{"id":"fbde9277-fe4f-4828-854f-386655fdef7e","arxiv_id":"2505.14719","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A spiking vision transformer with multi-scale attention fusion (MSVIT) reaches 85.06% top-1 ImageNet accuracy with a linear-complexity attention mechanism and reports state-of-the-art results among compared spiking transformers.","lead":"MSVIT is a spiking vision transformer that adds multi-scale attention fusion, and it reports top-1 accuracy of 85.06% on ImageNet, above prior spiking transformer models. The design is notable because spike-based networks promise lower energy use than ordinary AI models.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Attribution of the ImageNet gain to MSSA is not established: Eq. 7 implements per-token gating rather than cross-token attention, and Table 3 never removes the SPEMSF/hierarchical/hybrid confounds.","rationale":"The reader identified attribution as the weakest assumption; I agree, and I add a sharper reason why attribution is load-bearing. If the column-sum fusion in Eq. (7) does not actually mix information across tokens or scales, then both the name 'multi-scale spiking attention' and the claimed linear-complexity attention mechanism are overstated, and the empirical margins over QKFormer may be due to the other simultaneously introduced components. The concern is not that 85.06% is impossible; it is plausible for a well-tuned SNN transformer at 69.8M parameters. The problem is that the paper does not provide the experiment that ties that number to its stated mechanism. Table 3 is the only ablation, and it varies only the Q/P inputs within MSSA, leaving SPEMSF, the hierarchical design, and the hybrid placement constant; it is also on CIFAR100 rather than the dataset of the headline claim. The proposed test settles the matter by comparing the full model against no-MSSA, no-SPEMSF, and no-hierarchy variants. If no-MSSA matches the full model, the central novelty claim fails even though the architecture might still be a useful engineering result. This does not change the reader's CONDITIONAL verdict: the paper should be accepted only if such attribution evidence is supplied.","tokens_in":14964,"tokens_out":9381,"duration_ms":96585,"concrete_test":"Run a factorial ablation in the released implementation: (A) final MSVIT; (B) same but with every MSSA block replaced by SSA, keeping SPEMSF and the hierarchy; (C) same but with SPEMSF replaced by a standard patch embedding, keeping MSSA; (D) same but with the three stages flattened to a single scale, keeping MSSA and SPEMSF. Use the exact ImageNet-1K recipe of Section 4.2 with at least three seeds. If B is within seed noise of A, or if C/D account for as much degradation as B, then Eq. (7)'s fusion is not the source of the SOTA claim, and the central novelty claim should be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reported 85.06% ImageNet-1K top-1 accuracy (Table 1) is plausible as an empirical result, but the paper's central novelty claim—that MSSA's multi-scale attention fusion is what produces the gain—is unsupported. First, the mechanism itself does not perform cross-token attention in the usual sense. In Eq. (7)-(8), SUM_c(Q) and SUM_c(P) reduce each token's feature vector to a scalar, and the output is MSSA[n,:] = SN(α_q[n] + α_p[n])·V[n,:] (broadcast over features). With the pointwise Q/P/V projections described in Eq. (4)-(6), the output at token n depends only on token n; unless the projections are convolutional across tokens (which Section 3.2 does not specify), there is no Q-K-style interaction among tokens and no recombination of multi-scale features across spatial locations. The 'multi-scale attention' is therefore a per-token gain, not attention fusion. Second, the ablation in Table 3 only swaps which source feeds the column-sum (P, Q, P+P, Q+Q, P+Q) while holding SPEMSF, the hierarchical stages, and the hybrid MSSA/SSA placement fixed. It is run on CIFAR100 only, and it never compares against a model with MSSA removed or SPEMSF removed. The margins over QKFormer (0.84-1.29% at matched scales, Table 1) are therefore just as plausibly produced by the conv-based SPEMSF, the three-stage hierarchy, or the larger parameter count as by Eq. (7). Appendix E's firing-rate comparison (0.306 vs 0.285) does not isolate the mechanism.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"MSVIT is a hierarchical spiking vision transformer for image and neuromorphic classification. It introduces Multi-Scale Spiking Attention (MSSA), which computes per-token scalar values by column-summing two spike-form projections Q and P, adding them, passing through a spiking neuron, and element-wise multiplying with V. It also introduces a Spiking Patch Embedding with Multi-Scale Feature Fusion (SPEMSF) and a three-stage hierarchy that uses MSSA in stages 1-2 and standard spiking self-attention (SSA) in stage 3. The authors report 85.06% top-1 accuracy on ImageNet-1K (69.80M parameters, 224x224 input, 4 time steps, trained from scratch), surpassing QKFormer by 0.84-1.29% across three model sizes, and they claim state-of-the-art status among SNN transformer architectures. Additional results on CIFAR10/100, DVS128-Gesture, and CIFAR10-DVS are provided, along with an ablation study on CIFAR100 that varies which projections feed the column-sum operation.","tokens_in":15347,"tokens_out":6176,"duration_ms":57419,"significance":"If the reported ImageNet result is reproducible, it is a strong empirical data point for spiking vision transformers, narrowing the gap to ANN-based transformers while maintaining low theoretical energy consumption. The code is public and the energy estimates follow standard 45nm assumptions. However, the paper's central novelty claim—that the accuracy gain comes from multi-scale attention fusion in Eq. (7)—is not supported by the experiments as reported. The mechanism is a per-token gating operation rather than a cross-token attention mechanism, and the ablation does not isolate it from SPEMSF, the hierarchical design, or the hybrid attention placement. The significance of the work therefore depends on additional controlled experiments that are currently missing.","major_comments":[{"comment":"The operation in Eq. (7) is per-token gating, not multi-scale attention. With pointwise or small-kernel convolutions producing Q, P, and V, SUM_c collapses the feature dimension to one scalar per token; the output is MSSA[n,:] = SN(alpha_q[n] + alpha_p[n]) * V[n,:], which scales V[n,:] without any Q-K style interaction across tokens. The paper describes this as 'multi-scale interactions' and 'attention scores,' which overstates the mechanism. Unless the projections in Eqs. (4)-(5) are explicitly spatial convolutions with multi-scale kernels—and Section 3.2 does not specify this—the module does not implement attention over tokens. Please clarify what cross-token or multi-scale information is actually fused, or rename the module and adjust the novelty claims accordingly.","section":"§3.2, Eq. (7)-(8)"},{"comment":"The ablation does not isolate MSSA. It only changes which sources (P, Q, P+P, Q+Q, P+Q) feed the column-sum, while keeping SPEMSF, the three-stage hierarchy, and the MSSA/SSA placement fixed. No baseline removes MSSA entirely or replaces it with SSA in stages 1-2, and no baseline removes SPEMSF. As a result, the ImageNet-1K gains over QKFormer in Table 1 (0.84-1.29%) cannot be attributed to Eq. (7); they could come from the convolutional SPEMSF, the hierarchical architecture, or the larger parameter counts. The Appendix E firing-rate comparison (0.306 vs 0.285) is descriptive and does not resolve this attribution. Additionally, the prose misreports Table 3 values: the text cites 81.25 and 81.15 for rows that read 81.36 and 81.56, respectively.","section":"Table 3 and §5"},{"comment":"The state-of-the-art claim is based on a limited baseline set and single-run numbers. The comparison omits SpikingResformer (cited in the Related Work) and other recent SNN transformers, and no variance or seed information is reported. Given that the margin over QKFormer is under 1.3 percentage points, the reported differences may not be statistically meaningful. Please add multiple runs with confidence intervals and expand the baseline table, or soften the SOTA claim.","section":"Table 1"}],"minor_comments":[{"comment":"The citation for Spikformer is given as [Yao et al., 2022]; it should be [Zhou et al., 2023], as Spikformer is introduced by Zhou et al. and the reference list correctly attributes it to them.","section":"§4.2"},{"comment":"There are typos in the table and surrounding text: 'Comparision' should be 'Comparison' and 'Transfomer' should be 'Transformer'; the table formatting should be cleaned up to avoid overlapping values.","section":"Table 2"},{"comment":"The channel setting 'C is set to 2' is ambiguous; please specify whether this is a channel multiplier or an absolute width, since the same phrasing appears in both the main text and the limitation appendix.","section":"§3.1 and Appendix E"},{"comment":"The projections Q and P are described as 'convolutional layers' in Section 3.2, but the exact kernel sizes are only given in Table 3 as 1x1 and 3x1 convolutions. Please state the kernel sizes and whether these convolutions operate across spatial positions in the main text.","section":"§3.2 and Table 3"},{"comment":"The statement that a higher firing rate (0.306 vs 0.285) implies that neurons 'transmit more information' needs support; firing rate alone is not a direct measure of information content, and the comparison does not control for architecture differences.","section":"Appendix E"}],"recommendation":"major_revision","confidential_remarks":"The reported 85.06% ImageNet result appears credible and would be a useful empirical contribution to the SNN transformer literature. However, the central attribution claim is currently unsubstantiated by the ablations, and the mechanism itself is described in a way that is likely to mislead readers. I recommend requesting additional controlled experiments (e.g., replacing MSSA with SSA in all stages, removing SPEMSF, and reporting multi-seed variance) before the work can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"For the record: this is a real paper with real experiments. The 85.06% ImageNet-1K top-1 at 4 timesteps, direct training, 69.8M params is a legitimate data point that beats QKFormer at matched scale, and the gains are consistent across three model sizes (0.84–1.29%). The complexity reduction from O(N^2D) to O(ND) by replacing matrix multiplication with column sums is real, and the hybrid MSSA/SSA stage design plus SPEMSF are reasonable engineering choices. I believe the numbers are probably reproducible.\n\nWhat is new: the column-sum fusion of Q and P before applying to V, and the SPEMSF patch embedding. Both are incremental twists on known spiking attention designs, but they are not in Spikformer, SDT, or QKFormer as far as I can tell. That is enough for a conference paper in this subfield.\n\nWhere it wobbles: the attribution claim. Eq. (7)-(8) with pointwise projections gives you a per-token gating vector, not spatially-mixed attention. So calling it \"multi-scale attention fusion\" oversells the mechanism; it is more like a learned per-token gain with a multi-scale flavor from the conv-based Q/P projections. The bigger problem is Table 3: it only swaps which source feeds the column sum (P, Q, P+P, Q+Q, P+Q), holding SPEMSF, the hierarchy, and the hybrid placement fixed. It never ablates MSSA out entirely, nor SPEMSF out, nor uses a non-hierarchical baseline. So the ImageNet gain could plausibly come from the conv-based embedding or the hierarchy rather than the column-sum trick. The Appendix E firing-rate comparison (0.306 vs 0.285) does not isolate anything. Also: single-run numbers, no error bars, and a sloppy citation (Spikformer is referenced as [Yao et al., 2022] in the experimental text, which is inconsistent with the actual Spikformer citation by Zhou et al.). These are fixable.\n\nThe circularity burden is low: the architecture is hand-tuned, not fitted to the test set, and energy numbers are standard estimates. No invented entities. I do not see a load-bearing flaw that would make the results false; I see an overclaim about what the results prove.\n\nWho this is for: people working on spiking transformers and neuromorphic vision. A serious referee should engage with it, mainly to demand a cleaner attribution study and seed/error-bar reporting.\n\nMy recommendation: send it to peer review with a request for major revision. The empirical result is worth checking, and the central novelty can be salvaged if they isolate the mechanism and temper the SOTA-claim language.","headline":"A plausible, incremental SNN transformer paper whose reported gains likely hold but whose central attribution claim is not proven by the current evidence.","tokens_in":15859,"tokens_out":696,"would_cite":true,"duration_ms":8814,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A spiking vision transformer with multi-scale attention fusion reaches 85.06% top-1 accuracy on ImageNet-1K, surpassing the previous SNN-transformer state of the art.","keywords":["spiking neural networks","vision transformer","multi-scale attention","spike-driven transformer","image classification","neuromorphic datasets","energy efficiency","hierarchical architecture"],"falsifier":"Train the same MSVIT architecture on ImageNet-1K but replace the MSSA column-sum fusion in stages 1-2 with the standard spiking self-attention, or with a concatenation-based fusion, while keeping SPEMSF, stage depths, and the training recipe fixed; if top-1 accuracy stays around 85%, the multi-scale fusion is not the cause of the reported gain.","tokens_in":14771,"feed_emoji":"⚡","tokens_out":5044,"duration_ms":50043,"temperature":0.7,"pith_summary":"This paper argues that the performance gap between spiking and conventional vision transformers comes in part from a single-scale bottleneck in spiking attention, and that a spike-compatible multi-scale fusion can close part of that gap. It introduces MSVIT, a hierarchical spiking transformer whose attention heads summarize feature projections at different scales by column sums, add those summaries, and gate the value spikes with the result. Directly trained from scratch on ImageNet-1K with 69.80M parameters and four time steps, MSVIT reports 85.06% top-1 accuracy, surpassing QKFormer's 84.22% and, the authors claim, becoming the state of the art among SNN-transformer architectures. The same recipe improves CIFAR and neuromorphic benchmarks, so the paper's claim is that multi-scale attention fusion helps across static and event-driven vision tasks.","feed_headline":"Spiking transformer hits 85.06% on ImageNet","feed_subtitle":"MSVIT fuses multi-scale spike attention to beat the previous SNN-transformer record on ImageNet-1K.","key_machinery":"The load-bearing mechanism is MSSA, defined as $\\text{MSSA}(Q,P,V) = \\text{SN}(\\text{SUM}_c(Q) \\oplus \\text{SUM}_c(P)) \\otimes V$, where $\\text{SUM}_c$ is a column-wise summation that turns each spike-form projection into an $N \\times 1$ attention vector. The fusion of the low-level projection Q (from a $1\\times1$ convolution) with the high-level projection P (from a $3\\times3$ convolution) replaces the $QK^T$ matrix multiplication, reducing attention complexity from $O(N^2 D)$ to $O(ND)$. The surrounding design—SPEMSF's multi-scale patch embedding, the hybrid placement of MSSA in early stages and standard spiking self-attention in the final stage, and the hierarchical token downsampling—supports but does not itself carry the novelty claim.","core_discovery":"According to the paper, the central discovery is that multi-scale spiking attention (MSSA) can replace the dot-product attention used in spiking transformers with a column-sum fusion that preserves spike-driven computation and linear complexity while improving accuracy. Each attention head computes two spike-form projections, Q and P, from convolutional paths with different receptive fields; summing over the feature dimension gives per-token attention scores in spike form, their element-wise sum gates V. The paper claims this fusion of low-level and high-level features enriches spike representations and, together with a hierarchical three-stage design and a multi-scale spiking patch embedding (SPEMSF), lifts ImageNet-1K top-1 accuracy to 85.06%, exceeding QKFormer and earlier SNN transformers.","pith_inferences":["The column-sum fusion effectively removes token-token interactions from attention; if the reported gain is real, it would imply that diverse input projections matter more than pairwise similarity for spiking transformers—an inference the paper does not state.","Because the ablation varies only which projections (Q, P, or both) are fused and not the new patch embedding, hierarchical stages, or hybrid attention placement, the accuracy gain may be partly caused by those components; a controlled comparison with SPEMSF added to QKFormer would isolate MSSA's contribution.","The fusion mechanism is a natural candidate for temporal multi-scale attention in event-stream tasks, where summing over time windows could play the same role as summing over scale, though the paper tests only spatial scales."],"forward_implications":["If the ImageNet-1K result is reproducible, directly trained SNN transformers can reach roughly 85% top-1 accuracy, within striking distance of conventional ANN transformers while using spike-based accumulate operations and a fraction of the estimated energy.","QKFormer-style Q-K attention is not necessary for a strong spiking transformer; fusing two spike projections by column sums can match or exceed it.","Linear-complexity attention of this form would let spiking transformers process higher-resolution inputs or longer token sequences without the quadratic cost of standard self-attention.","On CIFAR100 the model surpasses a ViT baseline, suggesting the multi-scale fusion helps close part of the SNN-ANN gap on smaller static datasets as well as on large-scale ImageNet-1K."],"supporting_citations":[{"why":"QKFormer is the main baseline to beat; its 84.22% top-1 accuracy and hierarchical spiking transformer design define the comparison point, and the paper adopts a similar training recipe.","marker":"[Zhou et al., 2024a]"},{"why":"Spikformer supplies the standard spiking self-attention (SSA) used in stage 3 and serves as the origin of the spike-driven transformer lineage that MSVIT extends.","marker":"[Zhou et al., 2023]"},{"why":"The Spike-driven Transformer is a key prior SNN-transformer baseline whose linear-complexity spike attention MSVIT is compared against at multiple model sizes.","marker":"[Yao et al., 2024b]"},{"why":"Spike-driven Transformer v2 extends the SDT line and provides additional state-of-the-art SNN baselines on ImageNet-1K.","marker":"[Yao et al., 2024a]"},{"why":"MetaFormer establishes the token-mixer/channel-mixer block structure that the MSFormer blocks adopt.","marker":"[Yu et al., 2022]"},{"why":"Swin Transformer supplies the hierarchical multi-scale design principle and is an ANN transformer baseline for both accuracy and energy comparisons.","marker":"[Liu et al., 2021]"},{"why":"Multiscale Vision Transformers motivate the multi-scale feature extraction hypothesis that MSSA is built on.","marker":"[Fan et al., 2021]"},{"why":"The 45nm CMOS energy figures for MAC and AC operations underpin the paper's estimated energy advantage.","marker":"[Horowitz, 2014]"}],"fun_headline_variants":["MSVIT fuses multi-scale spike attention for 85.06% on ImageNet","Multi-scale spiking attention pushes spiking transformer to 85.06%","MSVIT: multi-scale fusion in spiking attention yields 85.06%","Spiking transformer record: 85.06% with multi-scale attention fusion","Fusing multi-scale spike attention: 85.06% on ImageNet-1K"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on the assumption that fusing the attention vectors from two different image scales is the reason accuracy improves, rather than the other changes introduced at the same time—the new patch embedding, the hierarchical stages, and where the two attention types are placed.","fun_headline_variants_meta":{"raw":{"variants":["MSVIT fuses multi-scale spike attention for 85.06% on ImageNet","Multi-scale spiking attention pushes spiking transformer to 85.06%","MSVIT: multi-scale fusion in spiking attention yields 85.06%","Spiking transformer record: 85.06% with multi-scale attention fusion","Fusing multi-scale spike attention: 85.06% on ImageNet-1K"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000638,"raw_usage":{"total_tokens":2896,"prompt_tokens":860,"completion_tokens":2036,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":476,"completion_tokens_details":{"reasoning_tokens":1927}},"tokens_in":476,"tokens_out":2036,"duration_ms":14668,"temperature":1.0,"reasoning_tokens":1927,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:18:40.582657+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same MSVIT architecture on ImageNet-1K but replace the MSSA column-sum fusion in stages 1-2 with the standard spiking self-attention, or with a concatenation-based fusion, while keeping SPEMSF, stage depths, and the training recipe fixed; if top-1 accuracy stays around 85%, the multi-scale fusion is not the cause of the reported gain.","supporting_citations":[{"cited_title":"Spikformer: When spiking neural network meets transformer","cited_arxiv_id":null,"evidence_quote":"Spikformer supplies the standard spiking self-attention (SSA) used in stage 3 and serves as the origin of the spike-driven transformer lineage that MSVIT extends."},{"cited_title":"Multiscale vision transformers","cited_arxiv_id":null,"evidence_quote":"Multiscale Vision Transformers motivate the multi-scale feature extraction hypothesis that MSSA is built on."}],"review_version":1}