{"id":"c4c49fa5-5ede-4831-bf6d-52836d1ce2e6","arxiv_id":"2411.09604","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"Local-Global Attention, a module fusing multi-scale local and global attention branches with learned weights, reports marginal accuracy gains (0.1 to 0.9 mAP points) over standard attention mechanisms on several detection benchmarks, but without error bars or code.","lead":"This preprint proposes Local-Global Attention, an attention module that combines small-kernel local convolutions with a global convolution branch and learned fusion weights, and tests it on object detection and classification datasets. The reported accuracy gains are small and the experiments lack error bars, code, and comparisons against the local-global attention methods the paper itself cites.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Experimental evidence for LGA's 'consistent outperformance' is unreliable: an SE baseline beats LGA on DOTAv1.0, MHSA baselines collapse on two datasets, and GFLOPs are identical across very different modules; the central claim is not established.","rationale":"The paper's central claim is empirical, so I looked for the weakest point in the empirical case. The most load-bearing concern is not primarily the underspecified global kernel, though that is a real reproducibility problem; it is that the reported numbers themselves are internally unreliable. Table 7 contains a direct counterexample to 'consistently outperforms' (SE beats LGA on DOTAv1.0 mAP50), and Table 6 contains baseline implausibilities (MHSA dropping 6.3 and 2.9 points on VOC2012 and COCOminitrain respectively). The identical GFLOPs across methods that cannot plausibly have identical computational cost further signals that the experimental logging is not trustworthy. Since the claimed LGA improvements are tiny, any one of these artifacts is enough to undermine the headline. I therefore agree with the reader's REJECT verdict and recommend no adjustment; a multi-seed re-run with correct baseline implementations and per-module cost profiling would settle whether LGA actually outperforms alternatives, but until that is done the central claim is not established.","tokens_in":11466,"tokens_out":7159,"duration_ms":71160,"concrete_test":"Run the Table 6 VOC2012 experiment using the paper's released code, with 5 seeds and a reference MHSA implementation, and log per-module GFLOPs. If the reference MHSA no longer collapses, if LGA's GFLOPs are not identical to baseline, or if LGA's margin over SE/CBAM falls within the seed standard deviation, the reported comparison does not support consistent outperformance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that LGA 'consistently outperforms' existing attention mechanisms is contradicted by the paper's own Table 7: on DOTAv1.0, SE reaches mAP50 54.4 while LGA reaches 54.3, so the consistency claim fails even before any re-analysis. More broadly, the comparisons show signs of unreliable baselines and measurements. In Table 6, adding MHSA to MobileNetV3 drops VOC2012 mAP50 from 43.7 to 37.4 and COCOminitrain mAP50 from 14.6 to 11.7, far outside plausible attention-module effects. The GFLOPs columns are identical across all methods in nearly every table (e.g., 2.8 in Tables 6 and 7) despite LGA adding multiple convolutions and attention branches. No seeds or error bars are reported, the promised code link is absent, and the global kernel size and positional encoding formula are unspecified. With reported LGA gains of only 0.1–0.9 mAP, these artifacts are as plausible an explanation as the mechanism itself. The evidence base for the headline claim is therefore not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Local-Global Attention (LGA), a plug-in attention module that combines multi-scale convolutions (local branch) with a larger-kernel global branch plus positional encoding, and fuses the two branches via learnable alpha parameters. The authors evaluate LGA by inserting it into MobileNetV3, ResNet18, and YOLOv8 backbones for classification (MNIST, Fashion-MNIST), object detection (VOC2012, VisDrone2019, COCOminitrain, DOTAv1.0, GWHD2020), and small-object detection (TinyPerson), comparing against MHSA, SE, CBAM, ECA, and SA. The central claim is that LGA 'consistently outperforms' existing attention mechanisms while maintaining computational efficiency (Abstract and Section 5). The mechanism itself is described with simple equations (Section 3), but several implementation details are unspecified, and the experimental evidence contains internal inconsistencies and lacks statistical support.","tokens_in":11676,"tokens_out":3085,"duration_ms":30376,"significance":"If the claims were substantiated, LGA would be a useful lightweight attention module for detection and classification, with the alpha-fusion idea being a sensible way to balance local and global features. However, the paper's central claim of consistent outperformance is directly contradicted by one of its own tables (SE beats LGA on DOTAv1.0 mAP50), and the reported baselines show signs of serious implementation problems (MHSA degrades VOC2012 by 6.3 mAP, an implausibly large drop). The computational-efficiency claim is unsupported because GFLOPs are identical across very different modules, and the global branch is incompletely specified. The study also reports no error bars, no multiple seeds, and no significance tests, while all improvements are 0.1–0.9 mAP, within typical run-to-run noise. These issues collectively mean the paper's empirical foundation does not establish its headline conclusion. No code is provided despite the statement that it is. The mechanism's design is reasonable, but the validation as presented is not reliable.","major_comments":[{"comment":"The claim that LGA 'consistently outperforms' existing attention mechanisms is contradicted by the paper's own results: on DOTAv1.0, SE achieves mAP50=54.4 while LGA achieves 54.3 (Table 7). Since the abstract and Section 5 specifically state consistency, this single counterexample invalidates the headline claim as currently worded.","section":"Table 7 (and Abstract/Section 5)"},{"comment":"The MHSA baseline results are implausible and suggest an implementation error. Adding MHSA to MobileNetV3 drops VOC2012 mAP50 from 43.7 to 37.4 (a 6.3-point decrease) and COCOminitrain mAP50 from 14.6 to 11.7 (a 2.9-point decrease). No mainstream attention module, especially a simple MHSA add-on, is known to cause such catastrophic degradation. If the MHSA baseline is incorrectly implemented, then every comparison against MHSA (and by extension the 'outperforms' claim) is unreliable.","section":"Section 4.5.1, Table 6"},{"comment":"The GFLOPs values are reported as identical across all methods in nearly every table (e.g., 2.8 for MobileNetV3, MHSA, SE, CBAM, and LGA in Tables 6 and 7), despite LGA adding multiple multi-scale convolution branches, two attention computations, and a 1x1 convolution as described in Section 3. Either the complexity figures were not actually measured (e.g., copied from a single run) or the module's cost is misreported. In either case, the statement that LGA 'maintains computational efficiency' is not supported by the data presented.","section":"Tables 6 and 7, with Section 3"},{"comment":"The global attention branch is not sufficiently specified: the 'larger convolution kernel' size kg is never given, and the positional encoding PE is defined only as 'aligns with the spatial dimensions' without providing the encoding formula or how it is combined with the features. Without these details, the method cannot be reproduced, and the claim that the global branch actually captures global context cannot be verified.","section":"Sections 3.4 and 3.6.2"},{"comment":"No error bars, multiple random seeds, or statistical significance tests are reported anywhere in the experimental section. The reported improvements range from 0.1 to 0.9 mAP points, which is well within the run-to-run variance typically observed in object detection. The paper's claim of consistent improvement across datasets is therefore not statistically substantiated.","section":"Section 4 (all experimental tables)"}],"minor_comments":[{"comment":"There is a typo in the heading: 'queeze-and-Excitation Attention' should be 'Squeeze-and-Excitation Attention'.","section":"Section 2.3"},{"comment":"The phrase 'we introduce a learnable parameters' (Abstract) and 'we introduce learnable α parameters' (Section 1) contain grammar errors; the intended meaning is 'learnable parameters'.","section":"Abstract and Section 1"},{"comment":"The tensor notation X ∈ R^{B×D×H×W} is unconventional: D is used for 'embedding dimension' but the standard channel axis is usually C. Also, the text says 'D: Embedding dimension' but then uses 'depth' interchangeably, which is confusing. Please clarify the notation.","section":"Section 3.1"},{"comment":"The paper states 'Code is available at the link' but provides no actual URL. Since the method is claimed to be reproducible, a working link or repository identifier should be given.","section":"Abstract and Section 1"}],"recommendation":"reject","confidential_remarks":"The central claim is not just under-supported; it is internally contradicted (Table 7) and the baseline implementation appears unreliable (Table 6 MHSA collapse). These are load-bearing issues that cannot be fixed by minor edits; the experiments would need to be redone with proper baselines, seeds, and significance testing. The scope of such a revision goes beyond what a major revision can reasonably require, so I recommend rejection. The idea of adaptive local-global fusion is not novel enough to offset the current validation problems."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the idea is a plausible recombination of known pieces, the equations are internally consistent, but the experimental case is not there. The paper claims LGA \"consistently outperforms\" existing attention mechanisms; its own Table 7 shows SE at 54.4 DOTA mAP50 against LGA's 54.3, and there are no error bars, no seeds, and no code. I'd treat the headline claim as unestablished.\n\nWhat is actually new: as far as I can tell, no one has published this exact module: multi-scale depthwise convs (kernels 3, 5, 7) plus a global larger-kernel conv with positional encoding, fused by learned alpha. That is a reasonable engineering contribution. Section 3 is written cleanly, the equations are consistent, and the paper attempts a wide set of datasets, which is more effort than many.\n\nSoft spots, in order of importance. First, the evidence base. Reported gains are 0.1 to 0.9 mAP points; without seeds or error bars, that is noise territory. Second, some baseline numbers look wrong: MHSA drops VOC2012 mAP50 from 43.7 to 37.4 and COCOminitrain from 14.6 to 11.7. That is far outside what an attention module should do, and it makes the comparisons untrustworthy. Third, the GFLOPs columns are identical across very different modules (2.8 everywhere), which makes me suspect the numbers were not actually measured. Fourth, the global branch is underspecified: kernel kg and the positional encoding formula are never given, so the module is not reproducible from the text. Fifth, the abstract says \"Code is available at the link\" but no link appears. Minor: the writing has typos (\"queeze\", \"num scalesX\").\n\nThe stress-test concern holds up; I checked the tables and it is accurate. The paper is not circular—alpha is trained normally, and the evaluation is on external benchmarks. The core problem is purely that the experiments are too weak and too inconsistent to support the claims.\n\nWho this is for: a reader tracking practical attention modules might skim Section 3, but anyone needing a reliable comparison should wait for a version with seeds, error bars, working code, and repaired baselines. My recommendation: desk reject, with the door open for a serious resubmission.","headline":"Plausible module, unproven claims: the experimental section fails to establish consistent gains, with contradicted baselines and missing details.","tokens_in":12228,"tokens_out":1808,"would_cite":false,"duration_ms":26201,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Local-Global Attention fuses fine-grained local features from three convolution scales with a larger-kernel global branch plus positional encoding, using learned alpha weights, and the paper reports consistent mAP gains over standard…","keywords":["object detection","attention mechanism","local-global attention","multi-scale feature integration","positional encoding","small object detection","YOLOv8","MobileNetV3"],"falsifier":"Inspect the released code for the global kernel size and positional encoding; then run the TinyPerson ablation with the global branch removed and with alpha fixed at 0.5 instead of learned, and if mAP does not drop or change, the global branch and the adaptive fusion are not the source of the reported gains.","tokens_in":11205,"feed_emoji":"🎯","tokens_out":7961,"duration_ms":65880,"temperature":0.7,"pith_summary":"This paper introduces Local-Global Attention, a plug-in module that runs a local attention branch and a global attention branch in parallel and then fuses them with learnable scalar weights. The local branch uses multi-scale convolutions with kernel sizes 3, 5, and 7 plus residual connections; the global branch uses a larger convolution kernel with positional encoding. The learned $\\alpha_\\text{local}$ and $\\alpha_\\text{global}$ weights let the network shift the balance between detail and context for each input. The author reports that, across nine object detection and classification benchmarks and three backbones, the module beats existing attention mechanisms such as SE, CBAM, ECA, MHSA, and SA while keeping FLOPs and parameters at roughly baseline levels. A sympathetic reader should care because the claim is that the long-standing local/global trade-off in attention can be handled cheaply and adaptively, with the clearest gains on small-object and multi-class datasets.","feed_headline":"Attention module lifts detection across scales with no extra FLOPs","feed_subtitle":"Multi-scale convolutions, positional encoding, and learnable alpha weights nudge mAP up on nine benchmarks.","key_machinery":"The load-bearing object is the two-branch attention head with a learned fusion. The local branch applies depthwise and $1\\times1$ convolutions at kernel sizes $k \\in \\{3,5,7\\}$ to a position-encoded, scale-weighted feature map, computes attention scores $Q_k K_k^\\top / \\sqrt{d}$, averages the three branch outputs, and adds a residual. The global branch applies the same attention formula with one larger kernel $k_g$ to the same input. The outputs are combined as $\\text{out} = \\alpha_\\text{local}\\,\\text{local\\_out} + \\alpha_\\text{global}\\,\\text{global\\_out}$, where the alphas are learned, and a final $1\\times1$ convolution compresses the fusion. This mechanism is what carries the argument: the residuals preserve identity features, the multiple kernel sizes supply granularity, the positional encoding supplies spatial structure, and the alphas make the local/global balance data-dependent without adding a learned projection or extra FLOPs.","core_discovery":"The paper's central claim is that combining multi-scale local attention with a positional-encoded global branch, fused by learnable alpha parameters, gives a better balance of fine detail and broad context than any single attention mechanism. On the paper's own terms this is established empirically: LGA posts the best or tied-best number on every benchmark, with the largest margins on small-object datasets such as TinyPerson (+0.92 mAP@50 on MobileNetV3, +0.7 on YOLOv8) and strong results on VOC2012 (+0.7 mAP@50-95), VisDrone2019, DOTAv1.0, and COCOminitrain, while GFLOPs stay equal to the baseline in nearly every row. The adaptive alpha fusion is the specific reason offered for these gains: the model can emphasize local detail when the input needs it and global context when it needs that.","pith_inferences":["A natural follow-up is repeated-seed evaluation, since the reported margins are 0.1 to 0.9 mAP and the paper does not report variance; such runs would show whether the advantage over CBAM and SE is stable.","Visualizing the learned alpha values per image would test the adaptivity claim directly: small-object-heavy inputs should weight the local branch more, and scene-heavy inputs should weight the global branch more.","The same two-branch fusion pattern could transfer to semantic segmentation and video object detection, where scale variance and the local/global context trade-off are equally central."],"forward_implications":["Plugging LGA into MobileNetV3, ResNet18, or YOLOv8 improves detection mAP on small-object benchmarks such as TinyPerson, VisDrone2019, and DOTAv1.0.","The module adds essentially no FLOPs and only a few hundred thousand parameters, so it can replace SE, CBAM, or other attention blocks in existing backbones without a compute budget change.","The same mechanism improves Top-1 accuracy on MNIST and Fashion-MNIST classification, indicating the local/global fusion is not specific to detection.","Because the alpha weights are learned, the module's behavior can adapt per dataset or per input, which is the mechanism behind the claimed multi-scale robustness."],"supporting_citations":[{"why":"YOLOv8 supplies the detection framework and default training settings used in every experiment.","marker":"[14]"},{"why":"MobileNetV3 is the primary backbone into which LGA and the baseline attention modules are inserted.","marker":"[10]"},{"why":"ResNet18 is the second backbone in the TinyPerson ablation, testing whether LGA generalizes beyond MobileNetV3.","marker":"[9]"},{"why":"TinyPerson is the small-object benchmark where LGA reports its largest mAP gains.","marker":"[34]"},{"why":"Squeeze-and-Excitation is one of the main attention baselines LGA must outperform.","marker":"[12]"},{"why":"CBAM is another baseline attention module, combining channel and spatial attention, used in all comparison tables.","marker":"[28]"},{"why":"Multi-head self-attention is the global-attention baseline in the comparisons.","marker":"[25]"},{"why":"COCO2017 defines the COCOminitrain benchmark and the Figure 1 comparison context.","marker":"[16]"},{"why":"MNIST is one of the two classification benchmarks used to evaluate LGA in MobileNetV3.","marker":"[4]"},{"why":"Fashion-MNIST is the harder classification benchmark where LGA is the only attention module with a positive accuracy gain.","marker":"[31]"}],"fun_headline_variants":["Adaptive local-global attention boosts multi-scale detection with zero extra FLOPs","Local-global attention: adaptive balance for multi-scale detection, no extra FLOPs","Self-balancing local-global attention lifts small-object detection at no FLOP cost","Multi-scale attention with adaptive local-global mixing, zero extra FLOPs","Adaptive local-global attention lifts small-object detection without extra FLOPs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes the larger-kernel branch actually captures distinct global context, but it never specifies the global kernel size or the positional encoding formula, so the two branches could be encoding overlapping information and the learned alpha weights could be adjusting noise rather than a real local/global trade-off.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive local-global attention boosts multi-scale detection with zero extra FLOPs","Local-global attention: adaptive balance for multi-scale detection, no extra FLOPs","Self-balancing local-global attention lifts small-object detection at no FLOP cost","Multi-scale attention with adaptive local-global mixing, zero extra FLOPs","Adaptive local-global attention lifts small-object detection without extra FLOPs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001033,"raw_usage":{"total_tokens":4349,"prompt_tokens":940,"completion_tokens":3409,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":3304}},"tokens_in":556,"tokens_out":3409,"duration_ms":20799,"temperature":1.0,"reasoning_tokens":3304,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:28:05.746277+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the released code for the global kernel size and positional encoding; then run the TinyPerson ablation with the global branch removed and with alpha fixed at 0.5 instead of learned, and if mAP does not drop or change, the global branch and the adaptive fusion are not the source of the reported gains.","supporting_citations":[{"cited_title":"Ultralytics yolov8, 2023","cited_arxiv_id":null,"evidence_quote":"YOLOv8 supplies the detection framework and default training settings used in every experiment."},{"cited_title":"Deep residual learning for image recognition","cited_arxiv_id":null,"evidence_quote":"ResNet18 is the second backbone in the TinyPerson ablation, testing whether LGA generalizes beyond MobileNetV3."},{"cited_title":"Scale match for tiny person detection","cited_arxiv_id":null,"evidence_quote":"TinyPerson is the small-object benchmark where LGA reports its largest mAP gains."},{"cited_title":"Squeeze-and-excitation net- works","cited_arxiv_id":null,"evidence_quote":"Squeeze-and-Excitation is one of the main attention baselines LGA must outperform."},{"cited_title":"Cbam: Convolutional block attention module","cited_arxiv_id":null,"evidence_quote":"CBAM is another baseline attention module, combining channel and spatial attention, used in all comparison tables."},{"cited_title":"Lawrence Zitnick, and Piotr Doll ´ar","cited_arxiv_id":null,"evidence_quote":"COCO2017 defines the COCOminitrain benchmark and the Figure 1 comparison context."},{"cited_title":"The mnist database of handwritten digit images for machine learning research","cited_arxiv_id":null,"evidence_quote":"MNIST is one of the two classification benchmarks used to evaluate LGA in MobileNetV3."},{"cited_title":"Fashion- mnist: a novel image dataset for benchmarking machine learning algorithms, 2017","cited_arxiv_id":null,"evidence_quote":"Fashion-MNIST is the harder classification benchmark where LGA is the only attention module with a positive accuracy gain."}],"review_version":1}