{"id":"55978f56-55bd-4276-924e-b957e5d6d75e","arxiv_id":"2506.13455","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Replacing the Conformer decoder with a bidirectional Mamba and asymmetric convolutions in a pretrained CNN14 PSELDnet improves stereo sound event localization F1 to 39.6% on the DCASE2025 Task 3 development set, over the Conformer's 38.2% and the baseline's 22.8%.","lead":"A team reports a sound-event localization system for the DCASE 2025 stereo task that swaps the standard Conformer audio decoder for a bidirectional Mamba module with asymmetric convolutions, and reports the best F1 score among their tested models. It shows a lightweight way to squeeze better spatial audio detection out of a pretrained model, relevant for anyone building hearing aids, robots, or smart speakers.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed BiMambaAC advantage over the Conformer decoder rests on a single-run comparison with unmatched hyperparameters; the 1.4-point F1 gap may reflect batch size, learning rate, and step count rather than architecture.","rationale":"The reader's weakest_assumption is correct and is the same concern I identify. I considered whether the pseudo-FOA construction in Eq. (4) or the near-identical MAC counts were more serious, but the pseudo-FOA transform is invertible and full fine-tuning can adapt to it, while the MAC issue is secondary to the performance claim. The uncontrolled hyperparameters directly undermine the headline architecture comparison. Since the paper is a technical report and the large gain over the baseline is likely robust, CONDITIONAL remains the appropriate verdict; the authors should provide a matched-hyperparameter, multi-seed comparison to support the specific BiMambaAC-over-Conformer conclusion.","tokens_in":7049,"tokens_out":6480,"duration_ms":67985,"concrete_test":"Train BiMambaAC and Conformer under identical hyperparameters, first with lr=3e-5, wd=5e-6, bs=32 and then with lr=1e-4, wd=1e-4, bs=256, for at least 3 random seeds each. Compare the mean and standard deviation of development-set F1 and DOAE. If the BiMambaAC F1 advantage over Conformer does not reproduce in both conditions or falls within seed noise, the architecture-specific superiority claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing assumption is that the architecture differences, not the training recipes, explain Table 1. Section 4 reports different learning rates, weight decays, and batch sizes per model (BiMambaAC: lr 3e-5, wd 5e-6, bs 32; Conformer: lr 1e-4, wd 1e-4, bs 256). With batch size 32 versus 256, BiMambaAC receives roughly 8x more parameter updates per epoch under a different learning-rate schedule, and no model is run with more than one seed. The abstract's 'significantly better' over the Conformer decoder therefore rests on a 1.4-point F1 and 0.1-degree DOAE difference with no error bars. If the recipe, not the BiMamba module, drives this gap, the central claim collapses. The MAC comparison (4.63 vs 4.69 G) also shows that the 'reduced computational complexity' claim is marginal in FLOPs, even though parameter count is much lower.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a stereo sound event localization and detection (SELD) system for the DCASE2025 Task 3 audio-only track. The authors convert stereo input into a pseudo-FOA representation, fine-tune pretrained PSELDnet models, and replace the Conformer decoder in the CNN14-Conformer architecture with a bidirectional Mamba (BiMamba) module combined with asymmetric convolutions, calling the resulting model BiMambaAC. On the development set, BiMambaAC is reported to achieve an F1 score of 39.6%, DOAE of 15.8 degrees, and RDE of 33% with 76M parameters and 4.63G MACs, compared with 38.2%, 15.9 degrees, 33%, 210M parameters, and 4.69G MACs for the Conformer variant, and 22.8%, 24.5 degrees, and 41% for the official baseline. The paper concludes that BiMambaAC is significantly better than the Conformer decoder while reducing computational complexity.","tokens_in":7229,"tokens_out":4426,"duration_ms":39898,"significance":"If the reported comparison were robust, the paper would provide useful evidence that state-space decoders can replace Transformer decoders in a pretrained SELD pipeline with fewer parameters and comparable or better accuracy. The use of public challenge data, a standardized evaluation protocol, and reporting of parameter counts and MACs are strengths. However, the central architecture comparison currently rests on a single run per model with deliberately different training hyperparameters, so the headline claim of significant improvement is not yet supported. The incremental nature of the technical change and the absence of statistical significance testing limit the contribution to a suggestive empirical result rather than a definitive architecture comparison.","major_comments":[{"comment":"The central comparative claim that BiMambaAC outperforms the Conformer decoder is supported only by a single run per model with uncontrolled training hyperparameters. Section 4 reports different learning rates, weight decays, and batch sizes for each model (BiMambaAC: lr 3e-5, wd 5e-6, bs 32; Conformer: lr 1e-4, wd 1e-4, bs 256). With a fixed 120-epoch budget, the batch size difference alone gives BiMambaAC roughly eight times more parameter updates per epoch, so the 1.4-point F1 gap (39.6% vs. 38.2%) and 0.1-degree DOAE gap (15.8 degrees vs. 15.9 degrees) in Table 1 cannot be attributed to the BiMamba architecture. Please run multiple seeds and either match the training recipe across models or perform a hyperparameter sweep; without this, the abstract's 'significantly better' is unsupported.","section":"Section 4 and Table 1"},{"comment":"The conclusion that BiMambaAC achieves 'the best overall performance across all metrics' is not consistent with Table 1: HTS-AT has a lower RDE (30% vs. 33%), and BiMambaAC's RDE is identical to that of Conformer and BiMamba. The advantage over Conformer is limited to F1 and DOAE, with DOAE differing by only 0.1 degree. Please revise the abstract and conclusion to state precisely which metrics are improved and by how much.","section":"Section 5 and Conclusion"},{"comment":"The 'reduced computational complexity' claim is overstated as written. Table 1 shows BiMambaAC at 4.63G MACs versus 4.69G for Conformer, a 1.3% reduction, while the parameter count drops from 210M to 76M. If the claim refers to parameter count, it should say so explicitly; if it refers to MACs, the difference is marginal and should not be described as a major advantage, especially since HTS-AT already operates at 2.88G MACs.","section":"Section 5, Table 1"},{"comment":"The paper does not state whether the BiMamba module and the asymmetric convolution layers are initialized from pretrained weights or randomly initialized, whereas the Conformer decoder inherits PSELDnet pretrained weights. If the BiMamba components are randomly initialized, the comparison mixes initialization and architecture effects. Please disclose the initialization scheme and, ideally, train the Conformer baseline from the same random initialization to isolate the contribution of the architecture change.","section":"Section 3.3 and Section 4"}],"minor_comments":[{"comment":"The Conformer DOAE is reported as 16.6 degrees in the text but 15.9 degrees in Table 1; please correct and unify the value.","section":"Section 5, paragraph 2"},{"comment":"Setting X(n)=Z(n)=0 discards all non-horizontal spatial information; please state explicitly that the pseudo-FOA representation is azimuth-only and discuss the implications for the evaluation, which already excludes elevation.","section":"Section 3.1, Eq. (4)"},{"comment":"The sentence 'the random seed is set to be 42' implies a single seed; please state the number of seeds used per configuration, consistent with the request for multiple runs in the major comments.","section":"Section 4"},{"comment":"The capitalization of PSELDnet/PSELDNet is inconsistent; please standardize the spelling.","section":"Throughout"},{"comment":"The baseline row is crowded ('Baseline0.7M57M'); please format the table for readability.","section":"Table 1"},{"comment":"Several references are listed as arXiv preprints; where peer-reviewed versions now exist, please update the citation details.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"This is a DCASE challenge technical report with a promising but preliminary empirical result. The main gap is experimental rigor: single-seed runs and per-model hyperparameters make the headline architecture comparison unreliable, and the computational-complexity claim is much weaker than the abstract suggests. These issues are fixable within the manuscript's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a modest engineering report from the DCASE2025 stereo SELD track. The genuinely new bit is the combination: PSELDnet pretraining with a BiMamba decoder plus asymmetric convolutions, applied to stereo input via a pseudo-FOA conversion. That conversion and the folded-azimuth ACS augmentation are sensible practical choices, and the paper is honest about what it did. The absolute result that pretrained models crush the small CRNN baseline is credible and likely robust. Credit is due for reporting parameter counts and MACs, and for comparing multiple decoder variants on the same development set.\n\nThe soft spots are real and one is load-bearing. Every model in Table 1 was trained with a different batch size, learning rate, and weight decay, on a single seed. BiMambaAC uses batch 32 and lr 3e-5; the Conformer uses batch 256 and lr 1e-4. With 120 epochs and a ReduceLROnPlateau schedule, those differences mean very different optimization dynamics. The abstract's 'significantly better' rests on a 1.4-point F1 gap and no error bars. That is not significant in any statistical sense. I agree with the stress-test note: the architecture advantage is not established until the training recipes are matched or the comparison is repeated across seeds. The computational-complexity claim is also weaker than advertised: MACs are 4.63G versus 4.69G, so you save parameters, not FLOPs. Missing code further limits reproducibility, though the pseudo-FOA trick is simple enough to reimplement.\n\nNone of this is a fatal flaw if the paper is positioned as a technical report with a provisional result. But it should not be read as a rigorous architecture comparison. For a workshop audience it is okay; for a journal or conference paper it needs multiple seeds, controlled hyperparameters, and ideally an ablation isolating the asymmetric convolution and the BiMamba module separately. The authors are clearly thinking about the problem and are not hiding their configurations; they just need stronger experimental discipline.\n\nWho gets value from this? Practitioners in SELD who want a compact decoder and are willing to re-run the comparison properly. As a desk editor, I would send this to peer review rather than desk-reject, because the direction is reasonable and the weaknesses are fixable with reasonable effort. The letter should ask for seed variance and matched hyperparameters before any claim of superiority is accepted.","headline":"Plausible engineering tweak; the 'significantly better' claim is undercut by uncontrolled hyperparameters.","tokens_in":7763,"tokens_out":2286,"would_cite":false,"duration_ms":24689,"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":"By swapping the Conformer decoder for a bidirectional Mamba with asymmetric convolutions, this paper reports the best stereo SELD scores on the DCASE2025 Task 3 development set, using roughly a third of the Conformer's parameters.","keywords":["stereo sound event localization and detection","SELD","BiMamba","selective state space model","asymmetric convolution","PSELDnet","DCASE 2025 Task 3","pseudo-FOA conversion"],"falsifier":"Retrain the Conformer and BiMambaAC models under identical hyperparameters (same seed, learning rate, batch size, weight decay) or a small shared hyperparameter grid, and check whether BiMambaAC keeps its 1.4-point F1 lead over Conformer; a Conformer configuration that matches or exceeds 39.6% F1 would falsify the paper's central claim.","tokens_in":6839,"feed_emoji":"🎧","tokens_out":7392,"duration_ms":64537,"temperature":0.7,"pith_summary":"The paper proposes a stereo sound event localization and detection (SELD) system that replaces the Conformer decoder in a pretrained PSELDnet with a bidirectional Mamba (BiMamba) sequence model augmented by asymmetric convolutions. Its central claim is that this BiMambaAC model outperforms both the original Conformer-based PSELDnet and the CRNN baseline on the DCASE2025 Task 3 development set, while using fewer parameters and fewer MACs than the Conformer. If true, it shows that selective state-space models are a computationally cheaper alternative to attention for SELD on ordinary stereo audio, and that pretrained FOA-oriented models can be transferred to stereo via pseudo-FOA reconstruction.","feed_headline":"BiMamba decoder beats Conformer on stereo sound-event localization","feed_subtitle":"A bidirectional Mamba plus asymmetric convolutions hits 39.6% F1 with 76M parameters in DCASE 2025 Task 3.","key_machinery":"The central object is the BiMamba module: a bidirectional selective state-space model that processes a sequence forward and backward and fuses the hidden states, combined with asymmetric convolutions that alternate between the time and frequency dimensions. This decoder replaces the Conformer in the CNN14-Conformer PSELDnet. A second load-bearing step is the conversion of stereo left-right signals into pseudo first-order-ambisonic (FOA) channels, $W = (L+R)/2$ and $Y = (L-R)/2$ with $X = Z = 0$, which lets the pretrained FOA-based PSELDnet be fine-tuned on stereo data. The mechanism is meant to capture both directions of temporal context and to decouple time-frequency features, handling overlapping events better than the CRNN baseline.","core_discovery":"On the paper's own terms, the discovery is that the CNN14-BiMamba hybrid — a pretrained CNN14 encoder with a BiMamba decoder plus asymmetric convolution — reaches a location-dependent F1 of 39.6%, a DOA error of 15.8°, and a relative distance error of 33% on the DCASE2025 Task 3 development set, with 76M parameters and 4.63G MACs. This beats the Conformer decoder (38.2% F1, 15.9° DOAE, 33% RDE, 210M parameters) and the CRNN baseline (22.8%, 24.5°, 41%, 0.7M), and is the best configuration the authors tested. The paper also claims that the asymmetric convolutions are responsible for part of the gain, since BiMamba without them scores 36.2% F1 with 178M parameters.","pith_inferences":["The same decoder swap could transfer to other audio tasks that currently pair a pretrained encoder with an attention decoder, though the paper does not test this.","Because the stereo-to-pseudo-FOA reconstruction discards the X and Z ambisonic components, an open question is whether a binaural or stereo-native pretraining objective would preserve spatial information that the current pipeline loses.","The azimuth folding to [-90°, 90°] exploits the fixed frontal perspective of this year's challenge; an obvious extension is testing whether the architecture handles full 360° azimuth when labels are not folded.","The channel-swap augmentation and pseudo-FOA trick could apply to other fixed-perspective SELD settings, such as smart displays or robotics, without collecting new spatial audio data."],"forward_implications":["If the result holds, BiMamba can serve as a drop-in Conformer replacement in pretrained SELD networks, cutting parameters by roughly a factor of three without losing accuracy.","The pseudo-FOA reconstruction trick suggests that large FOA-pretrained models can be adapted to stereo-only consumer content rather than requiring new stereo pretraining.","Asymmetric convolutions appear to be a cheap way to shave parameter count and improve discrimination of overlapping events, which may transfer to other time-frequency audio tasks.","The reported 39.6% F1 on the development set provides a concrete reference point that future stereo SELD systems can compare against."],"supporting_citations":[{"why":"Supplies the pretrained PSELDnet backbones (CNN14-Conformer, HTS-AT) that the paper fine-tunes and modifies.","marker":"[13]"},{"why":"Defines the selective state-space model (Mamba) that the BiMamba decoder is built from.","marker":"[18]"},{"why":"Shows that BiMamba can replace the MHSA part of a Conformer in speech enhancement, motivating the decoder swap in this paper.","marker":"[20]"},{"why":"Prior work replacing Conformer with BiMamba in the EINV2 SELD framework, the direct precedent this paper extends.","marker":"[21]"},{"why":"Source of the Audio Channel Swapping augmentation used to double the stereo training data.","marker":"[24]"},{"why":"The DCASE2025 Task 3 stereo SELD development dataset on which all reported metrics are measured.","marker":"[27]"},{"why":"Defines the MultiACCDOA output format and ADPIT training objective used by the models and baseline.","marker":"[12]"}],"fun_headline_variants":["BiMamba trims model size, wins SELD against Conformer","Asymmetric convs push BiMamba SELD to 39.6% F1","BiMamba SELD: 76M params, 4.63G MACs, top F1","BiMamba decoder beats Conformer with 2.8x fewer params"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes that the single training run per model, with different learning rates, batch sizes, and weight decay for each architecture, is a fair way to rank architectures; if those hyperparameters, rather than the BiMamba design, explain the score gap, the headline claim collapses.","fun_headline_variants_meta":{"raw":{"variants":["BiMamba trims model size, wins SELD against Conformer","Asymmetric convs push BiMamba SELD to 39.6% F1","BiMamba SELD: 76M params, 4.63G MACs, top F1","BiMamba decoder beats Conformer with 2.8x fewer params"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00026,"raw_usage":{"total_tokens":1559,"prompt_tokens":881,"completion_tokens":678,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":497,"completion_tokens_details":{"reasoning_tokens":586}},"tokens_in":497,"tokens_out":678,"duration_ms":6434,"temperature":1.0,"reasoning_tokens":586,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:00:13.197724+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the Conformer and BiMambaAC models under identical hyperparameters (same seed, learning rate, batch size, weight decay) or a small shared hyperparameter grid, and check whether BiMambaAC keeps its 1.4-point F1 lead over Conformer; a Conformer configuration that matches or exceeds 39.6% F1 would falsify the paper's central claim.","supporting_citations":[{"cited_title":"Multi-ACCDOA: Localizing and Detecting Overlapping Sounds from the Same Class with Auxiliary Duplicating Permutation Invariant Training,","cited_arxiv_id":null,"evidence_quote":"Defines the selective state-space model (Mamba) that the BiMamba decoder is built from."},{"cited_title":"PANNs: Large-Scale Pretrained Audio Neural Networks for Audio Pattern Recognition,","cited_arxiv_id":null,"evidence_quote":"Shows that BiMamba can replace the MHSA part of a Conformer in speech enhancement, motivating the decoder swap in this paper."},{"cited_title":"Ef- ficient Training of Audio Transformers with Patchout,","cited_arxiv_id":null,"evidence_quote":"Prior work replacing Conformer with BiMamba in the EINV2 SELD framework, the direct precedent this paper extends."},{"cited_title":"Mamba: Linear-Time Sequence Modeling with Selective State Spaces,","cited_arxiv_id":null,"evidence_quote":"Source of the Audio Channel Swapping augmentation used to double the stereo training data."},{"cited_title":"SELD-Mamba: Selective State-Space Model for Sound Event Localization and Detection with Source Distance Es- timation,","cited_arxiv_id":null,"evidence_quote":"The DCASE2025 Task 3 stereo SELD development dataset on which all reported metrics are measured."}],"review_version":2}