{"id":"21a57c10-e5ac-47ad-91c6-f6fb9de8a488","arxiv_id":"2601.12507","paper_version":2,"verdict":"UNVERDICTED","confidence":"UNKNOWN","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":1,"one_line_summary":"CoLR-Det is announced in the abstract, but the supplied full text is actually the authors' SDCoNet paper, so the claimed result is not present in the artifact.","lead":"The abstract announces CoLR-Det, a framework that uses super-resolution as a training-only regularizer for small-object detection, but the body of the paper is a different method, SDCoNet. As submitted, CoLR-Det's methods, experiments, and equations are absent, so no scientific claim from the abstract can be verified.","discovery_kind":"unclear","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The full text is SDCoNet, not CoLR-Det: no equation, algorithm, or table in the artifact instantiates CoLR-Det's latent-restoration claim, so the abstract's SOTA claim is unsupported as submitted.","rationale":"The reader's verdict is UNVERDICTED, and I agree with the central concern: the artifact is internally inconsistent, with CoLR-Det only in the title/abstract and SDCoNet in the body. I partially disagree with the reader's phrasing that the latent-regularization hypothesis is 'untested in the supplied text': Table V does provide an indirect test of the general idea that a training-only SR branch improves detection, under the SDCoNet architecture. However, that test is not for CoLR-Det, and it uses explicit image-level reconstruction via Eq. 4 and L_sr in Eq. 11, which actually conflicts with the abstract's claim of avoiding 'explicit image-level enhancement.' Thus the strongest claim—that CoLR-Det outperforms SOTA—is not supported by the artifact, and the appropriate verdict remains UNVERDICTED. No further objection is needed: the mismatch itself is the load-bearing issue.","tokens_in":23010,"tokens_out":5619,"duration_ms":59876,"concrete_test":"Run a full-text scan of the submitted PDF for 'CoLR-Det' and 'latent restoration' limited to the methods/experiments sections (excluding title/abstract). If neither term occurs in any equation, algorithm, or table—and no ablation is labeled CoLR-Det with the SR decoder removed at inference—the central claim is not supported by this artifact. As a positive control, confirm Table V's SR-branch rows are labeled 'SDCoNet'/'DINO', not CoLR-Det.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The submitted artifact is internally mismatched: the title, author list, and abstract describe CoLR-Det, but the body is a complete manuscript for SDCoNet. Sections III–IV give SDCoNet's equations (Eqs. 1–11), Algorithm 1, and experiments; the term 'CoLR-Det' does not appear in the method or experiment sections. Consequently, none of CoLR-Det's specific components are supported: the 'training-only restoration branch' and 'latent regularization' are not defined in any equation, the 'saliency-guided object-preserving token routing' is not the same as SDCoNet's saliency query filtering, and no ablation under the CoLR-Det name isolates the effect of removing the SR decoder at inference. The closest evidence, Table V, tests SDCoNet's SR branch and does show gains (e.g., DINO AP 0.766→0.816 on NWPU VHR-10-Split) with the SR head discarded at inference, so the general idea that auxiliary SR supervision can help detection is not entirely untested. But that does not rescue the abstract's claim, and it is arguably in tension with it: SDCoNet's SR branch reconstructs I_SR explicitly (Eq. 4) and is trained with pixel loss L_sr (Eq. 11), i.e., explicit image-level supervision, whereas CoLR-Det claims to avoid 'explicit image-level enhancement.' The artifact therefore cannot support the claimed CoLR-Det result.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The submission as received is internally inconsistent. The title, author list, and abstract describe \"CoLR-Det: Collaborative Latent Restoration for Small Object Detection in Low-Resolution Remote Sensing Images,\" whose central claim is that super-resolution supervision is used as a training-only latent regularizer on shared multiscale representations, with the SR decoder discarded at inference and no explicit image-level enhancement. However, the full text supplied is a complete manuscript for a different method, \"SDCoNet: Saliency-Driven Multi-Task Collaborative Network for Remote Sensing Object Detection.\" Sections I-V, all equations (Eqs. 1-11), Algorithm 1, and all experimental tables describe SDCoNet and never mention CoLR-Det. The reported results (e.g., Table II: AP 0.836 on NWPU VHR-10-Split; Table III: AP 0.419 on DOTAv1.5-Split; Table IV: AP 0.701 on HRSSD-Split) are SDCoNet results under explicit image-level SR supervision, not CoLR-Det results. Therefore the paper's stated contribution—latent restoration as an implicit regularizer—is not defined, derived, or evaluated in the submitted artifact.","tokens_in":23310,"tokens_out":2938,"duration_ms":36451,"significance":"If the CoLR-Det claim were validated—that a training-only auxiliary reconstruction constraint on shared features, with the SR decoder removed at inference, improves small-object detection without explicit image-level enhancement—it would be a meaningful contribution to low-resolution remote sensing detection. The existing SDCoNet body does provide some useful empirical material: Table V shows that adding an SR branch to Cascade R-CNN and DINO improves AP and especially AP_s on NWPU VHR-10-Split and DOTAv1.5-Split, with the SR head not used at inference, and Tables II-IV compare against several recent detectors under a controlled low-resolution protocol. These results give partial evidence for the broader idea that auxiliary SR supervision can help detection. However, none of this supports the specific latent-restoration mechanism claimed in the abstract. The artifact contains no derivation, algorithm, ablation, or experiment for CoLR-Det; the claimed mechanism is absent, and the one existing ablation (Table V) is for explicit pixel-level SR reconstruction (Eqs. 4 and 11), which is in tension with the abstract's claim of avoiding explicit image-level enhancement.","major_comments":[{"comment":"The manuscript body is SDCoNet, not CoLR-Det. The term \"CoLR-Det\" does not appear in the method section (Sections III.A-III.E), Algorithm 1, experimental section (Section IV), or conclusion (Section V). The equations describe SDCoNet's shared-encoder, U-Net SR decoder, saliency query filtering, and two-stage gradient routing. Consequently, none of the central CoLR-Det components—collaborative latent restoration, training-only restoration branch, or detection-oriented latent regularization—are defined or tested. The abstract's claim that CoLR-Det \"outperforms state-of-the-art methods\" is unsupported by any CoLR-Det experiment. This is not a presentation issue; the paper as submitted does not contain its claimed method.","section":"Abstract vs. full text; Sections III-V"},{"comment":"The claimed latent-restoration mechanism contradicts the actual equations in the body. The abstract states that CoLR-Det treats SR supervision as \"detection-oriented latent regularization rather than explicit image-level enhancement.\" But SDCoNet's SR branch explicitly reconstructs a high-resolution image: Eq. (4) defines I_SR = Conv(PixelShuffle(σ(Conv(F_dec)))), and Eq. (11) includes the pixel-level reconstruction loss β_sr L_sr in the stage-two objective. No equation or loss term applies reconstruction supervision to hidden shared representations as a latent regularizer. Thus the core mechanism advertised in the abstract is absent from the supplied method.","section":"Eq. (4), Eq. (11), Section III.C, Section III.E"},{"comment":"Table V is the only ablation involving removal of the SR branch, but it evaluates SDCoNet, not CoLR-Det. It shows AP improvements when the explicit SR branch is enabled during training (e.g., DINO on NWPU VHR-10-Split: AP 0.766→0.816; AP_s 0.487→0.552), with the SR head discarded at inference. This is evidence for the weaker, already known hypothesis that auxiliary SR training can help a detector. It does not test the CoLR-Det-specific claim that a latent-restoration branch without explicit image reconstruction outperforms explicit SR training, nor does it isolate the effect of \"latent regularization\" versus pixel-level loss. No table or figure in the manuscript instantiates CoLR-Det's proposed training-only restoration branch.","section":"Table V, Section IV.C"},{"comment":"Because the method and experiments belong to a different named framework, the internal consistency of the paper is broken at the level of its central claim. The reported SDCoNet results, even if technically valid, cannot be used to evaluate CoLR-Det. The manuscript would require a complete rewrite of the method, including a formal definition of the latent-restoration loss, an inference-time architecture, and experiments explicitly comparing CoLR-Det against SDCoNet and other baselines. This goes beyond what can be addressed through minor revisions.","section":"General assessment of soundness"}],"minor_comments":[{"comment":"The category distribution figures contain raw Unicode substitution strings such as \"/uni00000014/uni00000011/...\" instead of legible text labels, making Fig. 4 unreadable.","section":"Figure 4"},{"comment":"In the SDCoNet row, the AP_l column reads \"0.6190.680\" (likely two numbers concatenated without a separator). This should be corrected.","section":"Table IV, HRSSD-Split row"},{"comment":"The notation in Eq. (8), q_i = A(q_i + pos_i, q), is confusing: the attention operation uses q_i both as a query to be updated and as part of the argument list, while q is not prominently defined. Please clarify the query/key/value roles.","section":"Section IV.A, Equation (8)"},{"comment":"The reference list contains duplicates and inconsistencies: [55] is the same paper as [41]; [10] is cited as \"SR4IR\" but the title at the reference entry is \"Beyond image super-resolution for image recognition with task-driven perceptual loss,\" which should be checked for accuracy.","section":"References"}],"recommendation":"reject","confidential_remarks":"This submission appears to pair the title and abstract of CoLR-Det with the body of a separate SDCoNet manuscript. The disclosed code URLs also differ between the abstract (github.com/qiruo-ya/CoLR-Det) and the body (github.com/qiruo-ya/SDCoNet). This is not a case of a locally fixable gap in a derivation or experiment; the submitted paper does not describe the claimed method at all. I recommend rejection. If the authors believe the CoLR-Det results exist, they should submit a manuscript whose body actually defines and evaluates CoLR-Det."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — the one thing to know: this submission is two papers taped together. The title, author list, and abstract describe CoLR-Det, a latent-restoration framework that treats SR as training-only regularization. The full text is a complete, self-contained paper about SDCoNet, a shared-encoder multi-task network with saliency-guided token filtering and gradient routing. The word CoLR-Det appears nowhere in the method or experiments. I'm not reviewing a paper that exists.\n\nWhat is actually in front of me is SDCoNet, and on its own terms it's a decent piece of work. The idea of coupling SR and detection through a shared encoder, freezing the SR branch first and then routing gradients with a smaller SR learning rate, is sensible and reasonably well motivated. The experiments are extensive: three benchmarks, many baselines, ablations for the SR branch, saliency guidance, the α modulation, and token filtering efficiency. Table V is the key piece of evidence — removing the SR head at inference still gives DINO a +0.050 AP gain on NWPU VHR-10-Split — and that is a real result if it holds up.\n\nBut the soft spot is not minor. The abstract's central claim — CoLR-Det treats SR supervision as latent regularization rather than explicit image-level enhancement — is directly undermined by the body. SDCoNet reconstructs I_SR explicitly in Eq. 4 and trains it with pixel loss L_sr in Eq. 11. That is explicit image-level supervision. So even if we pretend the body is the paper, the abstract's conceptual framing doesn't match the concrete method. And the CoLR-Det components — the latent-restoration branch, the token-routing mechanism, the two-stage optimization as described — are never defined or tested under that name. The stress-test note is right.\n\nMy recommendation: send this back to the authors to fix the manuscript before any peer review. The mismatch is such that a reviewer cannot evaluate the claimed contribution. If the authors want to submit SDCoNet, they should submit SDCoNet. If they want to submit CoLR-Det, they need to actually write it up. As it stands, no serious referee should spend time on this artifact.","headline":"The submission is two different papers: the abstract describes CoLR-Det, the full text is a complete SDCoNet paper, so the claimed contribution is unsupported as submitted.","tokens_in":23799,"tokens_out":3294,"would_cite":false,"duration_ms":32052,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"CoLR-Det claims super-resolution can improve small-object detection as a training-only latent regularizer, not an image enhancer.","keywords":["remote sensing object detection","small object detection","super-resolution","latent regularization","multi-task learning","low-resolution imagery","token routing","DETR"],"falsifier":"Train the same detection head and shared encoder with the SR branch present versus absent (or with the SR loss weight set to zero) on the same low-resolution splits while keeping all other components fixed; if small-object AP does not clearly drop when the SR branch is removed, the latent-regularization claim is falsified. The paper's own ablations would constitute this experiment if they were run for CoLR-Det.","tokens_in":22879,"feed_emoji":"🛰️","tokens_out":6092,"duration_ms":58552,"temperature":0.7,"pith_summary":"The paper argues that the standard \"restore first, then detect\" pipeline for low-resolution remote sensing is misaligned: super-resolution amplifies texture and edges, while detection needs object semantics, so restored images can be visually sharper yet semantically misleading. It proposes CoLR-Det, which instead runs a super-resolution branch only during training, imposing reconstruction constraints on shared multiscale features and then discarding the branch at inference, so the deployment path is purely detection-driven. The claimed outcome is higher small-object detection accuracy on three resolution-degraded benchmark datasets without paying super-resolution compute at inference. A caveat is that the supplied full text actually describes a differently named architecture (SDCoNet) with the same broad strategy; the CoLR-Det-specific experiments and ablations asserted in the abstract are not present in the provided manuscript.","feed_headline":"Training-only super-resolution lifts small-object detection","feed_subtitle":"If correct, restoration can be discarded at inference, cutting compute while keeping the detection gains.","key_machinery":"The central mechanism is the training-only restoration branch: a super-resolution decoder that shares a Swin-Transformer encoder with the detection head, contributes an auxiliary reconstruction loss during training, and is removed at inference. Two supporting mechanisms carry the argument: saliency-guided token routing (tokens are ranked by a predicted saliency score, attention-based refinement is applied only to high-saliency tokens, and low-saliency tokens are preserved with a learnable background embedding), and detection-prioritized two-stage optimization (detection is trained alone with the SR branch frozen, then the SR branch is unfrozen at a smaller learning rate). Together they conve","core_discovery":"On the paper's own terms, the discovery is that super-resolution supervision can be redefined as detection-oriented latent regularization: a training-only restoration decoder attached to a shared encoder forces the shared features to retain small-object details, and can then be removed without harming detection at inference. The accompanying mechanisms—saliency-guided token routing, which concentrates attention on high-saliency tokens and keeps bypassed tokens through a learnable background embedding, and a two-stage schedule that stabilizes detection semantics before adding conservative SR updates—are presented as the means to make this regularization work. The abstract reports state-of-the","pith_inferences":["If this works, the same 'auxiliary pixel task as latent regularizer' trick may generalize to other pixel-level losses—denoising, deblurring, inpainting—so detection on degraded imagery could be improved without any restoration module at test time.","A natural extension is to test whether the regularization effect is tied to the specific reconstruction objective or just to the presence of an extra training-time task; replacing the SR loss with a random feature-matching loss would separate those.","The mismatch between the CoLR-Det abstract and the SDCoNet full text means the manuscript as submitted does not actually contain the described architecture; the abstract's performance numbers should be treated as unverified until the matching full paper and code are available."],"forward_implications":["If the latent-regularization claim holds, SR-assisted detection pipelines can remove the super-resolution network at inference entirely, eliminating its compute and latency while keeping the accuracy gain.","The framing implies reconstruction losses no longer need to be optimized for pixel fidelity (e.g., PSNR); only their effect on shared detection features matters, opening the door to cheaper or weaker restoration decoders.","The reported token-routing results suggest that in low-resolution remote sensing scenes, a large fraction of background tokens can be skipped without losing accuracy, which is a template for efficient Transformer-based detectors.","The two-stage training schedule indicates that freezing the auxiliary branch until detection semantics stabilize avoids gradient conflicts, a recipe that could transfer to other multi-task detection settings."],"fun_headline_variants":["Training-only SR lifts small-object detection","Latent restoration, not pixel SR, aids detection","Detect small objects with training-only restoration","No SR at inference: CoLR-Det still gains"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that training-only auxiliary reconstruction constraints on shared multiscale features improve detection accuracy once the super-resolution decoder is discarded at inference—a hypothesis the abstract asserts but the supplied full text (which describes a different architecture, SDCoNet) does not directly test.","fun_headline_variants_meta":{"raw":{"variants":["Training-only SR lifts small-object detection","Latent restoration, not pixel SR, aids detection","Detect small objects with training-only restoration","No SR at inference: CoLR-Det still gains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000235,"raw_usage":{"total_tokens":1365,"prompt_tokens":804,"completion_tokens":561,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":502}},"tokens_in":548,"tokens_out":561,"duration_ms":7731,"temperature":1.0,"reasoning_tokens":502,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T09:43:10.858197+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same detection head and shared encoder with the SR branch present versus absent (or with the SR loss weight set to zero) on the same low-resolution splits while keeping all other components fixed; if small-object AP does not clearly drop when the SR branch is removed, the latent-regularization claim is falsified. The paper's own ablations would constitute this experiment if they were run for CoLR-Det.","supporting_citations":[],"review_version":1}