{"id":"f1afc2b5-2ba5-41ad-b6e9-34d16be5cc0d","arxiv_id":"2508.12966","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"GazeDETR uses two disentangled decoders for head localization and gaze prediction, achieving state-of-the-art results on GazeFollow, VideoAttentionTarget, and ChildPlay.","lead":"GazeDETR is a new computer vision model that detects where people look by using two separate decoder modules, one for finding heads and one for predicting gaze direction. The model reports better accuracy than previous end-to-end gaze detection systems on three standard benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA margin may stem from increased decoder capacity rather than disentanglement; no capacity-matched ablation is visible.","rationale":"The full text is garbled/unreadable, so no internal inconsistencies can be verified. The reader's verdict of UNVERDICTED is appropriate. The strongest available claim is SOTA performance via disentangled decoders. The most concrete, testable risk is that the margin is confounded by decoder capacity: two decoders naturally have more parameters than one, and without a capacity-matched ablation, the 'disentanglement' mechanism is not established. This does not accuse the authors; it identifies a missing control that is standard for architecture claims. The proposed test would settle whether the concern lands, but because the full text is unavailable, the verdict remains UNCHANGED (UNVERDICTED).","tokens_in":14363,"tokens_out":6257,"duration_ms":64172,"concrete_test":"Recover a readable copy of the paper and locate Section 4.x ablation. Check whether there is an experiment training a single-decoder variant with the same encoder, the same number of decoder layers and hidden units (i.e., matched parameters) and the same training epochs. If the single-decoder variant matches or exceeds GazeDETR's metrics on GazeFollow, the disentanglement claim is falsified; if GazeDETR retains the margin, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that two disentangled decoders outperform single-decoder end-to-end models. A DETR decoder with its own query set adds substantial parameters; the abstract gives no indication that comparisons control for parameter count, training budget, or backbone. If the improvement disappears when a single decoder is given the same parameter budget (e.g., by increasing width/depth), the paper's contribution collapses to a capacity increase rather than the proposed disentanglement mechanism. The full text is corrupted, so no ablation table can be inspected to refute this. This is load-bearing because the paper's novelty is specifically the two-decoder design; if the margin is a capacity effect, the 'disentangled representation' argument is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents GazeDETR, an end-to-end architecture for gaze target detection that splits the task into head localization and gaze prediction using two disentangled Transformer decoders. The authors argue that this decomposition yields complementary attention patterns: local for head detection and local-plus-global for gaze prediction. The abstract claims state-of-the-art results on GazeFollow, VideoAttentionTarget, and ChildPlay, outperforming existing end-to-end models by a notable margin. However, the submitted full text is severely corrupted: most sections appear as mojibake, making equations, experimental tables, ablations, and implementation details unreadable. The only reliable content is the abstract and a handful of fragmentary phrases.","tokens_in":14538,"tokens_out":3040,"duration_ms":34725,"significance":"If the claimed results are verified, the core idea—decoupling head and gaze decoders to let each learn specialized representations—is a plausible and useful architectural contribution to gaze target detection. The claim of complementary attention fields is interesting and could inform future multitask DETR designs. The use of three public benchmarks provides independent grounding, and the paper addresses a practically important problem. Nevertheless, the significance cannot currently be assessed beyond this: the evidence needed to confirm the disentanglement claim and the reported margins is absent from the readable portion of the manuscript.","major_comments":[{"comment":"The full text of the manuscript is not legible: the overwhelming majority of the body is corrupted (mojibake), so no equations, experimental details, ablation tables, or implementation specifics can be inspected. The abstract is readable and states the main claims, but those claims are unverifiable in the current submission. This is a load-bearing issue: the paper's central contribution rests on quantitative comparisons and architectural analyses that are entirely unavailable to the reviewer.","section":"Full text (submitted PDF)"},{"comment":"The novel element of GazeDETR is the use of two separate decoders. The abstract claims a 'notable margin' over single-decoder end-to-end models, but it does not indicate whether comparisons control for model capacity, parameter count, FLOPs, or training budget. A single decoder with additional layers or width could plausibly match or exceed the proposed two-decoder design. Without a capacity-matched ablation, the reported gains cannot be attributed to disentanglement rather than to extra parameters. This concern is central because the paper's stated contribution is specifically the two-decoder mechanism.","section":"Abstract and architecture description"},{"comment":"The paper's secondary claim is that the head predictor uses local information while the gaze decoder uses both local and global information. The abstract states this as a finding, but no methodology or evidence is visible. How was this determined? Attention map visualization, receptive-field analysis, or a controlled experiment? Without such evidence, this claim is not supported. This analysis is part of the 'disentangled representations' thesis and must be substantiated with quantitative or at least systematic qualitative evidence.","section":"Attention-field analysis (visible in abstract)"},{"comment":"No experimental setup is visible: there are no tables, metrics definitions, data split details, backbone choices, number of queries, loss functions, or error bars. Even the abstract does not list concrete numeric results. The reader cannot assess whether the reported performance is statistically meaningful, whether the comparison protocols match those of prior work, or whether the results are reproducible. All these elements must be present for the SOTA claim to be evaluated.","section":"Evaluation methodology (unavailable)"}],"minor_comments":[{"comment":"The manuscript has severe encoding corruption in the PDF text layer; the authors should resubmit a clean and readable version. The repeated fragments in the visible portions also suggest possible duplication or layout errors.","section":"Production quality"},{"comment":"The abstract would be more informative with concrete numbers (e.g., AUC or AP) or at least explicit mention of the evaluation metrics used, so readers can gauge the 'notable margin' without accessing the full text.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The submitted full text is unreadable, likely due to an encoding or PDF extraction failure. I could not review the methods or experiments beyond the abstract. This is not a rejection on scientific grounds; it is a request for a clean, complete manuscript. The core idea has merit, but the proof lies in the full experimental section and capacity-matched ablations, which must be provided before the paper can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague—\n\nThe only thing worth knowing about this paper is that its central idea is clean but its evidence is, in the copy we have, invisible. GazeDETR replaces the usual single-decoder gaze-detection head with two disentangled decoders—one for head location, one for gaze—and claims SOTA on GazeFollow, VideoAttentionTarget, and ChildPlay. The abstract also reports an interpretable split: the head decoder attends locally, the gaze decoder attends both locally and globally. That is a sensible, testable architecture and a real departure from the single-decoder end-to-end models it compares against.\n\nWhat the paper does well on the evidence we have: the problem is worth addressing, the three benchmarks are standard, and the disentanglement claim is concrete enough to be falsified. No circularity is apparent; it is benchmarked against external datasets.\n\nThe soft spots are substantial, though some may be artifacts of the pipeline. The full text we received is mojibake; I cannot verify any ablation, table, equation, or protocol. So the central claim—“notable margin”—is taken entirely on the abstract's word. The second issue is the one the stress test flagged and it is legitimate: a second DETR decoder adds nontrivial parameter count and query capacity. The abstract gives no indication that the single-decoder baselines were matched in parameters, training budget, or backbone. If the margin collapses under a capacity-matched single decoder, the “disentangled representation” explanation is unsupported. That is a load-bearing omission, not a cosmetic one. There are also no error bars or per-dataset margins in the abstract, so we do not know how “notable” the margin actually is.\n\nIf the clean PDF is available, this deserves a proper referee. The hypothesis is good and the benchmarks are appropriate; the referee should demand a capacity-matched ablation, per-dataset numbers with variance, and ideally a visualization that supports the local-vs-global attention claim. Without the readable text, I would not desk reject the idea, but I also would not trust the SOTA claim yet.\n\nMy bottom line: worth serious peer review if the manuscript is intact; the current corrupted copy cannot be evaluated. I would not cite it until I see the ablations, but I would bring it to a reading group to discuss the decoder-disentanglement idea.\n\nRecommended: send to review.","headline":"Crisp architectural idea, but the copy we have is unreadable past the abstract, and the capacity-matched ablation is the make-or-break question.","tokens_in":14870,"tokens_out":2823,"would_cite":false,"duration_ms":32701,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"GazeDETR proposes that separating head localization and gaze prediction into two disentangled decoder branches, with local versus local-plus-global attention, yields state-of-the-art gaze target detection on three benchmarks.","keywords":["gaze target detection","gaze following","disentangled decoders","transformer attention","head localization","end-to-end learning","multi-task learning"],"falsifier":"Train the identical architecture but replace the two decoders with one shared decoder of the same total parameter count and training schedule on GazeFollow. If the accuracy of the shared-decoder version equals GazeDETR's, the disentanglement claim is not supported.","tokens_in":14356,"feed_emoji":"👀","tokens_out":4902,"duration_ms":45512,"temperature":0.7,"pith_summary":"This paper argues that gaze target detection improves when head localization and gaze prediction are pulled apart into two separate decoder branches instead of being generated by one shared multitask decoder. The proposed architecture, GazeDETR, gives the head decoder a local attention field and gives the gaze decoder both local and global attention, so each representation fits its own subtask. If this claim holds, end-to-end gaze detectors can become more accurate without additional supervision, which matters for human-computer interaction and digital phenotyping. The paper reports state-of-the-art results on GazeFollow, VideoAttentionTarget, and ChildPlay, beating earlier end-to-end models by a notable margin.","feed_headline":"Two separate decoders push gaze detection past prior end-to-end models","feed_subtitle":"GazeDETR splits head and gaze prediction into two decoder branches and reports state-of-the-art on three gaze datasets.","key_machinery":"The central mechanism is the split of the multitask prediction into two decoder branches, each with its own learned attention field: a head decoder that attends to local cues for locating heads, and a gaze decoder that fuses local and global scene cues for predicting the looked-at point. The disentanglement lets each branch avoid the compromise of a single entangled representation.","core_discovery":"GazeDETR is an end-to-end model that detects human heads and predicts where each person is looking using two disentangled transformer decoders rather than the single decoder used in prior end-to-end gaze target detection. The head-decoder branch is shown to rely on local visual information, while the gaze-decoder branch combines local and global information to locate the gaze point. The paper claims this disentanglement produces representations that are coherent with each subtask, yielding state-of-the-art performance on GazeFollow, VideoAttentionTarget, and ChildPlay and a notable margin over existing end-to-end models.","pith_inferences":["The same two-decoder design could benefit other joint tasks where one output is local and the other needs context, such as object detection plus future trajectory prediction.","A direct stress-test would be to train the single-decoder baseline under identical random seeds and schedules; the abstract does not report such an ablation.","Because the head decoder is independent, gaze-following models could be transferred to new scenes using only weak head-location labels, with gaze predictions adapted later.","Visualizing the attention fields of the two decoders could provide an interpretability tool for debugging gaze errors in social scenes."],"forward_implications":["Because head and gaze branches are disentangled, the head branch can be used alone for head localization without retraining.","The gaze branch's local-plus-global attention makes gaze following less dependent on the head crop only.","A single end-to-end forward pass yields both head locations and gaze targets, simplifying deployment for interaction and phenotyping tools.","The reported margin over end-to-end baselines suggests that shared multitask decoders were limiting accuracy; separate decoders lift the constraint."],"supporting_citations":[],"fun_headline_variants":["GazeDETR splits decoders to beat prior gaze models","GazeDETR: separate head and gaze decoders set SOTA","Head and gaze decoders go separate ways in GazeDETR","Two decoders beat one: GazeDETR outperforms prior","Disentangled decoders improve gaze target detection"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The reported gains come from the disentanglement of the two decoders, not from additional model capacity, different hyperparameters, or evaluation settings.","fun_headline_variants_meta":{"raw":{"variants":["GazeDETR splits decoders to beat prior gaze models","GazeDETR: separate head and gaze decoders set SOTA","Head and gaze decoders go separate ways in GazeDETR","Two decoders beat one: GazeDETR outperforms prior","Disentangled decoders improve gaze target detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000913,"raw_usage":{"total_tokens":3725,"prompt_tokens":677,"completion_tokens":3048,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":421,"completion_tokens_details":{"reasoning_tokens":2957}},"tokens_in":421,"tokens_out":3048,"duration_ms":21670,"temperature":1.0,"reasoning_tokens":2957,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T19:09:52.639088+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the identical architecture but replace the two decoders with one shared decoder of the same total parameter count and training schedule on GazeFollow. If the accuracy of the shared-decoder version equals GazeDETR's, the disentanglement claim is not supported.","supporting_citations":[],"review_version":1}