{"id":"d98ec514-96b2-4dd3-9e8a-d3cd3fb5f570","arxiv_id":"2607.01555","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Introduces AC-SLSIoU loss combining logit-domain margin constraint, adaptive boundary suppression, and false-alarm focal loss to improve discrimination and contour quality in infrared small target detection.","lead":"The paper proposes AC-SLSIoU, a plug-and-play loss for infrared small target detection that enlarges logit-space gaps between weak targets and clutter while using scale-aware boundary penalties to reduce halo effects. A smart generalist might read it to see how targeted loss engineering can tackle imbalance and boundary problems in low-signal imaging without extra test-time cost.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Whether logit-domain margin enlargement and scale-aware annular penalties fix discrimination and boundary issues without introducing training instability or dataset-specific tuning","rationale":"The reader's weakest_assumption directly identifies the empirical reliability of the two core mechanisms (LDMC and ABS) as the load-bearing point; the abstract-only limitation noted by the reader is now superseded by the instruction to treat the full manuscript as available, but the same mechanistic assumption remains the least-secured link in the argument for consistent, overhead-free gains.","tokens_in":1757,"tokens_out":368,"duration_ms":19301,"concrete_test":"Re-train the reported baseline detectors on the three main IRSTD benchmarks while sweeping the LDMC margin parameter over {0.5, 1.0, 2.0, 4.0} and the ABS annulus scale factor over {1.5, 2.0, 3.0}; if any sweep produces >5% drop in Pd or >10% rise in Fa relative to the paper's reported best setting on even one dataset, the assumption of reliable, tuning-free improvement is falsified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that post-activation saturation and missing contour constraints are the dominant bottlenecks, and that LDMC (enlarging target vs. hard-negative logit gaps) plus ABS (scale-aware annular penalties) will reliably widen usable margins and suppress halos without creating new modes such as logit explosion, gradient instability on tiny targets, or the need for per-dataset retuning of margin/annulus hyperparameters. If the logit-space constraint interacts poorly with the existing detector heads or if the annular penalty scale estimation fails on the smallest targets, the \"consistent improvement\" and \"seamless integration\" claims would not hold.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper claims that post-activation probability saturation and missing explicit contour constraints limit IRSTD performance, and introduces the plug-and-play AC-SLSIoU loss comprising Logit-Domain Margin Constraint (LDMC) to enlarge target vs. hard-negative logit gaps, Adaptive Boundary Suppression (ABS) via scale-aware annular penalties to refine contours and suppress halos, plus False-Alarm Focal Loss. It asserts seamless integration into existing detectors with no inference overhead and consistent gains in accuracy and shape quality, supported by extensive experiments and cross-backbone tests.","tokens_in":1874,"tokens_out":509,"duration_ms":18435,"significance":"If the logit-domain margin enlargement and adaptive annular penalties deliver robust gains without new instabilities or per-dataset retuning, the method would offer a lightweight, generalizable improvement to IRSTD pipelines by directly targeting discrimination and boundary issues in a manner compatible with standard detector heads.","major_comments":[{"comment":"§3.2, LDMC definition: the hard-negative selection criterion and margin hyperparameter are not shown to be dataset-independent; if selection thresholds or the margin value must be retuned per dataset or backbone, this contradicts the 'seamless integration' and 'no extra tuning' claims.","section":"§3.2"},{"comment":"§4.2, ABS formulation: the scale-aware annulus radius estimation relies on an initial target-size prediction that may be unreliable for the smallest targets; no ablation demonstrates that this does not introduce gradient instability or halo suppression failures on sub-3-pixel targets.","section":"§4.2"},{"comment":"Table 4, cross-backbone rows: while mean improvements are reported, the absence of per-run variance or statistical significance tests leaves open whether the gains are consistent or could be explained by hyperparameter sensitivity in LDMC/ABS.","section":"Table 4"}],"minor_comments":[{"comment":"The abstract supplies no quantitative metrics, dataset names, or backbone list; adding one representative result and the evaluation protocol would improve readability.","section":null},{"comment":"Notation for the annular penalty weights in Eq. (8) would benefit from an accompanying diagram showing the inner/outer radii relative to predicted target scale.","section":"Eq. (8)"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive feedback. We address each major comment below, indicating where revisions will be made to strengthen the manuscript.","responses":[{"response":"The manuscript describes the LDMC margin and selection criterion in §3.2 but does not include explicit cross-dataset ablations confirming fixed values suffice without retuning. We will add an ablation table in the revision that applies the same fixed margin and selection rule across all reported datasets and backbones to directly support the no-extra-tuning claim.","revision_made":"yes","referee_comment":"[§3.2] §3.2, LDMC definition: the hard-negative selection criterion and margin hyperparameter are not shown to be dataset-independent; if selection thresholds or the margin value must be retuned per dataset or backbone, this contradicts the 'seamless integration' and 'no extra tuning' claims."},{"response":"The manuscript does not provide a dedicated ablation isolating sub-3-pixel targets for the ABS module. We will insert such an ablation in the revised version, reporting both detection metrics and gradient-norm statistics on the smallest targets to verify stability.","revision_made":"yes","referee_comment":"[§4.2] §4.2, ABS formulation: the scale-aware annulus radius estimation relies on an initial target-size prediction that may be unreliable for the smallest targets; no ablation demonstrates that this does not introduce gradient instability or halo suppression failures on sub-3-pixel targets."},{"response":"Table 4 reports only mean values; variance and significance tests are absent. We will revise the table to include per-backbone standard deviations from repeated runs (where computational budget permits) or add a short discussion of observed run-to-run consistency.","revision_made":"partial","referee_comment":"[Table 4] Table 4, cross-backbone rows: while mean improvements are reported, the absence of per-run variance or statistical significance tests leaves open whether the gains are consistent or could be explained by hyperparameter sensitivity in LDMC/ABS."}],"tokens_in":1424,"tokens_out":453,"duration_ms":28026,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The one thing to know is that this paper offers AC-SLSIoU, a loss for infrared small target detection built from a logit-domain margin constraint, adaptive annular boundary suppression, and false-alarm focal weighting.\n\nThey start from the usual problems in this task: weak targets produce saturated probabilities that are hard to separate from clutter, and thermal diffusion plus tiny scale produce halo predictions. LDMC widens the logit gap between targets and hard negatives. ABS adds scale-aware annular penalties around each target to push down overflow responses. The focal term then hits persistent high-probability false alarms harder. The whole thing is written as a drop-in replacement that leaves inference unchanged.\n\nThe approach is practical for the IRSTD setting. The components line up with documented weaknesses in the literature, and testing across backbones is a fair check for whether the changes travel with different detectors.\n\nThe soft spot is that the abstract contains no metrics, ablation tables, or dataset details, so it is impossible to tell how large the claimed improvements are or whether they survive different conditions. The stress-test worry about training instability or the need for per-dataset margin and annulus tuning is not addressed in the summary. If the full paper shows stable training and gains that do not require extra fiddling, the method would be more convincing.\n\nThis paper is for researchers already working on infrared small target detection in computer vision. A reader who builds or tunes detectors for surveillance or defense imagery might find the specific loss terms worth trying.\n\nIt deserves a serious referee because the task is applied and the proposal is concrete enough to evaluate. I would recommend sending it to peer review so the experimental claims can be checked in detail.","headline":"The paper adds a logit-domain margin and annular shape penalties to an IR small target detection loss, but the abstract gives no numbers so the size of any gains is unclear.","tokens_in":2378,"tokens_out":420,"would_cite":false,"duration_ms":29228,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Enlarging logit gaps and applying scale-aware annular penalties improves infrared small target detection without added inference cost.","keywords":["infrared small target detection","logit domain margin","adaptive boundary suppression","plug-and-play loss","shape refinement","false alarm focal loss","IRSTD"],"falsifier":"Attach AC-SLSIoU to a baseline detector, train and evaluate on NUAA-SIRST or similar standard IRSTD set, and check whether precision, recall, or boundary IoU metrics show no gain or decline relative to the identical detector trained without the loss.","tokens_in":2645,"feed_emoji":"🎯","tokens_out":661,"duration_ms":21832,"temperature":0.7,"pith_summary":"Infrared small target detection faces saturation when weak targets and clutter produce similar post-activation probabilities, plus halo artifacts from missing contour rules. The paper introduces AC-SLSIoU, a loss that first widens response gaps directly in logit space between targets and hard negatives. It then applies adaptive annular penalties around predicted boundaries, scaled to target size, to sharpen edges and suppress overflow. A focal term further down-weights persistent high-confidence false alarms. The loss attaches to any existing detector and raises both detection accuracy and shape metrics on benchmarks.","feed_headline":"Logit margins plus annular penalties lift IR target detection","feed_subtitle":"Plug-in loss widens pre-activation gaps and refines contours to raise accuracy and shape quality at zero inference cost.","key_machinery":"AC-SLSIoU loss combining Logit-Domain Margin Constraint to separate pre-activation responses, Adaptive Boundary Suppression via scale-aware annular penalties, and False-Alarm Focal Loss.","core_discovery":"The Adaptive-Contrastive SLSIoU loss improves detection by replacing post-activation supervision with a logit-domain margin constraint that enlarges target-to-negative gaps before softmax, combined with scale-aware annular boundary penalties that penalize halo-like overflow and enforce contour fidelity, plus focal re-weighting of high-probability negatives; these terms integrate into existing networks and raise both accuracy and shape quality with no inference overhead.","pith_inferences":["The logit-domain and annular-penalty ideas could transfer to other small-object detection settings that suffer probability saturation or boundary blur.","Training dynamics might stabilize because the margin term reduces the dominance of easy negatives early in optimization.","If target scales vary widely within one scene, the adaptive scaling of annular penalties may still need dataset-level adjustment to avoid under- or over-suppression."],"forward_implications":["Existing IRSTD detectors gain better weak-target discrimination from the logit margin term.","Predicted shapes become tighter and less halo-prone due to the adaptive annular penalties.","No extra runtime cost occurs at inference because the loss affects only training.","Gains appear consistently across different backbone networks in cross-evaluation."],"fun_headline_variants":["Logit margins widen target to negative response gaps in IR detection","Annular penalties curb halo effects around small IR targets","Boundary suppression refines contours in infrared small target detection","Focal loss downweights persistent false alarms in IRSTD"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Core limits of earlier methods come from probability saturation after activation and absent explicit contour constraints, and that logit gap enlargement plus annular penalties will correct them without new failure modes or dataset-specific tuning.","fun_headline_variants_meta":{"raw":{"variants":["Logit margins widen target to negative response gaps in IR detection","Annular penalties curb halo effects around small IR targets","Boundary suppression refines contours in infrared small target detection","Focal loss downweights persistent false alarms in IRSTD"]},"model":"grok-4.3","cost_usd":0.005825,"raw_usage":{"total_tokens":2787,"prompt_tokens":699,"num_sources_used":0,"completion_tokens":64,"cost_in_usd_ticks":58249500,"prompt_tokens_details":{"text_tokens":699,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2024,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":699,"tokens_out":64,"duration_ms":19790,"temperature":1.0,"reasoning_tokens":2024,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-03T00:51:52.739414+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Attach AC-SLSIoU to a baseline detector, train and evaluate on NUAA-SIRST or similar standard IRSTD set, and check whether precision, recall, or boundary IoU metrics show no gain or decline relative to the identical detector trained without the loss.","supporting_citations":[],"review_version":1}