{"id":"642edc04-97ad-44c9-bb50-fb04f4c3e7f9","arxiv_id":"2607.08004","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Prompt-modulated content queries let a transformer detector focus on text-specified categories and outperform prior oriented detectors on DOTA.","lead":"LOGOS is a DETR-style detector that conditions oriented bounding-box queries on text prompts via FiLM modulation and text-aware cross-attention. On DOTA aerial images it reports higher mAP than prior oriented detectors, especially for dense or rotated objects.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"SOTA claim rests on an unablated closed-set perfect-prompt protocol that is not compared against a matched non-language baseline.","rationale":"The reader correctly isolates the perfect-prompt assumption of Sec. 5.1 as the weakest link. That assumption is not merely an implementation detail; it is the only setting in which the FiLM + class-mask machinery can produce the numbers that underwrite the SOTA claim. Because no matched non-language baseline or open-vocabulary experiment is provided, the paper cannot yet demonstrate that language guidance itself is responsible for the gains. The concern therefore leaves the verdict unchanged at CONDITIONAL: the engineering is competent and the numbers are real under the stated protocol, but the causal contribution of the language component remains unproven until the ablation above (or an equivalent open-vocabulary test) is run. No stronger objection (e.g., internal inconsistency or numerical error) is visible in the manuscript.","tokens_in":15075,"tokens_out":568,"duration_ms":5938,"concrete_test":"Train and evaluate an otherwise identical model on DOTA-v1.0 in which FiLM is replaced by a constant (or identity) modulation and the class mask of Eq. 4 is disabled (full C-way classification). If the mAP drops by less than ~1–2 points relative to the reported 81.32, the language-conditioning claim is not supported by the numbers; if it drops substantially, the claim holds under the closed-set protocol.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim (SOTA mAP of 81.32/69.97/66.04 on DOTA-v1.0/1.5/2.0 via prompt-modulated queries + text-aware cross-attention) is load-bearing on the protocol of Sec. 5.1: the textual prompt is exactly the set of ground-truth categories present in the image, which is then used both for FiLM modulation of content queries (Eq. 2) and for the hard class mask that zeros out all other logits (Eq. 4). Under this protocol the model never has to handle missing or extraneous labels, and every competing method in Tables 1–3 is a pure visual detector that receives no category list. Consequently it is impossible to tell whether the reported gains come from language guidance or simply from an oracle class filter plus a stronger DINO-style decoder. The paper supplies no ablation that (a) removes the prompt/FiLM/mask while keeping the same decoder and training recipe, or (b) supplies incomplete/open-vocabulary prompts. Without that control the causal link between the language mechanism and the SOTA numbers remains unestablished.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes LOGOS, a DETR/DINO-style transformer for oriented object detection in aerial imagery that conditions detection on a textual prompt. Content queries are modulated via FiLM using a pooled prompt embedding (Eq. 2), text-aware multi-head cross-attention attends jointly to visual and text tokens (Eq. 3), angles are encoded with sine/cosine, and a hard class mask derived from the prompt zeros out irrelevant logits (Eq. 4). The model is trained with a standard DETR-style combination of classification, L1, RotGIoU and contrastive denoising losses, and is evaluated on DOTA-v1.0/1.5/2.0 under a ResNet-50 backbone, reporting mAP of 81.32 %, 69.97 % and 66.04 % respectively and claiming state-of-the-art performance especially in dense/rotated scenes.","tokens_in":15378,"tokens_out":1018,"duration_ms":12868,"significance":"If the language-guidance mechanism is genuinely responsible for the gains, the work would constitute a useful step toward open-vocabulary or promptable oriented detection in remote sensing, an application domain where category lists are often known a priori and where angular discontinuity and density variation remain open problems. The numerical results under a common backbone are competitive with recent published detectors, the architecture is described with explicit equations, and the sine-cosine angle encoding plus RotGIoU matching are sensible engineering choices. The absence of any isolation of the language components, however, currently prevents the community from assessing whether the contribution is the claimed prompt modulation or simply a stronger decoder plus an oracle class filter.","major_comments":[{"comment":"Sec. 5.1 Setup and Eqs. (2)–(4): the evaluation protocol supplies, at both train and test time, a prompt that enumerates exactly the ground-truth categories present in the image; this prompt is used both for FiLM modulation of the content queries and for the hard class mask that sets all other logits to −∞. Every baseline in Tables 1–3 is a pure visual detector that receives no category list. Consequently the reported mAP gains cannot be attributed to language guidance versus an oracle class filter plus a DINO-style decoder. A matched ablation that (a) removes FiLM/mask while keeping the identical decoder and training recipe, and (b) evaluates incomplete or open-vocabulary prompts, is required to establish the central claim.","section":null},{"comment":"Tables 1–3: no error bars, no multiple random seeds, and no statistical significance test are provided for the 0.45–8.76-point mAP margins over the strongest prior methods. Given that early-stopping, learning-rate schedule and the precise 900-query configuration are free parameters (Sec. 5.1), the numerical superiority is currently under-supported.","section":null},{"comment":"Sec. 4.1–4.3: the paper asserts that prompt-modulated queries solve the fixed-query-size problem of DETR-style models, yet the implementation still uses a fixed 900 decoder queries (explicitly matched to DN-DETR). No experiment varies query count with or without the prompt, so the claimed advantage remains untested.","section":null}],"minor_comments":[{"comment":"Keywords and abstract contain the typo “Vision-Lanague Models”; correct to “Language”.","section":null},{"comment":"Fig. 2 caption and several figure references say “Best viewed in color and zoomed in” but the PDF does not embed high-resolution versions; supply vector or higher-dpi figures.","section":null},{"comment":"Eq. (6) uses RotGIoU while the surrounding text sometimes writes “GIoU for oriented bounding boxes”; make the notation consistent.","section":null},{"comment":"Sec. 2.3 claims that earlier methods are limited to a 0–90° angular range; several cited works (e.g., CSL, GWD) already address full-periodicity; the discussion should be more precise.","section":null},{"comment":"Table 2 reports a “CC” category that is never defined in the text or in the DOTA-v1.5 description given in Sec. 5.1; clarify.","section":null}],"recommendation":"major_revision","confidential_remarks":"The core novelty claim rests on an evaluation protocol that is unusually favorable to any language-conditioned model and is not disclosed as a limitation. Once the missing ablations are supplied the paper may still be publishable, but without them the SOTA numbers are not yet scientifically interpretable. Scope is appropriate for a remote-sensing / CV journal."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing to know is that LOGOS posts the highest mAP numbers I have seen on the three DOTA versions (81.32 / 69.97 / 66.04) with a ResNet-50 backbone, and the architecture is a clean marriage of DINO-style decoding, FiLM-modulated content queries, and a hard class mask derived from the text prompt. That combination is new for oriented aerial detection and the tables are thorough.\n\nWhat the paper does well: it correctly identifies the fixed-query and angular-discontinuity problems that plague earlier DETR-style oriented detectors, encodes angle with sine-cosine, uses RotGIoU + contrastive denoising, and shows clear qualitative wins in harbors and dense vehicle packs. The equations for FiLM (Eq. 2), text-aware MHCA (Eq. 3), and the class mask (Eq. 4) are explicit, and the citation pattern covers the relevant DETR and oriented-detection literature without obvious gaps.\n\nThe soft spot is real and load-bearing. Section 5.1 states that the prompt is exactly the set of ground-truth categories present in the image; that list is then used both to modulate the queries and to zero out every other class logit. No ablation removes the language path while keeping the same decoder and training recipe, and no experiment uses incomplete or open-vocabulary prompts. Every competitor in Tables 1–3 is a pure visual detector that never sees a category list. So we cannot yet separate “language guidance helps” from “oracle class filter + stronger DINO decoder helps.” That is the stress-test concern, and it holds up on a full reading. Minor additional gaps: no multi-run variance, no code, and slightly weaker numbers on Ship and Baseball Court.\n\nThis is for people already working on oriented remote-sensing detectors who want a language-conditioned baseline. It is not a new principle, but it is a competent engineering result that deserves a serious referee. I would send it to peer review with a clear request for the missing ablations and an open-vocabulary stress test; if those hold, the numbers become much more convincing. Worth reading if you care about DOTA or language-guided aerial detection; otherwise skip.","headline":"Solid SOTA numbers on DOTA via language-conditioned oriented DETR, but the gains rest on an unablated perfect-prompt protocol that no pure-vision baseline receives.","tokens_in":15939,"tokens_out":559,"would_cite":false,"duration_ms":5843,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Text prompts that modulate transformer queries improve oriented object detection in aerial images.","keywords":["oriented object detection","aerial imagery","transformers","language guidance","prompt modulation","remote sensing","DOTA"],"falsifier":"Re-run the DOTA evaluation with prompts that omit some present categories or add absent ones; if mAP falls to or below the strongest non-language baselines, the reported gains depend on oracle-category prompts.","tokens_in":15957,"feed_emoji":"🛰️","tokens_out":756,"duration_ms":31763,"temperature":0.7,"pith_summary":"Detecting ships, vehicles, and buildings in satellite and drone imagery is hard because objects appear at arbitrary angles, pack densely, and sit in cluttered backgrounds. Prior detectors struggle with fixed numbers of queries, angular discontinuity, and ambiguity for square-like boxes. LOGOS conditions a transformer detector on a short textual prompt that names the categories present in the image. The prompt modulates the content queries and is attended to jointly with visual features, so the model focuses only on the named objects. On the DOTA benchmarks the method reports higher mean average precision than previous oriented detectors, with the largest gains in crowded harbors and rotated scenes.","feed_headline":"Text prompts lift aerial object detection to new highs","feed_subtitle":"Language-conditioned queries beat prior oriented detectors on DOTA, especially in crowded harbors.","key_machinery":"Prompt-modulated content queries: learnable queries are FiLM-conditioned by a pooled text embedding of the prompt, then attend jointly to visual tokens and text tokens; angle is encoded as (sin θ, cos θ) and a class mask zeros logits outside the prompt’s categories.","core_discovery":"Conditioning a DETR-style oriented detector on a textual prompt that lists the object categories present in an aerial image, via FiLM-modulated content queries and text-aware cross-attention, yields higher mean average precision than prior oriented detectors on DOTA-v1.0, v1.5 and v2.0, especially for densely packed and rotated objects.","pith_inferences":["If the prompt is allowed to be free-form natural language rather than an exact category list, the same architecture becomes a candidate open-vocabulary oriented detector for remote sensing.","The failure cases of extreme density and unusual orientations suggest that prompt modulation alone may still need denser multi-scale sampling or orientation-equivariant backbones.","Because the method already injects text into every decoder layer, adding multi-spectral or SAR tokens would be a natural next multi-modal extension."],"forward_implications":["Oriented detectors can trade fixed query budgets for language-conditioned query selection, reducing wasted computation on empty regions.","Angular discontinuity can be mitigated by (sin, cos) regression plus language-guided class masking without specialized angle classification heads.","The same prompt interface can be reused for sparse versus dense aerial scenes simply by changing the text, without retraining separate models.","Applications such as harbor monitoring or urban inventory can filter detections to user-specified categories at inference time."],"fun_headline_variants":["Text prompts guide oriented detection of aerial objects","Language-conditioned queries lift DOTA oriented detection","Prompt-modulated DETR beats priors on dense rotated objects","LOGOS conditions content queries on text for aerial scenes","FiLM and text-aware attention boost oriented aerial detectors"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The textual prompt given at both train and test time is exactly the set of ground-truth categories that appear in the image, so the class mask and query modulation never see missing or extra labels.","fun_headline_variants_meta":{"raw":{"variants":["Text prompts guide oriented detection of aerial objects","Language-conditioned queries lift DOTA oriented detection","Prompt-modulated DETR beats priors on dense rotated objects","LOGOS conditions content queries on text for aerial scenes","FiLM and text-aware attention boost oriented aerial detectors"]},"model":"grok-4.5","effort":"low","cost_usd":0.00385,"raw_usage":{"total_tokens":1179,"prompt_tokens":710,"num_sources_used":0,"completion_tokens":78,"cost_in_usd_ticks":38500000,"prompt_tokens_details":{"text_tokens":710,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":391,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":710,"tokens_out":78,"duration_ms":4341,"temperature":1.0,"reasoning_tokens":391,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T13:52:21.024958+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Re-run the DOTA evaluation with prompts that omit some present categories or add absent ones; if mAP falls to or below the strongest non-language baselines, the reported gains depend on oracle-category prompts.","supporting_citations":[],"review_version":1}