{"id":"fb5acc09-a7a0-4f85-82fc-898410a295c8","arxiv_id":"2507.09806","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"LoRA-adapted Deep Prior networks reconstruct room impulse responses as accurately as full fine-tuning while training only about 30% of the network parameters.","lead":"This paper shows that a low-rank adaptation trick called LoRA can fine-tune a deep prior network for reconstructing room impulse responses almost as well as full retraining, while updating only a fraction of the network's parameters. The finding matters because it could make adapting sound field reconstruction to new source positions or rooms much cheaper.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline NMSE comparison is vulnerable to selection bias: the LoRA rank r=16 is chosen on the same single-room source-move scenario used to report the within-1.6 dB result, so the 'comparable to full fine-tuning' claim may not survive a blind rank selection.","rationale":"The reader's conditional verdict already captures the main weaknesses of the paper, and the additional rank-selection concern I identify does not overturn that verdict; it strengthens the case for requiring a proper validation protocol. The most load-bearing issue is not the transfer-learning assumption per se, although the multi-room results temper the general claim. Rather, the 'comparable to full fine-tuning' statement is the paper's headline quantitative contribution, and that particular number is produced under a hyperparameter selection protocol that is not blind to the evaluation scenario. The paper is otherwise a plausible empirical study: the Deep Prior formulation, the LoRA injection defined in Eq. (5) and Eq. (6), and the parameter-count comparisons are internally consistent, and the multi-room table provides an honest negative result that the authors partially acknowledge. A conditional acceptance requiring a validation split, multiple seeds, and error bars is the appropriate call. I therefore leave the reader's verdict unchanged, while flagging that the concrete test I propose is the minimal experiment that would settle whether the 1.6 dB claim is robust or an artifact of rank selection.","tokens_in":8503,"tokens_out":6237,"duration_ms":74534,"concrete_test":"Withhold one adaptation scenario as validation: pretrain on source r'_1, use adaptation to a held-out source or room (e.g., one of the rooms in [45]) for rank selection, and evaluate on a different held-out scenario; report NMSE for r ∈ {1,2,4,8,16,32,64} with multiple seeds and error bars. If LoRA with a validation-selected rank remains within 1.6 dB of full fine-tuning and beats scratch at M = 4, the claim survives; if the best rank shifts between scenarios, the reported comparison is partly an artifact of test-set hyperparameter selection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim, that LoRA achieves NMSE comparable to full fine-tuning (within 1.6 dB) while training about 30% of parameters, rests on a rank value that was selected by inspecting the test scenario itself. In §3.1, Fig. 2(a) sweeps r = {1,2,4,16,32,64} on the single-room source-move task, identifies r = 16 as best, and Fig. 2(b) then reports the LoRA-versus-FT comparison in that same scenario using r = 16. This is a classic selection-on-the-test-set protocol: the reported 1.6 dB gap is conditioned on a hyperparameter that had access to the evaluation data. The multi-room section (§3.2) simply states that r = 16 'has been experimentally found to be effective also in this scenario', but does not report a separate validation procedure. Without a validation split, a different rank could plausibly change the comparison: Fig. 2(a) itself shows that rank-1 behaves very differently in the first 100 iterations, and the ordering at convergence is not shown as stable. Additionally, no error bars or repeated runs are provided, so we cannot tell whether the 1.6 dB difference is even outside run-to-run noise. The paper's parameter-count advantage is not affected by this issue, but the accuracy-comparability claim is directly tied to the biased rank choice.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes integrating Low-Rank Adaptation (LoRA) into a MultiResUNet-based Deep Prior model for room impulse response (RIR) reconstruction. The method freezes pretrained convolutional weights and learns low-rank adapter tensors, aiming to adapt a pretrained model to new source positions or rooms with few trainable parameters. The authors evaluate the approach in two scenarios: same-room source relocation (single room, M=32 microphones) and cross-room adaptation (three rooms with M=100 microphones). The main reported results are that LoRA achieves NMSE close to full fine-tuning (within 1.6 dB in the single-room case) while training roughly 30% of the parameters, and that fine-tuning is especially beneficial when only four microphones are available. In the multi-room setting, the results are mixed: LoRA and full fine-tuning sometimes underperform training from scratch, depending on the pretraining room. The paper concludes that adaptation strategies are promising for RIR reconstruction, particularly for source-move scenarios.","tokens_in":8754,"tokens_out":5568,"duration_ms":57385,"significance":"If the claims hold, this is a useful empirical contribution to parameter-efficient transfer learning for acoustic inverse problems, an area with little prior work. The LoRA formulation for convolutional layers is clearly presented and the experimental setup uses real measured RIR datasets, which strengthens the practical relevance. The parameter-count reduction (about 70% fewer trainable parameters for r=16) is a robust and practically meaningful achievement that is independent of the absolute NMSE ranking. The observation that low-rank adapters behave differently at short training budgets (Fig. 2a) could inform future work on very fast adaptation. The main limitations are the apparent selection of the rank hyperparameter on the evaluation scenario and the lack of uncertainty estimates, both of which affect the strength of the central accuracy-comparability claim. The paper also does not offer a criterion for when transfer will help, which is important given the mixed multi-room results.","major_comments":[{"comment":"The LoRA rank r=16 is selected by sweeping r on the same single-room source-move scenario that is subsequently used in Fig. 2(b) to report the 'within 1.6 dB' comparison between LoRA and full fine-tuning. Because no validation split or separate development scenario is used, this is effectively a selection-on-the-test-set protocol; the reported gap may be optimistic relative to a blind rank choice. Section 3.2 then reuses r=16 in the multi-room case without reporting a separate validation for that scenario. This concern directly affects the central accuracy-comparability claim, although the parameter-count advantage is not affected.","section":"Section 3.1, Fig. 2"},{"comment":"No error bars or repeated runs are reported in Fig. 2 or Table 1. The Deep Prior optimization involves random initialization, a randomly sampled fixed input tensor z, and randomly selected microphone subsets during adaptation, so run-to-run variability could be comparable in magnitude to the 1.6 dB difference observed at M=32. Without multiple seeds or some form of uncertainty quantification, the reader cannot assess whether the reported differences among DP, FT, and LoRA are statistically meaningful.","section":"Section 3.1 and Table 1"},{"comment":"The comparisons are made at a fixed training budget of 500 iterations for all methods and ranks, yet Fig. 2(a) itself shows that convergence behavior is rank-dependent: rank-1 performs best under 100 iterations, and the text suggests that r=64 might improve with a larger iteration span. Thus the 'comparable to full fine-tuning' conclusion is conditional on a single training budget, and the paper does not show that the ranking is stable across budgets. A convergence-matched or multi-budget comparison would substantially strengthen the efficiency claim.","section":"Section 3.1, Fig. 2(a)"},{"comment":"The multi-room results show that pretraining can hurt adaptation relative to training from scratch; for example, when pretraining on Balder and adapting to Freja with M=33, DP achieves -7.2 dB while FT and LoRA achieve -6.3 dB and -6.9 dB, respectively. The paper acknowledges this but provides no diagnostic or criterion for when transfer is expected to help. Since the paper's motivation is the reuse of pretrained models, this omission limits the practical guidance; a simple proxy (e.g., T30 mismatch or direction-of-arrival mismatch) or an explicit discussion of failure modes would be needed.","section":"Section 3.2, Table 1"}],"minor_comments":[{"comment":"The rank set is stated as r=(1,2,4,16,32,64), but Fig. 2(a) also includes r=8 and the parameter-percentage list contains seven values (0.2,4,7,14,30,60,122)%. Please reconcile the enumeration.","section":"Section 3.1"},{"comment":"The reverberation time is given as 'T60 ≈ 0.4 s' in the single-room setup and 'T30' values in the multi-room setup; please use consistent terminology and specify which reverberation metric is reported.","section":"Sections 3.1 and 3.2"},{"comment":"The sentence 'the differences in NMSE between LoRA and the best results are limited in all the cases below −1.81 dB' is ambiguous; presumably the magnitude of the difference is at most 1.81 dB, but the phrasing should be clarified.","section":"Section 3.2"},{"comment":"The caption states that best values are in bold and italic indicates when LoRA outperforms FT, but the formatted table as rendered does not visibly distinguish these cases; please ensure the typesetting is clear.","section":"Table 1"},{"comment":"The NMSE definition would benefit from an explicit note that the norm is the Euclidean norm over time samples; the notation ∥·∥ is used without definition.","section":"Equation (7)"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid empirical study, but the central accuracy-comparability claim is weakened by the selection of r on the test scenario and by the lack of error bars. I recommend asking the authors to either use a separate validation scenario for rank selection or to show that the conclusions are robust across a range of ranks. The mixed multi-room results should also be interpreted with more caution. These are fixable within the scope of the manuscript, so I am not recommending rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a first, honest empirical look at using LoRA to adapt a Deep Prior RIR reconstruction network to new source positions and rooms. That alone is useful: fine-tuning strategies for deep priors are genuinely understudied, and the parametric efficiency result is real. When only the source moves within one room, LoRA lands within about 1.6 dB of full fine-tuning while training roughly 30% of the parameters, and at M=4 it beats training from scratch. The multi-room experiment is also honestly reported: pretraining on Balder can actually hurt adaptation, and LoRA sometimes falls behind a retrained-from-scratch DP. The authors do not bury that.\n\nThe qualitative conclusions hold up. The headline quantitative claim is shakier, and for a specific reason: the LoRA rank r=16 was selected by sweeping ranks on the same single-room source-move scenario that is then used for the main LoRA-vs-FT comparison (Fig. 2a feeds directly into Fig. 2b). That is selection on the test set. The 1.6 dB gap is conditioned on a hyperparameter that already saw the evaluation data. It is not fatal—the parameter-count advantage is independent of the rank choice, and Figure 2a shows other ranks give roughly similar accuracy at convergence—but it means the \"comparable accuracy\" claim is softer than the abstract suggests. The paper also gives no error bars or repeated runs, so we cannot tell whether the remaining difference is within run-to-run noise. The multi-room section says r=16 was \"experimentally found to be effective\" but gives no separate validation.\n\nA few smaller things: the training budget is fixed at 500 iterations, there is no ablation on the scaling factor or learning rate, and there is no code release. None of these are load-bearing; they just cap how much we can trust the numbers.\n\nOverall this is a sensible, modest conference-style paper. The authors overstate the strength of the comparison slightly, but they state the limitations themselves in the conclusions. It deserves a serious referee. For an archival journal I would push for a validation protocol—rank selection on a separate split, multiple seeds, error bars—and a short discussion of when transfer is expected to help given that the multi-room case shows it can hurt. For a workshop or short paper it is near the bar as is.\n\nRecommendation: send to peer review; conditional accept with requests for a clean rank-selection protocol and variance reporting.","headline":"Useful first empirical study of LoRA for Deep Prior RIR reconstruction, with honest results and a real parameter-count win, but the headline accuracy gap is undercut by test-set rank selection and missing error bars.","tokens_in":9319,"tokens_out":2335,"would_cite":true,"duration_ms":25138,"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 shows that Low-Rank Adaptation fine-tunes a Deep Prior for RIR reconstruction to near full fine-tuning accuracy with roughly 70% fewer trainable parameters.","keywords":["room impulse response reconstruction","deep prior","low-rank adaptation","LoRA","transfer learning","sound field reconstruction","sparse microphone measurements"],"falsifier":"In the single-room scenario, run the same LoRA adaptation protocol with the source moved by more than 90 cm in the same room; if the NMSE advantage of the pretrained initialization over training from scratch vanishes or reverses at larger displacements, the paper's single-room transfer claim would be falsified.","tokens_in":8266,"feed_emoji":"🎧","tokens_out":8193,"duration_ms":77677,"temperature":0.7,"pith_summary":"The paper asks whether a Deep Prior network for room impulse response (RIR) reconstruction can be reused across acoustic setups by adapting only a small set of low-rank parameters, using the LoRA technique. It claims that, in the single-room case where only the source position changes, LoRA fine-tuning of a pretrained MultiResUNet matches full fine-tuning within 1.6 dB while training roughly 30% of the parameters, and that with as few as four microphones it clearly outperforms training from scratch. In the multi-room case, LoRA performs comparably to full fine-tuning and to retraining, although pretraining on a room with different reverberation and direction of arrival can hurt. The paper thereby establishes that parameter-efficient transfer learning is a viable route for per-element Deep Prior sound-field reconstruction, with the caveat that the benefit depends on the acoustic similarity of the pretraining and target configurations.","feed_headline":"LoRA-adapted Deep Prior matches full fine-tuning at 30% of parameters","feed_subtitle":"Transfer learning with LoRA cuts trained parameters by 70% while keeping reconstruction accuracy close to full fine-tuning.","key_machinery":"The machinery is the LoRA convolutional adapter: for a pretrained weight $W \\in \\mathbb{R}^{C_{\\text{out}} \\times C_{\\text{in}} \\times k \\times k}$, the update is a low-rank tensor $\\Delta W = \\alpha \\cdot BA$ with $A \\in \\mathbb{R}^{r \\times C_{\\text{in}} \\times k}$ and $B \\in \\mathbb{R}^{C_{\\text{out}} \\times k \\times r}$, contracted over $r$ and reshaped to the size of $W$. The adapted layer computes $Y(x) = (W + \\Delta W) * x$, so only $A$, $B$, and the scalar $\\alpha$ are learned. This carries the argument by letting the pretrained Deep Prior act as a fixed backbone while the adapter encodes the new acoustic configuration; because different adapters can be swapped in, one network can serve multiple setups. The Deep Prior itself supplies the implicit prior: a fixed random input tensor $z$ is mapped through the network to RIRs, fitting only the sparse measured positions with an $\\ell^1$ loss.","core_discovery":"The central claim is that Low-Rank Adaptation makes Deep Prior RIR reconstruction reusable across acoustic configurations. The authors embed LoRA into every convolutional layer of a MultiResUNet by writing each weight tensor as $W + \\Delta W$, with $\\Delta W = \\alpha \\cdot BA$ where $A$ and $B$ are low-rank factors and $W$ stays frozen; only the adapters are optimized. They compare this against full fine-tuning (FT) of all parameters and against the standard Deep Prior (DP) trained from scratch. With a single room and a changed source position, LoRA with rank $r=16$ (about 30% of the trainable parameters) achieves NMSE within 1.6 dB of FT for the full microphone array, and for $\\tilde{M}=4$ it is clearly better than DP from scratch. In the multi-room scenario, LoRA and FT perform comparably to each other and to DP, with LoRA sometimes best, though pretraining on the Balder room degrades adaptation because of its lower reverberation time and different source direction. The paper also reports that rank-1 adapters are instead preferable for very short adaptation runs (under 100 iterations).","pith_inferences":["Because the adapter $\\Delta W$ is small and swappable, the same pretrained Deep Prior could serve a library of acoustic configurations, and the set of adapters might be interpolated or averaged to approximate unseen configurations without any optimization.","The harmful case (pretraining on Balder) suggests a practical selection rule: compare coarse room descriptors such as reverberation time and source direction of arrival before deciding whether to adapt from a given pretrained model or retrain from scratch.","The optimum at $r=16$ indicates an intrinsic dimension of the adaptation task; measuring how this rank scales with room complexity or source displacement could guide adapter sizing for larger problems.","Since LoRA performs best with very few microphones, a natural extension is to combine it with active sensor selection or sequential measurement acquisition, where the adapter can be refined as new microphones arrive."],"forward_implications":["In the single-room scenario, a pretrained Deep Prior adapted with LoRA uses about 70% fewer trainable parameters than full fine-tuning with NMSE degradation of at most 1.6 dB, so hardware-constrained systems can reuse prior knowledge at a fraction of the update cost.","With very sparse measurements ($\\tilde{M}=4$), transferring from a pretrained network clearly outperforms a randomly initialized Deep Prior, suggesting that adaptation is most valuable exactly when data are scarce.","Rank-1 LoRA adapters are the best choice for very short adaptation runs (fewer than 100 iterations), implying a compute-versus-accuracy trade-off that can be tuned per deployment.","Across different rooms, LoRA, full fine-tuning, and training from scratch land within about 1.8 dB of each other, so adaptation does not consistently beat retraining; in fact, pretraining on a dissimilar room can be harmful."],"supporting_citations":[{"why":"Supplies the Deep Prior MultiResUNet architecture and the per-element training scheme used as the backbone for RIR reconstruction.","marker":"[33]"},{"why":"Establishes the Deep Prior paradigm that the paper adapts to the acoustic setting.","marker":"[28]"},{"why":"Introduces LoRA, the low-rank adaptation method that the paper transplants to convolutional layers.","marker":"[35]"},{"why":"Provides the convolutional LoRA formulation with the tensor factorization used to adapt the conv weights.","marker":"[41]"},{"why":"Defines the MultiResUNet architecture employed in the Deep Prior model.","marker":"[42]"},{"why":"Supplies the single-room dataset with a 32-sensor uniform linear array used in the source-move scenario.","marker":"[44]"},{"why":"Supplies the multi-room dataset (Balder, Munin, Freja) with 100 channels used for the cross-room adaptation experiments.","marker":"[45]"}],"fun_headline_variants":["LoRA fine-tuning matches full fine-tuning in RIR reconstruction","Transfer learning via LoRA beats retraining for room impulse responses","30% parameters: LoRA matches full fine-tuning for RIR tasks","LoRA adapts Deep Prior to new source positions with minimal cost","LoRA keeps RIR fidelity at 30% of full fine-tuning parameters"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a network pretrained in one acoustic configuration provides a useful starting point for another; if the pretraining and target rooms differ too much, as the Balder case shows, transfer can be worse than training from scratch.","fun_headline_variants_meta":{"raw":{"variants":["LoRA fine-tuning matches full fine-tuning in RIR reconstruction","Transfer learning via LoRA beats retraining for room impulse responses","30% parameters: LoRA matches full fine-tuning for RIR tasks","LoRA adapts Deep Prior to new source positions with minimal cost","LoRA keeps RIR fidelity at 30% of full fine-tuning parameters"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000428,"raw_usage":{"total_tokens":2232,"prompt_tokens":1032,"completion_tokens":1200,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":648,"completion_tokens_details":{"reasoning_tokens":1108}},"tokens_in":648,"tokens_out":1200,"duration_ms":10049,"temperature":1.0,"reasoning_tokens":1108,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:46:32.419070+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In the single-room scenario, run the same LoRA adaptation protocol with the source moved by more than 90 cm in the same room; if the NMSE advantage of the pretrained initialization over training from scratch vanishes or reverses at larger displacements, the paper's single-room transfer claim would be falsified.","supporting_citations":[{"cited_title":"Deep image prior,","cited_arxiv_id":null,"evidence_quote":"Establishes the Deep Prior paradigm that the paper adapts to the acoustic setting."},{"cited_title":"Multiresunet : Rethinking the u-net architecture for multimodal biomedical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Defines the MultiResUNet architecture employed in the Deep Prior model."},{"cited_title":"Ray-space-based multichannel nonnegative matrix factorization for audio source separation,","cited_arxiv_id":null,"evidence_quote":"Supplies the single-room dataset with a 32-sensor uniform linear array used in the source-move scenario."},{"cited_title":"Compressed sensing of impulse responses in rooms of unknown properties and contents,","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-room dataset (Balder, Munin, Freja) with 100 channels used for the cross-room adaptation experiments."}],"review_version":1}