{"id":"aaf6b104-bfe1-40a8-a595-973e49bfa8e3","arxiv_id":"2502.04656","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"MHAF-YOLO reaches 48.9% AP on COCO with 7.1M parameters by combining auxiliary shallow-deep fusion paths and heterogeneous reparameterized convolutions.","lead":"MHAF-YOLO is a modified YOLO detector with a new neck that adds extra connections between shallow and deep feature layers, plus multi-scale reparameterized convolutions. On COCO it reports 48.9% average precision with 7.1 million parameters, beating several recent YOLO variants at comparable size.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Baseline recipe and val-set tuning are uncontrolled; the 1.9 AP gain over YOLO11-s could reflect training-protocol differences rather than architecture.","rationale":"The paper has genuine strengths: the module-level ablations in Tables 2, 3, and 5 are controlled and show consistent gains, the RepHConv reparameterization derivation is standard, and the proposed modules are described in enough detail to be testable. My concern is not with the architecture's internal consistency but with the measurement framework used for the headline comparison. The reader's weakest assumption identified the same spot: published baselines combined with val2017 serving as both selection and evaluation set. I agree. The most direct way to settle it is to run the key baselines under the authors' own recipe and evaluate all models on a held-out split. If the advantage survives, the CONDITIONAL verdict can be upgraded; if not, the central SOTA claim is overstated. I therefore leave the reader's verdict unchanged.","tokens_in":20366,"tokens_out":9436,"duration_ms":89276,"concrete_test":"Using the released MHAF code, retrain the three main baselines in Table 6 (YOLOv10-s, YOLO11-s, YOLOv12-s) under the exact protocol of Section 4.2, then evaluate MHAF-YOLO-s and all retrained baselines on COCO test-dev or on a held-out subset of val2017 that was never used for ablations. If MHAF-YOLO-s remains at least 1 AP ahead of both baselines under matched training and held-out evaluation, the central claim is supported; if the gap drops below roughly 1 AP or reverses, the SOTA claim is an artifact of uncontrolled baselines and val-set tuning.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the 1.9 AP gain of MHAF-YOLO-s (48.9) over YOLO11-s (47.0) and the 0.9 gain over YOLOv12-s in Table 6. These gaps are measured against published baseline numbers, not models run under the paper's protocol. Section 4.2 specifies MHAF is trained from scratch for 500 epochs with SGD, cached-mixup, and low-probability copy-paste, with the last 10 epochs disabling strong augmentation; published YOLO11/YOLOv12 numbers come from their own schedules and augmentation stacks. Sections 4.3-4.5 then use COCO val2017 to select SAF/AAF/RepHMS/GHFKS configurations, and Table 6 reports the final result on the same val2017 set, so the comparison also mixes model selection and evaluation. The nano-scale numbers reinforce the fragility: MHAF-YOLO-n is 42.3 while the authors' own MAF-YOLO-n is 42.4, so claimed SOTA status is not even consistent at the nano scale. If the 1.9 AP gap is mostly training-recipe and val-selection effects, the architecture's contribution is much smaller than claimed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MHAF-YOLO, a YOLO-family detector built on three main architectural ingredients: a Multi-Branch Auxiliary FPN (MAFPN) neck with Superficial Assisted Fusion (SAF) and Advanced Assisted Fusion (AAF), a Reparameterized Heterogeneous Multi-Scale (RepHMS) feature-extraction block using reparameterized heterogeneous depthwise convolutions, and a Global Heterogeneous Flexible Kernel Selection (GHFKS) mechanism that assigns larger kernels to different feature scales. The authors report ablations on COCO val2017, comparisons against published real-time detector results on COCO and VOC, and transfer experiments to instance segmentation and rotated object detection. The central quantitative claim is that MHAF-YOLO-s reaches 48.9 AP at 7.1M parameters and 25.3 GFLOPs, surpassing YOLO11-s (47.0 AP, 9.4M) and YOLOv12-s (48.0 AP, 9.3M), and the paper concludes that MHAF-YOLO achieves state-of-the-art performance in real-time object detection.","tokens_in":20631,"tokens_out":9850,"duration_ms":85809,"significance":"If the empirical claims were established under a controlled protocol, this would be a practically valuable architecture paper. The internal ablations in Tables 2, 3, and 5 are systematic and show that each proposed component contributes positively under the authors' training recipe; the reparameterization identity in Eq. (6) is algebraic, and the code is publicly released. The transfer demonstrations to segmentation and rotated detection support the generality of the modules. However, the central SOTA claim currently rests on published baseline numbers from heterogeneous training protocols and on model selection and evaluation using the same COCO val2017 split, so the magnitude of the architectural gain over YOLO11/YOLOv12 is not yet established. The paper also never reports inference latency while calling itself real-time.","major_comments":[{"comment":"The main SOTA claim is measured against published baseline numbers rather than models retrained under the paper's protocol. MHAF-YOLO is trained from scratch for 500 epochs with SGD, cached-mixup, and low-probability copy-paste, while YOLOv10, YOLO11, YOLOv12, Gold-YOLO, and MAF-YOLO results in Table 6 come from their own schedules and augmentation stacks; the 1.9 AP gap over YOLO11-s and 0.9 AP gap over YOLOv12-s therefore conflate architecture with training recipe. Please retrain at least YOLOv10-s, YOLO11-s, YOLOv12-s, Gold-YOLO-s, and MAF-YOLO-s under the identical 500-epoch recipe, or report official and same-recipe numbers side by side, and give full training hyperparameters such as batch size, input resolution, learning-rate schedule, and augmentation probabilities.","section":"§4.2, Table 6"},{"comment":"The experimental protocol selects architecture hyperparameters on COCO val2017 in Sections 4.3–4.5 (RepHMS depths, GHFKS kernel sizes, SAF/AAF configurations) and then evaluates the final models on the same val2017 split in Section 4.6, Table 6. This selection-on-evaluation procedure can inflate both the ablations and the SOTA comparison, and the paper reports only APval with no COCO test-dev numbers. Please report test-dev AP, or make an explicit train/val separation for model selection and final evaluation, and state which reported numbers were used for architecture decisions.","section":"§4.3–§4.6, Table 6"},{"comment":"The paper repeatedly frames MHAF-YOLO as a real-time detector, but no latency or FPS measurement appears anywhere, and the conclusion explicitly concedes that inference speed still lags behind cutting-edge models such as YOLOv10 and YOLO11 because of MAFPN complexity and large-kernel depthwise convolutions. The real-time claim is therefore unsupported by the evidence; please add latency measurements on identical hardware, batch size, and input resolution for all compared models and revise the SOTA claims to separate accuracy/parameter trade-off from speed.","section":"§5 (Conclusions), §1, Fig. 2"},{"comment":"The nano-scale result is internally inconsistent with the claim that MHAF-YOLO surpasses existing real-time detectors: MHAF-YOLO-n reaches 42.3 AP, while the authors' own MAF-YOLO-n in the same table reaches 42.4 AP despite having more parameters. If MHAF-YOLO is meant to improve on MAF-YOLO, the table should be explained in terms of training recipe or capacity target; otherwise the claim of uniform superiority across N/S/M scales should be qualified.","section":"§4.6.1, Table 6"},{"comment":"The plug-and-play claim for MAFPN is partly confounded in the YOLOv8n experiment: YOLOv8n-MAFPN is trained for 300 epochs while YOLOv8n-PAFPN is trained for 500 epochs, and its channel count is changed to make the model smaller, so the +1.7 AP gain cannot be attributed solely to the neck. Please provide an equal-epoch, equal-capacity comparison for YOLOv8n and specify the training schedule, resolution, and augmentation used for the Faster R-CNN comparison.","section":"§4.4.2, Table 4"},{"comment":"Section 4.2 states that all scales of MHAF-YOLO are trained from scratch without relying on large-scale datasets such as ImageNet or pre-trained weights, but Table 6 includes rows MHAF-YOLO-n* and MHAF-YOLO-s* whose footnote says they are trained with a pretrained backbone. The text never explains how these starred models are pretrained or why they are included, and the table presents their higher numbers alongside the from-scratch results. Please remove the starred rows or describe the pretraining protocol explicitly, and keep the from-scratch claim consistent with the tables.","section":"§4.2, Table 6"}],"minor_comments":[{"comment":"YOLO11-n and YOLO11-s are cited as reference [16] in Table 6, but reference [16] is the authors' MAF-YOLO paper; YOLO11 should be cited as [18] as in Table 7, otherwise the provenance of the baseline numbers is unclear.","section":"Table 6, References"},{"comment":"The implementation details list the optimizer, hardware, and epoch count but omit batch size, input resolution, learning rate, momentum/weight decay, and the actual cached-mixup/copy-paste probabilities; these details are needed to reproduce the 500-epoch runs and to verify the claimed parameter-accuracy trade-offs.","section":"§4.2"},{"comment":"The notation in Eq. (6) is ambiguous: the kernel indices K_{2n-1} and K_{2n-(2i+1)} are not defined precisely, and the step of zero-padding smaller depthwise kernels before summation should be stated explicitly, since the correctness of the reparameterization merge relies on that padding.","section":"Eq. (6)"},{"comment":"The phrase saying that the Cascade strategy achieves 'lossless performance improvement' is inaccurate if 'lossless' refers to accuracy, because adding Cascade changes AP from 41.9 to 42.3 in Table 2; if the intended meaning is 'no extra parameters/FLOPs', the wording should say so.","section":"§4.3.2, Table 2"},{"comment":"Table 9 contains the label 'YOLOv11-n-obb' while the model is YOLO11-n-obb, and the rotated-detection section does not state the exact crop overlap and test-time augmentation settings for the MHAF rows beyond the default description; please align notation and protocol descriptions.","section":"Table 9, §4.6.4"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of a computer-vision methods journal, but the state-of-the-art framing is currently stronger than the controlled evidence supports. I would ask the editor to require the revised manuscript to report same-recipe baseline runs, COCO test-dev results, and latency measurements before considering acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First impression: this is a capable, incremental architecture paper, and its own ablations are the most convincing part. The MAFPN neck (SAF + AAF) and the RepHMS block with GHFKS kernel scheduling are a real combination, though each piece extends known ideas: PAFPN/BiFPN fusion, ELAN/CSP blocks, YOLO-MS heterogeneous kernel selection, and RepLKNet-style reparameterization. The step-by-step ladder in Table 5 (YOLOv10n + MAFPN + RepHMS + GHFKS + RepHConv) is clean and internally consistent: AP goes 38.5 to 40.1 to 41.2 to 41.9 to 42.3. The RepHMS-versus-other-blocks comparison in Table 1 also shows a genuine parameter-efficiency advantage. The reparameterization identity in Eq. (6) is plain algebra, no circularity there. Code is linked, though I have not tried to run it.\n\nThe soft spot is not the architecture, it is the measurement. The headline 1.9 AP gain over YOLO11-s is taken from published numbers, not from YOLO11 retrained under the paper's protocol. The training recipe is YOLOv10-style with RTMDet cached-mixup and lower-probability copy-paste, and the authors tune SAF/AAF/RepHMS/GHFKS on COCO val2017 (Sections 4.3-4.5) and then report final numbers on that same split (Section 4.6). That mix of model selection and evaluation can inflate the apparent gain. The nano scale shows the fragility: MHAF-YOLO-n reaches 42.3 AP, which is 0.1 below the authors' own MAF-YOLO-n (42.4). If the architecture were clearly SOTA at every scale, that would be less worrying; here the flagship small model carries the claim.\n\nAnother real gap: for a paper that calls itself real-time, there is no latency or FPS measurement anywhere. The conclusion openly admits that inference speed lags behind YOLOv10 and YOLO11 because of MAFPN's complexity and large-kernel depthwise convolutions. That is honest, but it undercuts the real-time positioning and should have been in the experiments. The transfer results to segmentation and rotated detection are useful supporting evidence, but they suffer from the same uncontrolled-comparison issue.\n\nVerdict: this deserves peer review, not desk rejection. A serious referee should ask for (1) controlled retraining of at least YOLO11 and YOLOv12 under the paper's schedule, or a test-dev evaluation, (2) variance estimates or at least multiple seeds, (3) latency numbers on the actual hardware, and (4) an explicit delta or comparison with MAF-YOLO, since the novelty is largely an extension of that prior work. With those additions the empirical claim would be much more trustworthy. As it stands, the architecture is plausible and the internal ablations are solid, but the SOTA-over-YOLO11 statement is not fully supported.","headline":"Genuine incremental architecture with clean internal ablations; the SOTA claim over YOLO11 rests on uncontrolled baselines and val-set tuning.","tokens_in":21232,"tokens_out":3418,"would_cite":false,"duration_ms":29547,"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":"This paper claims that a new neck design, MAFPN, lets a 7.1M-parameter YOLO reach 48.9% AP on COCO, beating YOLO11-s by 1.9 points with 24.4% fewer parameters.","keywords":["object detection","YOLO","multi-scale feature fusion","feature pyramid network","reparameterization","large kernel convolution","model efficiency","COCO"],"falsifier":"Train MHAF-YOLO-s and YOLO11-s from scratch with the identical data pipeline, epochs, and optimizer on COCO train2017 and compare AP on COCO val2017; if the reported 1.9-point edge at 7.1M parameters does not replicate, the central advantage claim is not robust.","tokens_in":20149,"feed_emoji":"🎯","tokens_out":5118,"duration_ms":41607,"temperature":0.7,"pith_summary":"The paper claims that the standard PAFPN neck used in most YOLO detectors wastes the information that shallow backbone layers carry about small objects, and that this can be fixed with a new neck architecture. The proposed MHAF-YOLO adds two auxiliary fusion paths, SAF at shallow layers and AAF at deep layers, plus a reparameterized heterogeneous multi-scale convolution module, and reports 48.9% AP on COCO at 7.1M parameters, 1.9 points above YOLO11-s with 24.4% fewer parameters. The authors also report gains when the neck is plugged into YOLOv8n and Faster R-CNN, and improved instance segmentation and rotated object detection. If correct, it means feature-fusion design is a currently underused axis for making real-time detectors both smaller and more accurate.","feed_headline":"7.1M-parameter YOLO hits 48.9% AP on COCO","feed_subtitle":"New multi-branch neck and reparameterized kernels beat YOLO11-s by 1.9 AP with 24.4% fewer parameters.","key_machinery":"The load-bearing object is MAFPN (Multi-Branch Auxiliary FPN), a neck that adds Superficial Assisted Fusion (SAF) and Advanced Assisted Fusion (AAF) to the standard bottom-up and top-down paths. SAF injects shallow backbone features into the neck's first fusion blocks to preserve small-object spatial detail; AAF densely connects four feature layers at the deeper output stages so the head receives richer gradients. Alongside it, RepHMS (Reparameterized Heterogeneous Multi-Scale) runs parallel small and large depthwise convolutions during training and folds them into a single kernel at inference, while GHFKS (Global Heterogeneous Flexible Kernel Selection) chooses larger kernels on higher-resolution layers to widen receptive fields. The mechanism's work is to make multi-scale information flow cheaply: no extra inference cost for the multi-kernel branches, and no extra parameters for the fusion paths beyond a few 1x1 convolutions.","core_discovery":"On the paper's own terms, the central discovery is that a one-stage detector's neck can be rebuilt to fuse non-adjacent scales in both directions, letting shallow spatial detail survive into deep layers and giving small objects a disproportionate accuracy gain. Specifically, MHAF-YOLO-s achieves 48.9% AP (49.1% with one-to-many NMS training) at 7.1M parameters and 25.3 GFLOPs on COCO val2017, which the authors report as state-of-the-art among real-time detectors and as a 1.9 AP improvement over YOLO11-s at 24.4% fewer parameters. The same architecture transfers to instance segmentation (39.7% seg AP for the small variant) and rotated object detection (81.10% mAP on DOTA-v1.0 in multi-scale mode), which the authors offer as evidence of generalization.","pith_inferences":["The ablations select hyperparameters on COCO val2017 and the final comparisons are also on val2017, so part of the reported gap may reflect tuning to that split; a test-dev evaluation would show whether the SOTA claim holds outside the selection set.","The authors state that inference speed trails YOLOv10 and YOLO11; if latency rather than parameters is the deployment constraint, the practical advantage over those models is smaller than the AP-per-parameter numbers suggest.","The training recipe differs from baselines (cached-mixup from RTMDet and lower-probability copy-paste), so a recipe-matched controlled comparison would isolate how much of the gain is architectural.","Since MAFPN is plug-and-play and RepHMS can enter any network, the same design is a natural test in transformer-based detectors, not just convolutional YOLO variants."],"forward_implications":["Real-time object detectors can improve accuracy per parameter by redesigning the neck instead of scaling up the backbone or head.","MAFPN as a plug-in reports AP gains on other detectors: +1.7 on YOLOv8n (with fewer parameters and 200 fewer epochs) and +1.2 on Faster R-CNN.","RepHMS-style reparameterized heterogeneous kernels give multi-scale receptive fields at no additional inference cost, since training-time branches merge into a single convolution.","The same detector body transfers to instance segmentation and rotated object detection with gains over YOLO11 counterparts."],"supporting_citations":[{"why":"Defines the PAFPN neck that the paper identifies as the main limitation and replaces with MAFPN.","marker":"[21]"},{"why":"YOLOv10 is the base framework that MHAF-YOLO builds on and the primary baseline in the ablations.","marker":"[17]"},{"why":"YOLO11 is the latest YOLO baseline that MHAF-YOLO-s claims to beat by 1.9 AP with fewer parameters.","marker":"[18]"},{"why":"MAF-YOLO is the predecessor whose RepHELAN and auxiliary fusion ideas MAFPN extends.","marker":"[16]"},{"why":"YOLO-MS supplies the heterogeneous kernel selection and MSBlock ideas behind RepHMS and GHFKS.","marker":"[14]"},{"why":"RepVGG provides the reparameterization technique used to merge heterogeneous convolution kernels at inference.","marker":"[27]"},{"why":"RTMDet is both a compared real-time detector and the source of the cached-mixup augmentation used in training.","marker":"[11]"},{"why":"COCO is the dataset on which the central state-of-the-art detection claim is evaluated.","marker":"[30]"},{"why":"DOTA-v1.0 is the benchmark used to demonstrate transfer to rotated object detection.","marker":"[32]"}],"fun_headline_variants":["MHAF-YOLO hits 48.9 AP on COCO with 7.1M params","New neck fuses scales, beats YOLO11-s by 1.9 AP","YOLO upgrade: 48.9 AP, 24.4% fewer params","Multi-branch fusion neck lifts YOLO to 48.9 AP","Small-target win: MHAF-YOLO scores 48.9 AP"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim rests on the published baseline numbers being accurate and fairly comparable, and on COCO val2017 serving both to tune the design and to score the final result, so a baseline run with a different training recipe could change the reported gaps.","fun_headline_variants_meta":{"raw":{"variants":["MHAF-YOLO hits 48.9 AP on COCO with 7.1M params","New neck fuses scales, beats YOLO11-s by 1.9 AP","YOLO upgrade: 48.9 AP, 24.4% fewer params","Multi-branch fusion neck lifts YOLO to 48.9 AP","Small-target win: MHAF-YOLO scores 48.9 AP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000275,"raw_usage":{"total_tokens":1694,"prompt_tokens":1048,"completion_tokens":646,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":664,"completion_tokens_details":{"reasoning_tokens":531}},"tokens_in":664,"tokens_out":646,"duration_ms":5422,"temperature":1.0,"reasoning_tokens":531,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T21:57:54.940449+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train MHAF-YOLO-s and YOLO11-s from scratch with the identical data pipeline, epochs, and optimizer on COCO train2017 and compare AP on COCO val2017; if the reported 1.9-point edge at 7.1M parameters does not replicate, the central advantage claim is not robust.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the PAFPN neck that the paper identifies as the main limitation and replaces with MAFPN."},{"cited_title":"Jocher, A","cited_arxiv_id":null,"evidence_quote":"YOLO11 is the latest YOLO baseline that MHAF-YOLO-s claims to beat by 1.9 AP with fewer parameters."},{"cited_title":"Multi-Branch Auxiliary Fusion YOLO with Re-parameterization Heterogeneous Convolutional for accurate object detection","cited_arxiv_id":"2407.04381","evidence_quote":"MAF-YOLO is the predecessor whose RepHELAN and auxiliary fusion ideas MAFPN extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"RepVGG provides the reparameterization technique used to merge heterogeneous convolution kernels at inference."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"COCO is the dataset on which the central state-of-the-art detection claim is evaluated."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DOTA-v1.0 is the benchmark used to demonstrate transfer to rotated object detection."}],"review_version":1}