{"id":"a74ce9d1-5031-4e20-98af-0da4b64f7227","arxiv_id":"2508.20955","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"E-ConvNeXt combines CSPNet, batch-normalized ConvNeXt blocks, a stepped stem, and ESE attention to reach 78.3-81.9% ImageNet top-1 at 0.9-3.1 GFLOPs.","lead":"E-ConvNeXt is a lightweight version of the ConvNeXt vision network that uses cross-stage partial connections and other efficiency tweaks. It reports 78.3% to 81.9% ImageNet top-1 accuracy at 0.9 to 3.1 GFLOPs, challenging heavier baselines.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"FLOPs derivation in §3.1.1 is internally inconsistent, so E-ConvNeXt's reported GFLOPs—and the claimed accuracy-efficiency edge—are not yet established.","rationale":"The reader's CONDITIONAL verdict is the right one; my concern sharpens the condition. The strongest claim depends on E-ConvNeXt's own FLOPs, and Section 3.1.1 contains checkable arithmetic errors: the third term of the CSPConvNeXt block as printed is ~115.6M, not ~29.5M, and the stated totals do not match the claimed reduction. This is concrete and falsifiable, and it precedes the external-baseline comparisons. I therefore only partially agree with the reader's choice of weakest assumption: quoted baseline comparability matters, but the paper's own GFLOPs are even less secure. If the FLOPs recomputation matches the table, the errors are typographical and the main conclusion stands; if not, the central 'state-of-the-art accuracy-efficiency' claim is unsubstantiated. No code or architecture configs are verifiable from the text, so the test requires release of exact variant definitions. I would keep the verdict at CONDITIONAL: the condition should explicitly include a corrected FLOPs derivation or profiler output.","tokens_in":18236,"tokens_out":10688,"duration_ms":101480,"concrete_test":"Recompute whole-network FLOPs for E-ConvNeXt-mini/tiny/small from the released configs (Fig. 8, stage ratios [3,3,9,3], 224×224 input) with a standard counter such as fvcore or ptflops, and compare with Table 8. Independently re-evaluate the §3.1.1 block: 56×56×(49×48 + 48×192 + 192×48) ≈ 65.2M, not 66.5M; the printed '384×56×56×96' term is ≈115.6M. If corrected model FLOPs differ from 0.93/2.04/3.12 G by more than ~10%, the state-of-the-art claim fails; if they match, the §3.1.1 equations are transcription errors and the conclusion can stand.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is an accuracy/FLOPs frontier (Table 8), so E-ConvNeXt's own FLOPs are load-bearing. Section 3.1.1's arithmetic cannot be right. For the stated CSPConvNeXt block (DW 7×7 on 48 channels; MLP 48→192→48), the three terms should be 56×56×49×48 ≈ 7.4M, 48×3136×192 ≈ 28.9M, and 192×3136×48 ≈ 28.9M, totaling ≈65.2M. The paper instead writes the second 1×1 term as '384 × 56 × 56 × 1² × 96 ≈ 29.5M'; that product is ≈115.6M, not 29.5M. The preceding ConvNeXt-block total is also mis-stated (~246M, not 257M), so the claimed '60% reduction' does not follow. Without code or configs, Table 8's 0.93/2.04/3.12 GFLOPs cannot be traced to either a correct profiler or the inconsistent derivation. If the true FLOPs are higher, the points shift right on the accuracy-efficiency curve and the state-of-the-art comparisons fail. This is a more direct threat than the also-valid baseline comparability concern because it affects the paper's own measurements.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes E-ConvNeXt, a lightweight ConvNeXt variant that integrates CSPNet-style cross-stage partial connections, a stepped stem, Batch Normalization in place of Layer Normalization, and an ESE channel-attention module in place of Layer Scale. Three variants (mini/tiny/small) are reported on ImageNet-1K with claimed accuracies of 78.3% at 0.93 GFLOPs, 80.6% at 2.04 GFLOPs, and 81.9% at 3.12 GFLOPs. Transfer experiments on sonar and underwater optical object detection (PP-YOLOE and YOLOv10) are reported. The central claim is a superior accuracy-efficiency frontier relative to comparable lightweight models.","tokens_in":18604,"tokens_out":3553,"duration_ms":27843,"significance":"If the reported accuracy-efficiency numbers are correct and reproducible, the paper offers a useful practical contribution: a simple way to adapt a well-known architecture family to resource-constrained settings, with measured ImageNet-1K results and transfer detection results. The architecture choices are plausible and the experimental protocol is conventional. However, the significance is currently conditional on resolving the FLOPs accounting and on the comparability of Table 8 baselines. The paper ships no code or configs and the GitHub link is not verifiable from the manuscript.","major_comments":[{"comment":"The FLOPs derivation is internally inconsistent. For the stated CSPConvNeXt block (7×7 DW on 48 channels; MLP 48→192→48 at 56×56), the terms should be roughly 56×56×49×48 ≈ 7.4M, 48×3136×192 ≈ 28.9M, and 192×3136×48 ≈ 28.9M, totaling ≈65.2M. The manuscript instead writes the second 1×1 term as '384 × 56 × 56 × 1² × 96 ≈ 29.5M'—a product of ≈115.6M—and the third term as '384 × 56 × 56 × 1² × 96 ≈ 29.5M', which is also inconsistent with the 48/192 channel configuration described in the text and Figure 2. The ConvNeXt-block total (≈257M) is also wrong (the three listed terms sum to ≈247M), so the claimed '60% reduction' does not follow from the shown arithmetic. Since Table 8’s reported GFLOPs (0.93/2.04/3.12) are central to the SOTA claim, the paper needs either a corrected derivation or—preferably—profiler-based FLOPs and fully specified configs to verify the reported numbers.","section":"§3.1.1"},{"comment":"The comparisons quote accuracy and FLOPs for baseline models from their original papers without a unified training or measurement protocol. Differences in training recipe (epochs, augmentation, EMA, input resolution) and FLOPs conventions (including or excluding multiply-add details, detection of efficient operators) can shift the frontier substantially. The claim that E-ConvNeXt establishes a new SOTA accuracy-FLOPs balance is therefore not yet established. At minimum the authors should disclose the exact FLOPs-measurement tool and conventions used for their own models, and for close baselines such as StarNet-S4 (78.4% at 1.07G vs. 78.3% at 0.93G), report whether the comparison is within noise. This is a load-bearing point because the paper’s main conclusion is based on this table.","section":"Table 8 / §4.6"},{"comment":"The object-detection comparisons mix detector scales in a way that undermines the stated conclusions. Table 11 compares YOLOv10-L with PaNet against YOLOv10-M with ConvNeXt-T and E-ConvNeXt-T; the reported FLOPs differences reflect both the backbone and the detector head, so the 'higher mAP with smaller FLOPs' claim is not attributable to the backbone. The same mixing appears in Table 15 (YOLOv10-L PaNet vs. YOLOv10-M E-ConvNeXt-T). To support the backbone-generalization claim, comparisons should hold the detector scale fixed (e.g., YOLOv10-M with PaNet vs. YOLOv10-M with E-ConvNeXt-T) or at least report paired configurations. As written, these tables support only a weaker claim that E-ConvNeXt works as a backbone, not that it outperforms PaNet at equal total FLOPs.","section":"§5.3.1 / Table 11 and §5.4.1 / Table 15"},{"comment":"There is an unresolved discrepancy in the reported FLOPs trajectory. §3.1.3 says the final CSPConvNeXt has 'the same FLOPs' as original CSPConvNeXt (1.5G per Table 3), yet the final E-ConvNeXt models are reported at 0.93G–3.12G with stage ratios and channel widths that do not match the 'input channels [65, 128, 256, 512], output channels [128, 256, 512, 1024]' configuration, since channel counts must be integers and the stem/block changes alter the totals. The paper never gives a complete configuration table (channel widths and block counts per stage) for mini/tiny/small, so the reader cannot reproduce the reported FLOPs or accuracies. This is a load-bearing documentation gap for a paper whose contribution is an efficiency frontier.","section":"§3.1.3 and Table 3 vs. §3.4"}],"minor_comments":[{"comment":"Many typos and inconsistent naming: 'ConNeXt' vs 'ConvNeXt', 'CSPConNeXt', 'Dpethwise', 'proformence', 'Tint', 'PaNet' presumably 'PANet', 'Table 9/10 both say training data' for the test table, and 'E-ConvNeXt-Small' vs 'E-ConvNeXt-S'. Copy-editing is needed.","section":"Throughout"},{"comment":"The claim that Layer Scale is replaced by channel attention to improve accuracy needs a baseline ablation: Table 7 compares attention modules on ImageNet-100, but there is no comparison of ESE Block against the original Layer Scale (or against no scaling) at the same dataset and budget. The causal attribution in the abstract is therefore not directly supported.","section":"§2.3 and §3.3"},{"comment":"Training details are incomplete: the paper says 'for more detailed settings, please refer to the github repository' but the repository link is not shown to be accessible and no code/checkpoints are provided. ImageNet-100 is described as a subset but the class-selection procedure is not given.","section":"§4.1 and Table 2"},{"comment":"The trade-off curve is difficult to read from the text-only manuscript; the figure is not included in the text. If it is included in the camera-ready, please ensure all Table 8 models are plotted with consistent FLOPs conventions.","section":"Figure 9"},{"comment":"The sonar dataset and DUO are niche; the DUO results are labeled 'COCO underwater' in Table 15 but DUO is not a COCO subset. Please correct the label and provide the DUO test/evaluation details (e.g., whether the official metric was used).","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The paper is a plausible engineering contribution, but the FLOPs derivation in §3.1.1 contains arithmetic errors that directly affect the central efficiency claim. I would encourage the editor to require a complete architecture configuration table and either corrected derivations or profiler outputs. The GitHub link should be checked for reproducibility. The detector results also need paired comparisons to be interpretable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper is an empirical architecture study that combines CSPNet with ConvNeXt, adds a transition hyperparameter, a stepped stem, BN in blocks, and ESE attention. The specific configuration is new, though the CSP+ConvNeXt combination already exists in CSPNeXt (RTMDet), which the paper does not cite. The ImageNet accuracy numbers (78.3/80.6/81.9 at 0.93/2.04/3.12 GFLOPs) look plausible, and the ablations are systematic with a standard 300-epoch recipe. That part is fine.\n\nThe problem is the FLOPs derivation in Section 3.1.1. The arithmetic does not hold up: the 7x7 depthwise term uses 58 channels in the text (should be 48), and the third 1x1 term is written as 384x56x56x96 ≈ 29.5M, but that product is about 115.6M (they likely meant 192x56x56x48). The ConvNeXt block total is also about 246M, not 257M. From their own corrected numbers, the reduction is roughly 73%, not the claimed 60%. None of this necessarily means the reported GFLOPs are wrong—those might come from a profiler—but the paper does not say so, and the derivation is presented as evidence. The central claim of a superior accuracy-efficiency frontier rests on those FLOPs numbers, so they need to be verified or corrected.\n\nTwo other soft spots, in proportion. Table 8 quotes baselines from their own papers without a unified training recipe, which weakens the state-of-the-art claim. The object-detection experiment in Table 11 compares YOLOv10-L with PaNet against YOLOv10-M with ConvNeXt-T and E-ConvNeXt-T—different detector heads, so the mAP differences are confounded. The later DUO experiment (Table 15) is fairer. Also, the 36-epoch detection protocol is short for YOLOv10, though it is applied consistently.\n\nWho should read this: people building lightweight backbones for edge deployment, and researchers interested in ConvNeXt variants. It is not a new branch of knowledge, but the specific configuration could be useful. The paper deserves a serious referee, not a desk reject, but only if the authors fix the FLOPs accounting and release code. I would not cite it until those numbers are independently verified. Take the accuracy-FLOPs curve with a grain of salt until then.","headline":"A plausible lightweight ConvNeXt variant, but the FLOPs arithmetic in Section 3.1.1 is wrong, so the central trade-off claim is not yet established.","tokens_in":716,"tokens_out":966,"would_cite":false,"duration_ms":53140,"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":"E-ConvNeXt makes ConvNeXt lightweight by splitting every stage into two streams","keywords":["lightweight networks","ConvNeXt","CSPNet","cross-stage partial connections","ImageNet classification","channel attention","object detection","efficient CNNs"],"falsifier":"Retrain FasterNet-T2, EdgeNeXt-S, MobileViT-S, StarNet-S4, and ConvNeXt-T under the exact E-ConvNeXt training schedule and count FLOPs with one tool at 224x224; if any baseline matches or beats E-ConvNeXt at equal or lower FLOPs in the same band, the paper's accuracy-efficiency claim fails. A second check: measure wall-clock latency for E-ConvNeXt-tiny versus ConvNeXt-tiny on a representative mobile CPU; the paper predicts a 20-25% speedup for the BN block, so a substantially slower result would undercut the efficiency claim.","tokens_in":18164,"feed_emoji":"⚡","tokens_out":4475,"duration_ms":40679,"temperature":0.7,"pith_summary":"This paper tries to make ConvNeXt usable where compute is scarce. It grafts Cross Stage Partial (CSP) connections onto ConvNeXt, splitting each stage's feature map so only part flows through the expensive blocks, and couples that with a stepped stem, batch normalization in place of layer normalization, and a lightweight channel-attention (ESE) module replacing Layer Scale. The result, E-ConvNeXt, claims network complexity reductions of up to 80% versus ConvNeXt-tiny at comparable or better accuracy for its size. The authors report ImageNet-1K top-1 accuracy of 78.3% at 0.93 GFLOPs (mini), 80.6% at 2.04 GFLOPs (tiny), and 81.9% at 3.12 GFLOPs (small), and show the backbone transfers to object detection on sonar and underwater optical images. A sympathetic reader should care because the paper is a direct recipe for turning a strong general-purpose backbone into lightweight variants without abandoning its design.","feed_headline":"CSP split shrinks ConvNeXt compute by up to 80%","feed_subtitle":"Mini hits 78.3% top-1 at 0.93 GFLOPs; tiny 80.6% at 2.04 GFLOPs; small 81.9% at 3.12 GFLOPs.","key_machinery":"Cross-stage partial (CSP) routing is the load-bearing mechanism. Each CSPConvStage splits the incoming feature map into two streams; one stream goes through N ConvNeXt-style blocks (7x7 depthwise conv plus two 1x1 convs, inverted-residual structure, with BN and GELU), the other stream bypasses the blocks; a final 1x1 conv merges the streams. The transition hyperparameter ch_mid = (ch_in + ch_out)/2 is what makes the expensive stream narrower, and the bypass stream is what preserves information flow at reduced cost. The stepped stem, the BN-for-LN substitutions, and the ESE channel-attention module are supporting modifications that recover accuracy and speed lost when the network is thinned.","core_discovery":"The paper's central claim is that ConvNeXt's accuracy-efficiency deficit in small models is structural, not a scaling problem, and that integrating CSPNet's cross-stage partial routing into ConvNeXt repairs it. In E-ConvNeXt, each stage's feature map is split: one part passes through 7x7 depthwise conv plus two 1x1 convs, the other bypasses the blocks, and a 1x1 conv merges the streams; a transition hyperparameter, ch_mid = (ch_in + ch_out)/2, sets the width of the expensive stream and concentrates the FLOP reduction. Three supporting changes recover accuracy and speed: a stepped stem (2x2 conv followed by two 3x3 convs) instead of a single 4x4 patchify stem; batch normalization replacing la","pith_inferences":["The CSP split with ch_mid creates an asymmetric information path that may behave like a learned residual gating: one testable hypothesis is that the bypass stream carries global or low-frequency context while the block stream refines details, which could be checked with feature visualizations.","Because the paper reports FLOPs savings but only a small FPS comparison, a practical deployment test on mobile CPUs or GPUs with measured latency would be a stronger claim than FLOPs alone; BN and 1x1-conv blocks may or may not translate to real speedups on different hardware.","The same recipe (CSP split + stepped stem + BN + ESE attention) could plausibly be applied to other ConvNeXt-derived or hybrid transformer backbones; the paper does not test this, but the mechanism is architecture-agnostic.","The 0.7% accuracy gain attributed to stage-ratio adjustment suggests that how FLOPs are distributed across stages matters as much as total FLOPs; a systematic depth-width allocation study could extend this result beyond the three variants presented."],"forward_implications":["E-ConvNeXt variants give a ready-made small-model family for resource-constrained deployments: 7.6M parameters at 0.93 GFLOPs, 13.2M at 2.04 GFLOPs, and 19.4M at 3.12 GFLOPs, all trained with the same 300-epoch ImageNet-1K recipe.","As a backbone for PP-YOLOE and YOLOv10 detectors, E-ConvNeXt-tiny raises mAP on sonar images and underwater optical images relative to the default PaNet backbone, with lower or comparable FLOPs than the larger detector configuration.","If the ImageNet-1K results hold under matched training conditions, E-ConvNeXt becomes a strong pure-convolution lightweight option in the 1-3 GFLOP band, making the CSP-split-plus-stepped-stem recipe a template for other heavy backbones.","Replacing Layer Scale with a normalized squeeze-and-excitation block suggests that channel attention can substitute for per-channel learned scaling in ConvNeXt-style blocks without destabilizing training, opening that substitution to other architectures.","The reported BN-for-LN swap is claimed to add no FLOPs or parameters while improving speed by about 20-25%, which would make it a nearly free efficiency upgrade for ConvNeXt-derived blocks."],"supporting_citations":[{"why":"Supplies the base ConvNeXt architecture that E-ConvNeXt modifies and the ConvNeXt-tiny baseline for comparison.","marker":"[7]"},{"why":"Supplies the cross-stage partial connection mechanism and the claim that CSP reduces computation while preserving accuracy.","marker":"[16]"},{"why":"Provides FasterNet, a key lightweight baseline that E-ConvNeXt must beat in the accuracy-FLOPs comparison.","marker":"[15]"},{"why":"Provides MobileViT-XS and MobileViT-S baselines used in the state-of-the-art comparison table.","marker":"[39]"},{"why":"Provides EdgeNeXt-XS and EdgeNeXt-S baselines used in the state-of-the-art comparison table.","marker":"[40]"},{"why":"Provides StarNet-S4, the strongest comparable baseline in the mini group that E-ConvNeXt-mini claims to match.","marker":"[41]"},{"why":"Supplies the squeeze-and-excitation channel attention paradigm and the SENet integration style adopted for the ESE block.","marker":"[22]"},{"why":"Supplies the ESE block itself, the channel-attention module chosen over SE, ECA, and CBAM in the paper's comparison.","marker":"[29]"},{"why":"Provides MobileNetV2, a widely used lightweight baseline referenced both for context and in the comparison table.","marker":"[13]"},{"why":"Identifies the layer-scaling technique in ConvNeXt that E-ConvNeXt replaces with ESE channel attention.","marker":"[30]"}],"fun_headline_variants":["E-ConvNeXt: CSPNet cuts ConvNeXt FLOPs 80%","Cross-stage split makes ConvNeXt slim and fast","ConvNeXt slims down with CSPNet: 80% less FLOPs","Split-stage trick trims ConvNeXt to 80% fewer FLOPs","Mini hits 78.3% top-1 at 0.9 GFLOPs with CSP"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The comparison table assumes that accuracy and FLOPs quoted from other papers are directly comparable to E-ConvNeXt's numbers without re-running those baseline models under a unified training recipe and a single FLOPs convention.","fun_headline_variants_meta":{"raw":{"variants":["E-ConvNeXt: CSPNet cuts ConvNeXt FLOPs 80%","Cross-stage split makes ConvNeXt slim and fast","ConvNeXt slims down with CSPNet: 80% less FLOPs","Split-stage trick trims ConvNeXt to 80% fewer FLOPs","Mini hits 78.3% top-1 at 0.9 GFLOPs with CSP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001202,"raw_usage":{"total_tokens":4829,"prompt_tokens":823,"completion_tokens":4006,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":3898}},"tokens_in":567,"tokens_out":4006,"duration_ms":28989,"temperature":1.0,"reasoning_tokens":3898,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T14:40:28.255794+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain FasterNet-T2, EdgeNeXt-S, MobileViT-S, StarNet-S4, and ConvNeXt-T under the exact E-ConvNeXt training schedule and count FLOPs with one tool at 224x224; if any baseline matches or beats E-ConvNeXt at equal or lower FLOPs in the same band, the paper's accuracy-efficiency claim fails. A second check: measure wall-clock latency for E-ConvNeXt-tiny versus ConvNeXt-tiny on a representative mobile CPU; the paper predicts a 20-25% speedup for the BN block, so a substantially slower result would undercut the efficiency claim.","supporting_citations":[{"cited_title":"Run, don’t walk: chasing higher flops for faster neural networks","cited_arxiv_id":null,"evidence_quote":"Provides FasterNet, a key lightweight baseline that E-ConvNeXt must beat in the accuracy-FLOPs comparison."},{"cited_title":"Edgenext: efficiently amalgamated cnn-transformer archi- tecture for mobile vision applications","cited_arxiv_id":null,"evidence_quote":"Provides EdgeNeXt-XS and EdgeNeXt-S baselines used in the state-of-the-art comparison table."},{"cited_title":"Rewrite the stars","cited_arxiv_id":null,"evidence_quote":"Provides StarNet-S4, the strongest comparable baseline in the mini group that E-ConvNeXt-mini claims to match."},{"cited_title":"Squeeze-and-excitationnetworks","cited_arxiv_id":null,"evidence_quote":"Supplies the squeeze-and-excitation channel attention paradigm and the SENet integration style adopted for the ESE block."},{"cited_title":"Centermask: Real-time anchor- free instance segmentation","cited_arxiv_id":null,"evidence_quote":"Supplies the ESE block itself, the channel-attention module chosen over SE, ECA, and CBAM in the paper's comparison."},{"cited_title":"Mobilenetv2: Inverted residuals and linear bottlenecks","cited_arxiv_id":null,"evidence_quote":"Provides MobileNetV2, a widely used lightweight baseline referenced both for context and in the comparison table."},{"cited_title":"Original approach for the localisation of objects in images.IEE Proceedings- Vision, Image and Signal Processing, 141(4):245–250, 1994","cited_arxiv_id":null,"evidence_quote":"Identifies the layer-scaling technique in ConvNeXt that E-ConvNeXt replaces with ESE channel attention."}],"review_version":1}