{"id":"65f67ef3-90d0-4cb7-ba67-2e8dffcb585c","arxiv_id":"2608.06205","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":9,"one_line_summary":"CFGPNet combines RepViT-based backbones, cross-modal attention gating, and attention-based fusion to achieve SOTA mAP on FLIR, M3FD, LLVIP, VEDAI, and MFAD.","lead":"CFGPNet is a new RGB-thermal object detector that fuses visible and infrared images with cross-attention and attention-based fusion, reporting state-of-the-art accuracy on five benchmarks. It is a practical improvement for low-light and adverse-weather perception, offering three model sizes with different speed-accuracy trade-offs.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper internally contradicts itself on the final training loss: §3.6 specifies MSE+SSIM+CE-KL, while ablation Table 12 shows that exact combination collapses performance and that CE-KL alone is selected; the objective behind the reported SOTA is therefore unidentified.","rationale":"The reader's verdict of REJECT is well supported. The reader identified the loss-configuration contradiction in the rationale, while the formal 'weakest_assumption' focused on training-budget and post-processing unfairness. I agree with the rejection but locate the most load-bearing problem earlier in the chain of evidence: the paper's own ablation invalidates the loss configuration described as final. This is not a disagreement with a consensus view or an efficiency concern; it is an internal inconsistency that cannot be resolved by re-reading the paper. The architecture description is unusually detailed, and the training schedule, augmentation, NMS thresholds, and per-class results are explicitly specified, which is genuine supporting material. However, those details do not help because the training objective itself is ambiguous. A single focused reproduction of CFGPNet-m on MFAD with the two candidate losses would settle whether Table 12 is trustworthy and which configuration yields the reported headline. Until that check is done, the reported SOTA numbers and the architectural attribution should not be accepted. My recommendation is UNCHANGED because the reader already reached REJECT, and my independent reading strengthens rather than alters that verdict.","tokens_in":60805,"tokens_out":2647,"duration_ms":30233,"concrete_test":"Reconstruct CFGPNet-m from Appendix A.18a, train on MFAD for 600 epochs under the §4.3 protocol with the §3.6 combined loss (MSE+SSIM+CE-KL, weights 300/3/1), and separately with CE-KL only at weight 1.0. Compare final mAP50/mAP50:95 against Table 12 rows 'MSE+SSIM+CE-KL' (74.3/51.5) and 'CE-KL' (79.6/56.7). If the combined-loss run does not reproduce 74.3/51.5 or the CE-KL run does not reproduce 79.6/56.7, the published ablation cannot be trusted; if neither matches the headline values, the final model's training objective is unidentified and the central claim is unreproducible.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that CFGPNet's architecture produces the reported state-of-the-art results—is not verifiable because the manuscript specifies two mutually inconsistent final loss configurations. Section 3.6 defines the deployed objective as the YOLOv9 terms plus all three proposed alignment losses, with λ_CE-KL=1.0, λ_SSIM=3.0, and λ_MSE=300.0. Section 4.5, Table 12, however, reports that this exact combination (MSE+SSIM+CE-KL) degrades CFGPNet-m on MFAD to 74.3 mAP50 and 51.5 mAP50:95, far below the paper's headline 79.6/56.7 for the same model and dataset. The same ablation identifies CE-KL alone as best (79.6/56.7), and the text states that CE-KL is selected for the final model. No passage reconciles these statements: if the §3.6 combined loss was used, the headline results contradict the paper's own ablation; if CE-KL-only was used, §3.6 and its hyperparameters do not describe the final training objective. Because the loss function is the core optimization setup, this ambiguity makes the reported gains impossible to attribute to CrossCEA, ASAF, or PGI, and it blocks independent reproduction of the experiments. The baseline-fairness issue raised by the reader is real, but this internal contradiction is more fundamental: even with identical training budgets, one cannot know what was trained.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CFGPNet, a multispectral (RGB-T) object detection framework combining a RepViT-based GELAN backbone, a CrossCEA cross-modal attention module, an ASAF fusion module, and a programmable-gradient auxiliary branch (PGI). Three model scales (m/c/e) are designed and evaluated on FLIR, M3FD, LLVIP, VEDAI, and MFAD, where the paper reports state-of-the-art or near-state-of-the-art results. The authors also provide detailed layer-wise configuration tables, release code/data/models, and conduct ablation studies on MFAD to support the contribution of individual components.","tokens_in":61174,"tokens_out":7664,"duration_ms":80609,"significance":"If the reported results are reproducible, CFGPNet would represent a practical advance in multispectral detection, combining competitive accuracy with a favorable efficiency trade-off. The paper is commendable for its detailed architectural descriptions, the release of code and fine-tuned models, and the broad evaluation across five benchmarks with per-class breakdowns. The ablations are extensive and mostly controlled, and the disclosure of dataset-specific NMS thresholds and the use of a single VEDAI split is transparent. However, the current manuscript contains a load-bearing internal contradiction in the training-loss specification that prevents verification of which configuration produced the headline results; this issue, together with a lack of same-protocol baselines, makes the central claim not yet defensible as written.","major_comments":[{"comment":"The training objective is specified inconsistently. In Section 3.6, the total loss is defined as the YOLOv9 detection losses plus all three proposed alignment terms, with λ_CE−KL=1.0, λ_SSIM=3.0, and λ_MSE=300.0, and the text states that these coefficients are kept fixed across datasets. However, Table 12 of the ablation study reports that this exact combination (MSE+SSIM+CE-KL) yields only 74.3 mAP50 / 51.5 mAP50:95 on MFAD, while CFGPNet-m is reported at 79.6 / 56.7. The same table identifies CE-KL alone as the best loss, and the text in Section 4.5 explicitly states that \"CE-KL is selected for the final model.\" These two statements are mutually contradictory. If the final models were trained with the Section 3.6 loss, the headline results are contradicted by the paper's own ablation; if the final models used CE-KL only, then Section 3.6 and its hyperparameters do not describe the actual training objective. Since the loss function is central to the optimization, the reported state-of-the-art numbers cannot be attributed to the proposed architecture, and independent reproduction is blocked. The authors must correct this inconsistency and specify exactly which loss was used for each reported result, ideally by providing training logs or rerunning the affected experiments.","section":"Section 3.6 vs. Table 12"},{"comment":"The comparison with prior methods is not conducted under equal conditions. All CFGPNet variants are trained from scratch for 600 epochs, use dataset-specific NMS IoU thresholds selected on a 10% validation subset (Table 2), and are evaluated with the best checkpoint over the training schedule (as stated for VEDAI). In contrast, baseline results are copied from their original publications, which may use different training budgets, data splits, input resolutions, and post-processing settings. Consequently, the reported gains cannot be isolated to the architectural contributions (CrossCEA, ASAF, PGI). To support the claim of architecture-level superiority, the authors should provide at least one strong baseline (e.g., a YOLOv9-based dual-stream detector, or the best performing published method) trained under the same 600-epoch schedule, same data augmentation, same NMS calibration, and same checkpoint-selection protocol. Without such a control, the state-of-the-art claim is not convincingly established.","section":"Sections 4.3 and 4.4"},{"comment":"Even if the loss-contradiction is resolved, the ablation provides only weak support for the proposed feature-alignment loss. In Table 12, the \"None\" configuration achieves 79.7 mAP50 and 56.6 mAP50:95, while CE-KL alone achieves 79.6 and 56.7, i.e., a 0.1-point improvement on mAP50:95 and a 0.1-point drop on mAP50. The text acknowledges this but nevertheless selects CE-KL. Given that one of the paper's stated contributions is the auxiliary feature-alignment supervision of CrossCEA, the negligible and sometimes negative effect of the alignment loss should be discussed more candidly. The authors should either provide additional ablations (e.g., different λ values, other datasets) showing a consistent benefit, or moderate the claim that this loss is a key enabler of the reported performance.","section":"Table 12 and Section 3.6"}],"minor_comments":[{"comment":"The MBatt branch ablation reports results for 2 to 6 branches, but the text refers to the case of a single branch (\"When only one branch is used...\"); either add the N_b=1 row to the table or adjust the discussion to avoid referencing a configuration that is not tabulated.","section":"Table 14 / Section 4.5"},{"comment":"The FPS values differ by an order of magnitude between the main comparison (e.g., CFGPNet-m: 52.2 FPS) and the ablation tables (e.g., 1463.4 FPS for the same model). Although the batched timing protocol is explained, the presentation is confusing. Please clearly label ablation FPS as \"relative batched timing\" in every table caption and in the text wherever such values appear.","section":"Tables 3 vs 10–16"},{"comment":"The sentence \"each variant is evaluated using its best-performing checkpoint over the same 600-epoch training schedule\" should specify whether the best checkpoint is chosen using the validation subset or the test split. If the test set was used for checkpoint selection, the reported numbers would be optimistically biased.","section":"Section 4.4.4 (VEDAI)"},{"comment":"The claim of \"strong and consistent performance\" is weaker on FLIR, where the mAP50 gain over the best published method (ERFF, 80.6) is only 0.1 point for CFGPNet-e. Please present a more nuanced interpretation of the margin of improvement on each dataset.","section":"Abstract and Section 4.4"}],"recommendation":"major_revision","confidential_remarks":"The loss specification contradiction in Section 3.6 versus Table 12 is severe enough that, in my view, the paper should not be accepted until the authors clearly state which objective produced each result and provide supporting training logs or a rerun. The baseline-fairness issue is also important; even a corrected manuscript would need at least one same-protocol baseline to substantiate the architecture-level attribution. I would encourage the editor to demand these changes explicitly in a revision decision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a real architecture paper, not a toy. The framework — RepViT-GELAN backbones, CrossCEA cross-modal attention-map swapping, ASAF dense-plus-max-selected attention fusion, and a PGI-style auxiliary branch — is coherent, and the empirical work is unusually thorough: five benchmarks, three model scales, detailed ablations, per-class results, and a public repo. I'd also give credit for the honest limitation paragraphs: they openly state single-run point estimates, best-checkpoint selection, held-out NMS tuning, and the single-split VEDAI evaluation. That's more transparency than most papers in this area.\n\nBut the reader and the stress-test note are right about the load-bearing flaw. Section 3.6 defines the deployed objective as the YOLOv9 losses plus all three alignment terms with λ_CE-KL=1.0, λ_SSIM=3.0, λ_MSE=300.0. Table 12 shows that exact combination collapses MFAD to 74.3 mAP50 / 51.5 mAP50:95, while CE-KL alone gives 79.6/56.7 — precisely the headline numbers for CFGPNet-m. The text then says CE-KL is selected. The two statements cannot both be true. Either the reported numbers came from a different objective than the one specified in §3.6, or the ablation is inconsistent. Either way, an independent researcher cannot reproduce the training setup, and the attribution of gains to CrossCEA/ASAF/PGI is unverifiable. This is not a minor typo: the loss function is the optimization objective.\n\nThe baseline-fairness issue is secondary but real: 600 epochs from scratch with per-dataset NMS tuning and best-checkpoint selection makes comparisons against published single-run baselines uneven. That alone wouldn't kill the paper, but combined with the loss ambiguity it means the numbers are currently not evidence.\n\nRecommendation: send it to review, but expect heavy revision. The architecture is plausible and the ablations are informative. The authors need to resolve the loss contradiction, clearly report the actual final configuration, and ideally provide same-budget baselines. If they do, this could be a solid practical contribution. As written, it should not be accepted.","headline":"A detailed, transparent architecture paper whose headline numbers are undermined by an internal contradiction between the specified final loss and the ablation table; the reported SOTA is not reproducible until that is resolved.","tokens_in":61752,"tokens_out":1853,"would_cite":false,"duration_ms":21646,"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":"CFGPNet claims that visible–thermal object detection reaches leading results on five benchmarks by having each modality re-weight the other through exchanged spatial attention maps, with a gradient-programmed auxiliary branch carrying the…","keywords":["multispectral object detection","cross-modal feature fusion","cross-modal interaction","programmable gradient information","cross-attention","RGB-T detection","visible-infrared fusion","re-parameterized backbone"],"falsifier":"Retrain the strongest published RGB–T baselines under CFGPNet's exact protocol — 600 epochs from scratch at 640×640, the same batch-size-adjusted schedules, the same dataset-specific NMS IoU thresholds chosen on 10% held-out validation subsets, and best-checkpoint selection — and compare on the five benchmarks. If a baseline matches or exceeds CFGPNet under that protocol, the architecture-attributed gains are refuted; a complementary check replaces CrossCEA's swapped attention maps with concatenation or early feature mixing at matched computation and asks whether the mAP50:95 margin survives.","tokens_in":60548,"feed_emoji":"🎯","tokens_out":12841,"duration_ms":108732,"temperature":0.7,"pith_summary":"This paper claims that visible-plus-thermal object detection works best when the two modalities cooperate through exchanged spatial attention maps rather than raw features: each stream computes a compact reliability map and the other stream is re-weighted by it, preserving texture and thermal evidence while suppressing single-modality bias. Around that idea the paper assembles a full detector, CFGPNet, using a re-parameterized YOLOv9-style backbone, a selective aggregation fusion module, and a gradient-programming auxiliary branch that keeps a second detection path alive during training and inference. CFGPNet reports state-of-the-art or leading results on all five public RGB–T benchmarks it evaluates, including 80.7% mAP50 / 45.0% mAP50:95 on FLIR and 83.4% / 61.8% on MFAD, across three model scales from 21.0M to 180.9M parameters. A sympathetic reader would care because low-light and adverse-weather perception for driving, surveillance, and robotics depends on exactly this kind of fusion, and the paper argues it can be done without the heavy cost of dense Transformer-style cross-attention.","feed_headline":"Fusing RGB and thermal by swapping attention maps tops five benchmarks","feed_subtitle":"Each modality guides the other's spatial attention, so the detector skips heavy transformer costs.","key_machinery":"The load-bearing mechanism is CrossCEA, a cross-attention module that swaps compact spatial reliability maps between the visible and thermal branches. Inside each CEA unit, the input feature map is split into channel groups; axis-pooled descriptors pass through depthwise 1D convolutions, group normalization, and sigmoid to form height- and width-wise gates, while a parallel global branch builds channel queries; a temperature-scaled interaction between the local and global paths produces a single 2D attention map $\\Omega$ per modality. CrossCEA then gates each stream with the other stream's map, $\\tilde{F}^I = F^I \\odot \\Omega^V$ and $\\tilde{F}^V = F^V \\odot \\Omega^I$, so that only bounded attention weights, not raw feature values, cross between modalities. The second mechanism is ASAF, which concatenates the enhanced streams and runs two parallel paths — a DenseNet-style dense aggregation refined by CBAM, and an ELA-guided multi-branch selector (MBatt) whose element-wise max keeps the strongest attention candidate at each location — before a 1×1 projection yields the compact fused tensor. The third mechanism is optimization-level: the PGI auxiliary branch instantiates a second detection path whose loss (weighted 0.25) sends programmed gradients back through tapped intermediate features of the main branch, and the branch stays active at inference.","core_discovery":"The paper's central claim is that a multispectral detector built from three interacting mechanisms — a RepViT-block GELAN backbone with separate visible and infrared weights, a Cross Computation Efficient Attention (CrossCEA) stage that exchanges compact spatial attention maps between the two streams, and an Attention Selection and Aggregation Fusion (ASAF) network that condenses the enhanced streams into one tensor — outperforms recent RGB–T detectors on every benchmark it evaluates. The best reported configurations reach 80.7% mAP50 / 45.0% mAP50:95 on FLIR, 89.9% / 63.4% on M3FD, 97.8% / 68.9% on LLVIP, 83.3% / 56.9% on VEDAI, and 83.4% / 61.8% on MFAD. The author would add that the framework is trained from scratch in three scales (21.0M, 71.7M, and 180.9M parameters), that the programmable-gradient auxiliary branch remains active at inference and contributes candidate detections merged by non-maximum suppression, and that the MFAD ablations attribute the gains to the combination of the RepViT backbone, CrossCEA, ASAF, and the PGI branch rather than to any single module.","pith_inferences":["The paper never runs the clean attribution experiment: retraining strong published baselines under its exact 600-epoch, from-scratch protocol with identical NMS tuning and checkpoint selection; without that control, the size of the architecture-level margin over prior work is an open question.","The attention-map-swap principle is a transferable hypothesis for any aligned sensor pair with location-dependent reliability — RGB-depth, RGB-event, or multi-spectral satellite imagery — and could be tested at matched compute against concatenation, learned-gate, or Transformer fusion.","The ablations show that pointwise and structural alignment losses (MSE, SSIM) actively degrade accuracy, which suggests forcing modalities toward one another is counterproductive; a useful follow-up is to test whether any cross-modal alignment term is needed once the backbone and fusion are strong.","The abstract's LLVIP headline pair mixes scales: 97.8% mAP50 comes from CFGPNet-m and 68.9% mAP50:95 from CFGPNet-e, and no single variant achieves both; with accuracy reported as single-run point estimates, small margins on saturated benchmarks may fall within run-to-run variance, which the conclusion itself flags as future work."],"forward_implications":["If the reported results hold, the 'exchange attention maps, not features' principle offers a lightweight alternative to dense Transformer cross-attention for RGB–T fusion, since CrossCEA adds no token-wise attention and the three scales report 18.6–52.2 FPS at 640×640 input.","The compact 21.0M-parameter variant claims to match or beat much larger published fusion detectors on FLIR, M3FD, and LLVIP, suggesting strong parameter-efficient RGB–T detection is possible without ImageNet-pretrained initialization.","The MFAD ablation quantifies the optimization contribution: removing the PGI branch drops mAP50 from 79.6 to 74.1 and mAP50:95 from 56.7 to 53.3, so the gradient-programming path is a substantial part of the reported gain.","The paper's VEDAI evaluation uses Split 1 for all variants and includes only single-split reports from prior work, so its VEDAI comparisons are conditional on that split being representative of the benchmark."],"supporting_citations":[{"why":"YOLOv9 supplies the GELAN backbone template, the PGI gradient-programming scheme, the dual detection head, and the detection losses that CFGPNet adapts to the multispectral setting.","marker":"[36]"},{"why":"RepViT provides the re-parameterized convolutional block that replaces the RepVGG baseline inside GELAN; the controlled ablation credits it with a 3.2 mAP50 gain.","marker":"[40]"},{"why":"EMA supplies the grouped multi-scale attention and cross-dimension interaction ideas that the CEA unit adapts for lightweight attention computation.","marker":"[49]"},{"why":"ELA supplies the axis-wise positional encoding with 1D convolution and group normalization used both in CEA and in ASAF's MBatt branch.","marker":"[50]"},{"why":"Provides the aligned FLIR variant that filters misregistered RGB–T pairs; the FLIR results reported by CFGPNet are measured on this aligned version.","marker":"[65]"},{"why":"DenseNet supplies the dense feature aggregation structure used in the dense branch of ASAF.","marker":"[53]"},{"why":"CBAM supplies the channel-plus-spatial refinement applied to the dense branch output inside ASAF.","marker":"[54]"},{"why":"Supplies the scaled dot-product multi-head attention formulation inserted in the neck's RepViTBottleneck blocks.","marker":"[43]"}],"fun_headline_variants":["Swapped attention maps lift RGB–T detection on five benchmarks","Cross-modal attention fusion beats five RGB-T benchmarks","Gradient-programmed fusion network tops multispectral sets","Efficient cross-attention RGB-T detector wins five benchmarks","CFGPNet: RepViT + CrossCEA + ASAF beats five multispectral datasets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported superiority over prior detectors assumes that CFGPNet's longer training schedule (600 epochs from scratch), per-dataset NMS IoU threshold tuning, and best-checkpoint selection are not what produce the gains, since no published baseline is retrained under the same protocol.","fun_headline_variants_meta":{"raw":{"variants":["Swapped attention maps lift RGB–T detection on five benchmarks","Cross-modal attention fusion beats five RGB-T benchmarks","Gradient-programmed fusion network tops multispectral sets","Efficient cross-attention RGB-T detector wins five benchmarks","CFGPNet: RepViT + CrossCEA + ASAF beats five multispectral datasets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000464,"raw_usage":{"total_tokens":2441,"prompt_tokens":1194,"completion_tokens":1247,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":810,"completion_tokens_details":{"reasoning_tokens":1160}},"tokens_in":810,"tokens_out":1247,"duration_ms":11401,"temperature":1.0,"reasoning_tokens":1160,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:35:47.060912+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the strongest published RGB–T baselines under CFGPNet's exact protocol — 600 epochs from scratch at 640×640, the same batch-size-adjusted schedules, the same dataset-specific NMS IoU thresholds chosen on 10% held-out validation subsets, and best-checkpoint selection — and compare on the five benchmarks. If a baseline matches or exceeds CFGPNet under that protocol, the architecture-attributed gains are refuted; a complementary check replaces CrossCEA's swapped attention maps with concatenation or early feature mixing at matched computation and asks whether the mAP50:95 margin survives.","supporting_citations":[],"review_version":1}