{"id":"4d0ff196-91ff-4791-bbee-3bb167f9b823","arxiv_id":"2504.14371","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Spiking Point Mamba combines Mamba state-space layers with spiking neurons and dynamic encoding to achieve state-of-the-art SNN accuracy on point cloud classification and part segmentation.","lead":"This paper builds a spiking neural network that uses the Mamba sequence model to analyze 3D point clouds, claiming top accuracy among spiking models with lower estimated energy. It introduces a dynamic encoding method and a custom spiking Mamba block, plus a spike-based pre-training scheme.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported gains rest on a self-reproduced SPT baseline; the +6.2/+6.1/+7.4 margins need an official-SPT re-run before they are treated as established.","rationale":"The paper is internally coherent and the ablations support the proposed components, so I do not see a mathematical objection to the SMB or HDE designs. The weakest point is empirical: the comparison baseline is self-reproduced and under-specified, and the authors are also the authors of SPT, so the baseline is not independent. The reader identified exactly this concern. In addition, Table 4 contains a likely transcription error in the ANN Mamba row, which reduces confidence in the tables but does not by itself overturn the central architecture claims. My recommendation is unchanged: conditional acceptance, contingent on code release, baseline re-run with variance reporting, and correction of the energy citation and table inconsistency.","tokens_in":17061,"tokens_out":10955,"duration_ms":100464,"concrete_test":"Obtain the official SPT implementation and run it under the SPM evaluation protocol (same input points, tokenization, T=4, optimizer, and 300 epochs), for at least three seeds, and report mean and standard deviation. Then recompute the ScanObjectNN and ShapeNetPart deltas against this distribution. If the official SPT mean lies within 1% of the starred 78.0/82.8/83.4 PB-T50-RS/OBJ-BG/OBJ-ONLY values, the +6.2/+7.4/+6.1 margins hold; if it lies above SPM's 84.2 on PB-T50-RS, the SOTA claim fails. The same run also resolves whether Table 4's ANN Mamba row was mis-transcribed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"SPM's headline improvement is quantified relative to SPT (Spiking Point Transformer), and all SPT entries in Tables 1-3 are starred as self-reproduced by the authors, who also authored SPT. No code, random seeds, or variance estimates are provided, and the supplementary gives only single starred numbers. If the re-run used a different point count, token budget, timestep, or training schedule than the original SPT, the advertised margins shrink or vanish. This is load-bearing because the central claim is specifically a +6.2%, +6.1%, and +7.4% OA gain and a +1.9% instance mIOU gain over the prior SNN SOTA; the entire SOTA claim is a delta over this baseline. Even if SPT moves, SPM may remain best among the listed SNNs, but the specific headline numbers would not be substantiated. A secondary consistency issue: in Table 4 the ANN Mamba row lists 90.2/89.5/84.2/92.3, which are SPM's own numbers, while Table 1 reports PointMamba as 90.2/89.8/85.4/92.4; this suggests the comparison tables need an audit before the energy/performance trade-off is taken at face value.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Spiking Point Mamba (SPM), a Mamba-based spiking neural network for 3D point cloud classification and segmentation. It introduces three components: Hierarchical Dynamic Encoding (HDE), which makes direct encoding time-dependent by varying the point sampling across time steps; a Spiking Mamba Block (SMB) that adapts the Mamba gate and SSM branches to spike-based computation with a time-flip bidirectional strategy; and an asymmetric SNN-ANN masked pre-training scheme with an ANN decoder. On ScanObjectNN, ModelNet40, and ShapeNetPart, the authors report SNN state-of-the-art results, with gains of +6.2%, +6.1%, and +7.4% OA over Spiking Point Transformer (SPT) on the three ScanObjectNN variants and +1.9% instance mIoU on ShapeNetPart. They also report a theoretical energy reduction of at least 3.5x relative to the ANN counterpart. Component-wise ablations in Tables 4-6 support the individual design choices, and the supplementary provides additional ablations on neurons and masking ratios.","tokens_in":17373,"tokens_out":7432,"duration_ms":63970,"significance":"If the empirical claims are confirmed, SPM would be a useful contribution: it is the first Mamba-based 3D SNN, it combines temporal dynamics with state-space modeling, and its ablations isolate the effect of each proposed component. The energy model in Eqs. (18)-(20) is transparent as a first-order analytical estimate. However, the practical significance depends on the exact deltas against SPT, and those deltas currently rest on self-reproduced numbers without code or seeds, with theoretical rather than measured energy figures, and with an internal inconsistency in the reported ANN baseline. The manuscript is within scope for a computer-vision venue, but the headline claims need to be substantiated more carefully.","major_comments":[{"comment":"The headline improvements of +6.2%, +6.1%, and +7.4% OA and +1.9% instance mIoU are deltas against Spiking Point Transformer (SPT), and all SPT entries are marked as self-reproduced by the authors while Ref. [52] is the authors' own prior work. No code, random seeds, or reproduction details are provided, so the reader cannot determine whether the margin reflects the proposed architecture or a favorable re-implementation. I request the original SPT numbers (or a public checkpoint), a detailed reproduction protocol, and results from multiple seeds with mean and variance for both SPT and SPM before the SOTA claim can be accepted.","section":"Tables 1-3, §5.2.1"},{"comment":"Tables 1, 4, and 7 report inconsistent numbers for the same ANN baseline: PointMamba is listed as 90.2/89.8/85.4/92.4 in Table 1, 'ANN Mamba' is listed as 90.2/89.5/84.2/92.3 in Table 4, and Table 7 uses 90.2/85.4/92.4 (only three columns). Because Table 4 is used to claim that SPM's SMB performs on par with Mamba while reducing energy, these discrepancies must be corrected and the exact configuration (including whether the stars indicate self-reproduction) clarified.","section":"Tables 1, 4, and 7"},{"comment":"The 'at least 3.5x lower energy' claim is a theoretical estimate computed from FLOPs and fixed per-operation pJ constants, not a hardware measurement. The text does not say how the mJ values in Table 7 are obtained (per sample, including memory/data movement, with which firing-rate statistics), and the energy constants are attributed to Ref. [1], which is a neuron-model paper rather than the usual source for these values. I recommend relabeling the claim as an analytical estimate and, if the energy advantage is a central contribution, adding measurements on a neuromorphic simulator or GPU.","section":"§4.3, Eqs. (18)-(20), Table 7"},{"comment":"The pre-training comparison with SPT is uncontrolled. Table 2 lists SPT at 82.6/92.5, but the text does not state whether SPT underwent the same masked pre-training on ShapeNet with the same mask ratio, epochs, and fine-tuning schedule, or whether this is the from-scratch number. Without this control, the claims that spike-based pre-training adds +2.3/+0.8 points over SPM from scratch and beats SPT by +3.9/+1.4 points are not substantiated.","section":"Table 2, §5.2.1"}],"minor_comments":[{"comment":"Algorithm 1 uses S both as the input size to FPS and as the sampled point set, and the quantities L, M, R, E, and F are not defined consistently (for example, M is reassigned as M+L). Please restate the algorithm with distinct symbols and explicit index bounds.","section":"Algorithm 1"},{"comment":"In Eq. (12), the left-hand side assigns two tensors U'_n and U'_t from a single SSM call; the time-flip indexing and the resulting shapes need to be stated explicitly, and Eq. (13) should explain how SN(EAP(Z_n)) broadcasts after average pooling over the token dimension.","section":"Eqs. (12)-(13)"},{"comment":"The reference for the energy constants is incorrect: Ref. [1] (Brette and Gerstner) does not report 4.6 pJ/MAC and 0.9 pJ/AC. Please cite the original source (for example, Horowitz 2014) or remove the citation.","section":"§4.3, Ref. [1]"},{"comment":"There are several typographical errors and reference issues: 'obversed' in Sec. 4.1.1, 'Segmantation' in Fig. 2, 'classfication' in Table 8, and Refs. [53] and [54] appear to duplicate the same paper. A careful proofread is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The central risk is that the SOTA deltas are measured against the authors' own SPT reimplementation. I would ask the editor to require official SPT numbers or a shared-code reproduction before publication. The Table 4/Table 1 inconsistency and the absence of energy-measurement methodology are further grounds for major revision. The manuscript is otherwise in scope for a computer-vision venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is the first Mamba-based SNN for 3D point clouds, and it brings real components—HDE dynamic encoding, SMB with time-flip, and a spike-based pretraining scheme—rather than a trivial swap. The ablations in Tables 4-6 support each piece, and the supplementary's FPS-stage analysis gives a reasonable rationale for HDE. Credit where due: the authors identify an actual mismatch between static direct encoding and temporal feature extraction, and the design choices (time-flip rather than token-flip, Gate II) are empirically motivated.\n\nThe soft spots are real but not fatal. The headline margins over SPT (+6.2/+6.1/+7.4 OA) are computed against a self-reproduced baseline from the authors' own prior work, with no official SPT numbers, no code, no seeds, no error bars. That does not prove foul play, but it does mean the deltas are not established; a re-run with the official SPT release (or code from this paper) is needed. A related editing issue: Table 4's 'ANN Mamba' row lists 90.2/89.5/84.2/92.3, which are exactly SPM's own numbers, while Table 1 reports PointMamba as 90.2/89.8/85.4/92.4. That looks like a copy-paste error, and it needs an audit before anyone takes the energy/performance trade-off at face value.\n\nThe energy figures are theoretical estimates (MAC vs AC energy with assumed firing rates), not hardware measurements—standard for the SNN literature, but worth labeling clearly as estimates. The ablation across time steps shows the familiar plateau, and the claim of 3.5x lower energy is tied to those estimates.\n\nOverall, the central architectural claim is plausible and the paper is coherent. The comparison issue is load-bearing only for the specific delta numbers, not for the qualitative conclusion that SPM moves the SNN state of the art on ScanObjectNN. This deserves serious peer review, with a request for code, variance estimates, and a corrected Table 4. I'd bring it to a reading group as a good example of SNN-Mamba adaptation, but I wouldn't cite the headline numbers until the baseline question is resolved.","headline":"First Mamba-based SNN for point clouds with real components and plausible ablations; headline deltas over SPT need a baseline re-run and a table audit before they are treated as established.","tokens_in":17870,"tokens_out":2399,"would_cite":true,"duration_ms":20237,"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":"Spiking Point Mamba claims to be the first Mamba-based spiking neural network for point clouds, improving on prior spiking models by 6.2–7.4 percentage points while using at least 3.5x less energy than its non-spiking counterpart.","keywords":["Spiking neural networks","Point cloud analysis","Mamba","State space models","Hierarchical dynamic encoding","Spiking Mamba Block","Energy-efficient inference","Masked pre-training"],"falsifier":"Re-run the previous state-of-the-art spiking point transformer with its original hyperparameters and data splits, and compare directly with SPM; if its accuracy reaches or exceeds the self-reproduced numbers SPM improves on (82.8, 83.4, and 78.0 on the three ScanObjectNN variants), the claim of +6.2 to +7.4 point gains collapses. Separately, measure SPM's energy on real neuromorphic hardware or with an independent MAC/AC counter; if the ratio to PointMamba falls below 3.5x at four time steps, the efficiency claim is not general.","tokens_in":16917,"feed_emoji":"⚡","tokens_out":8178,"duration_ms":68118,"temperature":0.7,"pith_summary":"The paper tries to show that a spiking neural network (SNN) can get the long-range sequence modeling of Mamba without giving up event-driven efficiency, and that this combination is the right way to push 3D point-cloud accuracy in the SNN world. It proposes Spiking Point Mamba (SPM), built from a Hierarchical Dynamic Encoding (HDE) that makes the input tokens change over time and a Spiking Mamba Block (SMB) that adapts Mamba's gating and state-space branches to binary spikes. On top of that it adds masked spike-based pre-training with a small ANN decoder. The reported results are large gains over the previous SNN state of the art—up to +7.4% overall accuracy on ScanObjectNN and +1.9% instance mIoU on ShapeNetPart—with at least 3.5x lower energy than the ANN counterpart. A sympathetic reader cares because if these numbers hold, energy-efficient spiking models become a realistic option for point-cloud tasks rather than a niche with a large accuracy penalty.","feed_headline":"Spiking Mamba tops point-cloud SNNs by up to 7.4 points","feed_subtitle":"First Mamba-based spiking network for 3D point clouds uses at least 3.5x less energy than its non-spiking twin","key_machinery":"The load-bearing mechanism is the Spiking Mamba Block (SMB): a Mamba block rewritten so that its state-space branch and gate branch operate on binary spikes from Leaky Integrate-and-Fire (LIF) neurons. A LIF neuron accumulates a leaky membrane potential, emits a spike when the potential crosses a threshold, then resets. In SMB, the SSM branch reverses the time dimension of the spike feature before passing it through the selective state-space model, so the model sees past and future time steps, while the gate branch uses element-wise average pooling over the token dimension before the Hadamard product, a change the paper claims prevents the information loss that raw spike-gating causes. HDE supplies the SMB with a time-varying input stream, and the asymmetric SNN-ANN encoder-decoder pre-training keeps the SMB encoder spike-driven during inference.","core_discovery":"The central claim is that the poor performance of previous 3D SNNs comes from three fixable problems: static input encoding, missing long-range dependencies, and information loss when spike tensors pass through gates and state-space updates. SPM fixes the first with HDE, which splits farthest-point sampling into early, middle, and late stages and slides or extends the selected points across time steps so each time step sees a slightly different, hierarchically meaningful subset. It fixes the second and third with SMB, in which the SSM branch scans spike features with time-dimension reversal (Time Flip) rather than token reversal, and the gate branch averages gating activations across tokens before the Hadamard product to avoid dropping sparse spike information. The paper reports that the full SPM reaches 90.2/89.5/84.2 overall accuracy on the three ScanObjectNN variants and 92.3% on ModelNet40, exceeding the self-reproduced Spiking Point Transformer by 7.4/6.1/6.2/0.9 points, and 84.8% instance mIoU on ShapeNetPart, with 12.6x and 3.5x energy reductions at one and four time steps versus PointMamba.","pith_inferences":["Editorial inference: HDE's stage-wise treatment of farthest-point sampling should transfer to other ordered point-cloud tokenizers, such as octree or grid orderings, so the dynamic-encoding benefit is likely not specific to FPS.","Editorial inference: the paper's energy math assumes a 4.6 pJ MAC versus 0.9 pJ AC at 32-bit precision; under lower-precision ANN MACs the 3.5x margin would shrink, so the efficiency claim is hardware-precision-dependent.","Editorial inference: the mask-ratio peak at 0.6 and the Time-Flip result suggest a simple recipe for future spiking sequence models: keep the temporal axis bidirectional, keep the token axis unidirectional, and pre-train with a moderate mask.","Editorial inference: the supplementary neuron ablation shows that swapping in ILIF raises SPM to 93.0% on ModelNet40 and 85.2% on PB-T50-RS, so adopting ILIF as the headline configuration would widen the reported margins over prior SNNs."],"forward_implications":["If the reported accuracies hold, a 12-layer spiking Mamba backbone can match or approach ANN point-cloud classifiers on ModelNet40 (92.3% versus 92.4% for PointMamba) while running spike-driven, event-based computation.","The +2.3 and +0.8 point gains from masked spike pre-training on ScanObjectNN and ModelNet40 imply that SNN encoders can be improved by ANN decoders without sacrificing spike-driven inference.","The 3.5x (12.6x at one time step) energy advantage suggests that at equal accuracy, the spiking alternative is the lower-energy choice on neuromorphic or sparse-accumulation hardware.","Because Mamba's state-space scan is linear in sequence length, SPM inherits a path to longer point sequences and larger point clouds than quadratic-attention spiking transformers.","The time-flip rather than token-flip ablation indicates that the temporal dimension is the useful axis for bidirectionality in spike-based point-cloud Mamba, a design rule for future SNN sequence models.","The asymmetric SNN-ANN pre-training recipe—spiking encoder, lightweight ANN state-space decoder, Chamfer Distance reconstruction—provides a ready-made template for self-supervised learning in other 3D SNN backbones."],"supporting_citations":[{"why":"This is the Mamba paper that introduces the selective state-space model on which SMB's SSM branch is based, providing the long-range sequence-modeling capacity SPM claims to bring to SNNs.","marker":"[11]"},{"why":"PointMamba is the ANN architecture SPM adapts to the spiking domain and the direct energy-comparison counterpart used in Tables 1, 3, 4, and 7.","marker":"[24]"},{"why":"Spiking Point Transformer is the previous SNN state of the art; the starred self-reproduced results in Tables 1 through 3 are the baseline whose improvements define SPM's headline margins.","marker":"[52]"},{"why":"Spiking PointNet is an earlier SNN point-cloud classifier whose static direct encoding and results are compared against SPM, motivating HDE's dynamic encoding design.","marker":"[42]"},{"why":"P2SResLNet-B is an energy-oriented spiking point-cloud baseline used in the classification comparison table.","marker":"[54]"},{"why":"Point-MAE is a masked autoencoder point-cloud pre-training baseline that motivates the masked modeling with an asymmetric encoder-decoder adopted for spike-based pre-training.","marker":"[36]"}],"fun_headline_variants":["First spiking Mamba for 3D point clouds reduces energy 3.5x","Spiking Mamba lifts 3D point cloud accuracy by up to 7.4%","Mamba-powered spiking network tops 3D SNNs, uses 3.5x less energy","Spiking Point Mamba: first Mamba-based SNN for point clouds","3D point cloud spiking Mamba boosts accuracy 7.4%, cuts energy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline gains are measured against a Spiking Point Transformer baseline that the authors re-ran themselves; if that re-run was under-tuned relative to the original model, the reported +6.2, +6.1, and +7.4 point improvements could be inflated.","fun_headline_variants_meta":{"raw":{"variants":["First spiking Mamba for 3D point clouds reduces energy 3.5x","Spiking Mamba lifts 3D point cloud accuracy by up to 7.4%","Mamba-powered spiking network tops 3D SNNs, uses 3.5x less energy","Spiking Point Mamba: first Mamba-based SNN for point clouds","3D point cloud spiking Mamba boosts accuracy 7.4%, cuts energy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000324,"raw_usage":{"total_tokens":1886,"prompt_tokens":1084,"completion_tokens":802,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":700,"completion_tokens_details":{"reasoning_tokens":686}},"tokens_in":700,"tokens_out":802,"duration_ms":6376,"temperature":1.0,"reasoning_tokens":686,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:49:40.026354+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the previous state-of-the-art spiking point transformer with its original hyperparameters and data splits, and compare directly with SPM; if its accuracy reaches or exceeds the self-reproduced numbers SPM improves on (82.8, 83.4, and 78.0 on the three ScanObjectNN variants), the claim of +6.2 to +7.4 point gains collapses. Separately, measure SPM's energy on real neuromorphic hardware or with an independent MAC/AC counter; if the ratio to PointMamba falls below 3.5x at four time steps, the efficiency claim is not general.","supporting_citations":[{"cited_title":"Spiking point transformer for point cloud classification,","cited_arxiv_id":null,"evidence_quote":"Spiking Point Transformer is the previous SNN state of the art; the starred self-reproduced results in Tables 1 through 3 are the baseline whose improvements define SPM's headline margins."},{"cited_title":"Spiking pointnet: Spik- ing neural networks for point clouds.Advances in Neural Information Processing Systems, 36, 2024","cited_arxiv_id":null,"evidence_quote":"Spiking PointNet is an earlier SNN point-cloud classifier whose static direct encoding and results are compared against SPM, motivating HDE's dynamic encoding design."},{"cited_title":"Point-to-spike residual learning for energy- efficient 3d point cloud classification","cited_arxiv_id":null,"evidence_quote":"P2SResLNet-B is an energy-oriented spiking point-cloud baseline used in the classification comparison table."},{"cited_title":"Masked autoencoders for point cloud self-supervised learning","cited_arxiv_id":null,"evidence_quote":"Point-MAE is a masked autoencoder point-cloud pre-training baseline that motivates the masked modeling with an asymmetric encoder-decoder adopted for spike-based pre-training."}],"review_version":1}