mechanism: rank-$r$ matrices $B\\in\\mathbb{R}^{i\\times r}$, $A\\in\\mathbb{R}^{r\\times j}$ update each MLP weight as $W' = W + \\frac{1}{r}BA$, so only a small parameter subset is trained for a new room.","core_discovery":"On the paper's own terms, DANF is the first neural acoustic field to estimate Ambisonic-format room impulse responses, predicting $h\\in\\mathbb{R}^{4\\times T}$ from source position, listener position and orientation, and a set of bounce points describing room geometry. The direction-aware component is an intensity-vector loss computed from the short-time Fourier transform of the four Ambisonic channels, averaged over time and frequency and compared to the target via cosine distance. Experiments on simulated rooms show this loss reduces direction-of-arrival error at all tested weights and, at $\\lambda=10$, also improves T60, C50, and EDT metrics; at very large weights the model over-focuses on the direct sound and distorts late reverberation. For adaptation, warm-start fine-tuning and rank-3 LoRA both outperform cold-start and zero-shot baselines, with LoRA using $2.9\\times 10^4$ parameters, less than 1% of the full model's $3.5\\times 10^6$, while roughly matching warm-start performance.","pith_inferences":["Editorial inference: the same intensity-vector loss could serve as a self-supervised objective for unlabeled Ambisonic recordings, since it is computed from the predicted and measured signals alone without needing room geometry.","Editorial inference: if the early-stopping holdout and test set are genuinely disjoint, the few-shot result is stronger than if they overlap; checking the split is the first test a reader should run.","Editorial inference: extending the loss to higher-order Ambisonics would test whether directional accuracy scales with spherical-harmonic order, likely requiring frequency-dependent weighting because higher-order channels are quieter.","Editorial inference: conditioning on bounce points suggests the model could be driven by estimated room geometry from depth sensors, enabling few-shot adaptation in real rooms rather than simulated ones."],"forward_implications":["Ambisonic RIRs predicted by DANF can be convolved with dry sources to render directional sound, so spatial audio can be synthesized at arbitrary source-listener positions without new measurements.","The intensity-vector loss gives a single auxiliary term that improves both direction-of-arrival and conventional room-acoustic metrics at a moderate weight, potentially replacing separate DoA post-processing.","Pre-training across rooms plus LoRA adaptation makes per-room customization feasible with under 1% of the model parameters, which matters in memory-limited devices.","Because the model is continuous over source and listener locations, a single trained field can interpolate directional impulse responses across an entire room from a finite measurement grid.","Even a single training example yields better T60 and C50 errors than zero-shot transfer for warm-start and LoRA models, indicating background spatial knowledge transfers across rooms."],"supporting_citations":[{"why":"Establishes neural acoustic fields as continuous RIR representations and provides the MSE training baseline.","marker":"[16]"},{"why":"Supplies the bounce-point spatial embedding and spatio-temporal decoding architecture that DANF extends to Ambisonics.","marker":"[17]"},{"why":"Defines the pre-training and fine-tuning benchmark that the few-shot adaptation comparison follows.","marker":"[20]"},{"why":"Defines the Ambisonics intensity vector used to build the direction-aware loss.","marker":"[24]"},{"why":"Complements the intensity-vector definition for B-format and Ambisonic source localization.","marker":"[25]"},{"why":"Provides the geometric acoustic simulation that generated the first-order Ambisonic RIR dataset.","marker":"[28]"},{"why":"Introduces low-rank adaptation, the mechanism for few-shot tuning with a constrained parameter count.","marker":"[29]"}],"fun_headline_variants":["Neural acoustic field tracks sound direction from few room samples","Direction-aware AI maps sound direction in rooms with few samples","Few-shot Ambisonic field learns room sound direction","DANF: neural field captures room sound direction from sparse data","Ambisonic neural field needs few samples to pinpoint sound direction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The few-shot comparison assumes the 200-sample holdout used for early stopping is a separate set from the 200-point test set, so the training epoch was not chosen using the test data's error.","fun_headline_variants_meta":{"raw":{"variants":["Neural acoustic field tracks sound direction from few room samples","Direction-aware AI maps sound direction in rooms with few samples","Few-shot Ambisonic field learns room sound direction","DANF: neural field captures room sound direction from sparse data","Ambisonic neural field needs few samples to pinpoint sound direction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000669,"raw_usage":{"total_tokens":3040,"prompt_tokens":927,"completion_tokens":2113,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":543,"completion_tokens_details":{"reasoning_tokens":2030}},"tokens_in":543,"tokens_out":2113,"duration_ms":13200,"temperature":1.0,"reasoning_tokens":2030,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:11:51.883342+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the data splits: if the early-stopping holdout and the 200-point test set share indices, the reported warm-start and LoRA errors were selected on test data and are optimistic; re-running with a disjoint holdout and test set would settle whether the few-shot gains remain.","supporting_citations":[{"cited_title":"Analysis of mutiple-shepers radiation and scattering problems by using a null-field integral equation approach,","cited_arxiv_id":null,"evidence_quote":"Establishes neural acoustic fields as continuous RIR representations and provides the MSE training baseline."},{"cited_title":"Parametric wave field coding for precomputed sound propagation,","cited_arxiv_id":null,"evidence_quote":"Supplies the bounce-point spatial embedding and spatio-temporal decoding architecture that DANF extends to Ambisonics."},{"cited_title":"Spatial extrapolation of early room impulse responses in local area using sparse equivalent sources and image source method,","cited_arxiv_id":null,"evidence_quote":"Defines the pre-training and fine-tuning benchmark that the few-shot adaptation comparison follows."},{"cited_title":"A V-NeRF: Learning neural fields for real-world audio-visual scene synthe- sis,","cited_arxiv_id":null,"evidence_quote":"Defines the Ambisonics intensity vector used to build the direction-aware loss."},{"cited_title":"Neural acoustic context field: Rendering realistic room impulse response with neural fields,","cited_arxiv_id":null,"evidence_quote":"Complements the intensity-vector definition for B-format and Ambisonic source localization."},{"cited_title":"Physics-informed machine learning for sound field estima- tion: Fundamentals, state of the art, and challenges,","cited_arxiv_id":null,"evidence_quote":"Introduces low-rank adaptation, the mechanism for few-shot tuning with a constrained parameter count."}],"review_version":1}