{"id":"4bf2e5cb-7e99-4950-b030-b30367b0e8eb","arxiv_id":"2411.09180","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"LEAP:D swaps manual prompts for 8 learnable tokens in a CLIP-guided Faster R-CNN, improving mAP50 from 39.7 to 42.1 on VisDrone validation, with no cross-domain evaluation.","lead":"This paper replaces hand-written text prompts with learnable prompt vectors trained jointly with a CLIP-based aerial object detector, reporting small accuracy gains on the VisDrone validation set. The method is called domain-generalized, but only in-domain results are presented, so the generalization claim is untested.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The domain-generalization claim is unsupported: LEAP:D is evaluated only on VisDrone validation, same data distribution as training, and the learned-prompt objective is never tied to an actual domain shift.","rationale":"The reader's weakest_assumption correctly identifies the core problem: the method is only evaluated on VisDrone validation, which is within the training distribution, so the domain-generalization claim is not demonstrated. This is the single most load-bearing concern because the paper's title, abstract, and conclusion all assert domain generalization, and the proposed method is explicitly framed as an improvement over manual prompts for handling diverse shooting conditions. If the only evaluation is in-domain, then even a perfect result on VisDrone validation cannot support the central claim. The paper's own Section 4 admits the single-dataset limitation, and per the review rules this admission must be weighted as evidence against the claim. My analysis also points to a second, more technical weakness: the training objective in Eqs. (7)-(10) has no explicit tie to real domain factors. The learnable prompt is a single vector trained on one dataset, and the dissimilarity loss merely pushes features away from that vector in embedding space. There is no mechanism ensuring that what is removed is domain-specific variation rather than dataset-specific bias. This reinforces the experimental concern rather than replacing it. Because the central claim requires evidence of generalization to unseen aerial domains and no such evidence is provided, the appropriate verdict is REJECT. I agree with the reader that the same-distribution evaluation is the decisive flaw; the mechanism issue is supplementary. No ad hominem is intended; the critique is on the argument's evidence and internal logic.","tokens_in":5510,"tokens_out":2608,"duration_ms":34674,"concrete_test":"Take the trained LEAP:D model and, without any fine-tuning, evaluate it on an unseen aerial object detection dataset such as UAVDT or DOTA, comparing mAP50 against Faster R-CNN and LGNet trained on VisDrone only. Alternatively, split VisDrone into disjoint source and target subsets according to altitude/weather/view metadata (or use a held-out collection of drone scenes) and retrain under the current protocol. If LEAP:D does not outperform the baselines by a meaningful margin on those unseen target distributions, the domain-generalization claim is unsupported.","verdict_should_be":"REJECT","load_bearing_attack":"The paper's central claim is domain-generalized aerial object detection, but the only quantitative evidence is mAP on the VisDrone validation set (Section 3.2, Table 1). VisDrone train and validation are from the same dataset and the same underlying distribution, so this evaluates in-domain performance, not generalization to unseen domains. Section 4 explicitly admits that 'the experiments were limited to a single dataset,' and this stated limitation directly undercuts the title, abstract, and conclusion. In addition, the mechanism itself does not establish domain separation: the learnable prompt t_lp is a single vector optimized on VisDrone training data, and Lds in Eq. (10) simply pushes the FSN-aligned feature f' toward negative cosine similarity with t_lp. There are no domain labels, no partitioning of training data into source/target environments, and no evidence that this push removes altitude/view/weather-specific nuisances rather than arbitrary dataset-specific information. The reported gains of 0.7-2.4 percentage points over LGNet on the same-distribution validation set, without error bars or multiple seeds, do not support the claim of robustness across diverse environments. The load-bearing premise that similarity to CLIP's visual embedding is domain-invariant while dissimilarity to one learned prompt is domain-specific is assumed, not tested.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LEAP:D, a learnable-prompt variant of the LGNet detector for aerial images, replacing manually defined prompts with n=8 learnable tokens and training the prompt jointly with the object detector via CLIP-based auxiliary losses. The method is evaluated on the VisDrone validation set, reporting mAP50/75/50:95 of 42.1/25.5/24.8, which is 0.7-2.4 points above the LGNet baseline. An ablation varies the number of learnable prompts. The paper concludes that LEAP:D is a domain-generalized detector, while Section 4 acknowledges that experiments were limited to one dataset.","tokens_in":5749,"tokens_out":7240,"duration_ms":74829,"significance":"If the method genuinely generalized across shooting conditions, the use of learnable prompts would be a practical simplification over manual prompt engineering in LGNet, and the one-step training procedure is a reasonable efficiency contribution. The ablation over prompt count is also a useful starting point. However, the experimental evidence is entirely in-domain: all results come from the VisDrone validation set, which shares the distribution of the training set, and no cross-dataset or domain-shift evaluation is presented. The internal inconsistency in the reported baseline numbers further weakens the quantitative claims. The paper's central claims of domain generalization and robustness across diverse environments are therefore not currently substantiated.","major_comments":[{"comment":"The evaluation is entirely on the VisDrone validation set, which is sampled from the same dataset and distribution as the training set. The manuscript's own Section 4 states that 'the experiments were limited to a single dataset,' yet the Abstract claims the method 'enhances model robustness and adaptability across diverse environments' and the title claims domain-generalized detection. These claims are unsupported by the experimental design; the reported mAP gains over LGNet are same-distribution improvements and cannot be interpreted as evidence of domain generalization.","section":"§3.1–3.2, §4"},{"comment":"The core assumption that s = (1+sim(v,f'))/2 captures domain-invariant features while ds = (1−sim(t_lp,f'))/2 captures domain-specific features is offered without domain labels or a domain-shift experiment. Since t_lp is a single learnable vector optimized on VisDrone training data, L_ds in Eq. (10) can at best encourage dissimilarity to a training-distribution-specific direction; there is no evidence that this removes variations in altitude, angle, weather, or other factors in unseen domains. A concrete test would be to train on one aerial dataset or on a subset excluding certain conditions and evaluate on another, showing that LEAP:D's advantage persists under domain shift.","section":"§2.2, Eqs. (7)–(10)"},{"comment":"Table 1 lists LGNet with mAP75 = 24.8, while Table 2 lists manual(32) with mAP75 = 25.5; if manual(32) is the same configuration as LGNet, the claimed +0.7%p improvement in mAP75 is contradicted. Even if they differ, Section 3.4 calls manual(32) 'the baseline network,' which is ambiguous. In addition, no error bars, standard deviations, or multiple-seed runs are reported, so the small differences (0.7–2.4 points) are not established as statistically meaningful.","section":"Table 1 vs. Table 2, §3.2, §3.4"},{"comment":"The number of learnable prompts n=8 and the loss weights λ1..λ4 in Eq. (11) are set without a sensitivity analysis; the ablation in Section 3.4 varies only the prompt count, not the weights. The claimed efficiency of 'one-step training' is also not quantified, as no training-time or convergence comparison with LGNet's two-step procedure is provided. These omissions do not by themselves invalidate the method, but they leave the reported results dependent on unexamined choices.","section":"§2.3, Eq. (11), §3.4"}],"minor_comments":[{"comment":"The probability expressions are malformed: the denominator is missing or typeset as a stray 'P' in Eq. (3), and Eq. (6) has the same formatting issue. Please correct the typesetting.","section":"Eqs. (3) and (6)"},{"comment":"The text contains a typo: 'resembe' should be 'resemble'. Also, the phrase 'the similarity to t, generated through domain-specific prompts' is imprecise, since t here is one single learnable prompt rather than a set of domain-specific prompts.","section":"§2.2"},{"comment":"The column headers use inconsistent spacing ('mAP 50' vs 'mAP50' vs 'mAP 50:95'); please standardize for clarity.","section":"Table 2"},{"comment":"Reference [15] cites an overview of gradient descent optimization algorithms rather than a primary source for SGD; a more targeted citation would be appropriate.","section":"References"}],"recommendation":"reject","confidential_remarks":"The paper's claims clearly exceed its evidence. The single-dataset evaluation is acknowledged as a limitation but is still used to support domain-generalization claims in the title, abstract, and conclusion. The inconsistency between Tables 1 and 2 for the same baseline configuration is a red flag that the numbers may not be reliable. The most constructive path for the authors would be to either reframe the paper as an in-domain improvement with carefully reported uncertainty, or conduct a proper cross-domain evaluation (e.g., train on VisDrone and test on another aerial dataset, with multiple seeds and error bars) before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"LEAP:D is a reasonable engineering contribution: it replaces LGNet's manual prompts with CoOp-style learnable prompt vectors and collapses the two-step training into one joint update. That simplification is legitimate and the ablation over prompt counts (4–32) is a nice touch. The writing is clear, and the authors honestly admit in the limitations section that experiments were limited to a single dataset. Credit where due: on its own terms, the method is coherent and easy to reconstruct from the text.\n\nThe soft spots are serious, not minor. The title and abstract claim domain generalization across altitude, angle, and weather, but every quantitative result comes from the VisDrone validation set, which shares the training distribution. There is no cross-domain evaluation: no other aerial dataset, no synthetic-to-real shift, no split by metadata. The mechanism in Eq. (10) pushes the aligned feature away from one learned text vector, but there is no evidence this removes domain-specific nuisance rather than arbitrary dataset-specific information. The reported gains are 0.7–2.4 mAP points over LGNet on the same distribution, with no error bars and no multiple seeds, so calling them \"significant performance improvements\" and \"superior generalization\" is overclaiming. There is also a small technical gap: N_sc in Eq. (10) is never defined.\n\nNone of this makes the paper incoherent. The method is plausible and the experiments, while insufficient for the claim, are clean and reproducible in structure. But the current evidence supports an in-domain accuracy claim, not a domain-generalization one. If the authors added a proper cross-domain evaluation, the paper would be substantially stronger.\n\nWho this is for: researchers working on prompt-based aerial detection who want a baseline, and anyone teaching a lesson on why domain-generalization claims need actual domain shift in the evaluation. It deserves a serious referee because the method is real and the flaw is fixable, but I would reject the current version. If it comes back with cross-domain results, re-evaluate.","headline":"A plausible learnable-prompt variant of LGNet, but the domain-generalization claim is unsupported by a same-distribution VisDrone-only evaluation.","tokens_in":6286,"tokens_out":2337,"would_cite":false,"duration_ms":25084,"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":"LEAP:D claims that replacing hand-written drone prompts with eight learnable tokens improves domain-generalized aerial object detection on the VisDrone validation set.","keywords":["aerial object detection","domain generalization","learnable prompts","vision-language model","CLIP","drone imagery","VisDrone","feature disentanglement"],"falsifier":"Train LEAP:D and the manual-prompt baseline on VisDrone and evaluate both on an unseen aerial dataset collected under different cities, seasons, altitudes, or weather; if LEAP:D does not exceed the baseline's mAP50 under that domain shift, the claimed domain-invariant separation is not doing the work. A second check is to remove the domain-invariant and domain-specific losses from the total loss and re-run: if mAP50 stays at 42.1, the improvement comes from the prompt update alone rather than the feature-separation mechanism.","tokens_in":5300,"feed_emoji":"🚁","tokens_out":8011,"duration_ms":80721,"temperature":0.7,"pith_summary":"This paper proposes LEAP:D, a method for detecting objects in drone images when shooting conditions vary. It replaces the hand-written prompt templates used by a prior vision-language detector with eight learnable prompt tokens, and trains those tokens together with the detector in a single pass instead of a two-stage pipeline. The authors claim this removes domain-specific knowledge contained in fixed prompts and lets the vision-language feature alignment keep domain-invariant features, yielding mAP50 42.1, mAP75 25.5, and mAP50:95 24.8 on the VisDrone validation set, above Faster R-CNN, Cascade R-CNN, and their manual-prompt baseline. A sympathetic reader would take the contribution to be evidence that learnable prompts can substitute for manual prompt engineering in aerial domain generalization.","feed_headline":"Eight learnable prompt tokens lift aerial detection accuracy","feed_subtitle":"One-step training on VisDrone beats the manual-prompt baseline on mAP50, mAP75, and mAP50:95.","key_machinery":"The central object is the learnable prompt $x_{lp} = [e_1]\\ldots[e_8]$, a sequence of eight trainable token embeddings passed through the text encoder of a vision-language model to produce $t_{lp} = G(x_{lp})$. The argument is carried by a contrastive alignment between the aligned detection feature $f' = FSN(f)$, the image embedding $v = F(x)$, and the prompt embedding $t_{lp}$: the similarity score $s = (1 + \\mathrm{sim}(v, f'))/2$ and dissimilarity score $ds = (1 - \\mathrm{sim}(t_{lp}, f'))/2$ feed domain-invariant and domain-specific losses, and the total loss combines these with the detection loss and a prompt loss. This machinery lets a fixed template like 'altitude/view/weather' be replaced by embeddings learned at the same time as the detector, so the claimed domain separation is driven by similarity structure rather than by hand-labeled shooting conditions.","core_discovery":"On its own terms, the paper's central claim is that domain generalization for aerial object detection can be achieved by replacing manual prompts with a learnable prompt that is optimized jointly with the detector. The method uses Faster R-CNN with FPN for detection, a vision-language model with separate visual and textual encoders to supply embeddings, and an alignment network so that intermediate detection features can be compared with those embeddings. The trainable prompt's textual embedding $t_{lp}$ plays double duty: the detector's aligned feature $f'$ is encouraged to be similar to the image embedding $v$ (domain-invariant content) and dissimilar to $t_{lp}$ (domain-specific content). The authors report that this one-step approach improves over the two-step manual-prompt baseline by 2.4, 0.7, and 1.1 percentage points in mAP50, mAP75, and mAP50:95, respectively, on VisDrone validation, and they interpret the gain as better filtering of diverse shooting conditions.","pith_inferences":["A direct test of the domain-separation mechanism would be to inspect which visual patterns the learned prompt embedding repels; if the prompt merely encodes dataset-level brightness or scale statistics, the method may be a dataset-specific shortcut rather than a general domain filter.","The reported single-dataset evaluation leaves open whether the same margins hold under genuine domain shift; a cross-dataset evaluation, such as training on VisDrone and testing on a different aerial benchmark, would be the natural next experiment.","If the one-step learnable prompt generalizes, the same recipe could be applied to other base detectors and to per-object prompting, the object-level extension the authors list as future work, allowing domain filtering at instance granularity rather than scene granularity.","Because the vision-language model is only needed during training, LEAP:D could be viewed as a way to distill coarse vision-language knowledge into a conventional detector backbone without adding inference cost."],"forward_implications":["Aerial detectors can be adapted to variable shooting conditions without hand-crafted prompt templates; the eight-token learnable prompt reaches the reported mAP gains while training in one step.","Because the prompt and the detector are trained jointly, the two-stage pipeline of fine-tuning text embeddings before detection training is no longer necessary.","The learned alignment is used only during training, so the trained detector can run without the vision-language model at inference time, preserving the deployment characteristics of the base detector.","The ablation suggests performance is relatively stable across prompt lengths in the 4 to 32 token range, with 8 tokens giving the best reported mAP50 and mAP50:95, making prompt length a cheap hyperparameter rather than a fixed design constraint."],"supporting_citations":[{"why":"Supplies the baseline architecture and the two-step manual-prompt training that LEAP:D replaces.","marker":"[11]"},{"why":"Provides the detector whose loss contributes to the total objective.","marker":"[12]"},{"why":"Provides the feature pyramid features that the alignment network maps to the vision-language embedding space.","marker":"[13]"},{"why":"Supplies the visual and textual encoders used to define the image embedding, prompt embedding, and similarity losses.","marker":"[8]"},{"why":"Provides the VisDrone dataset used for training and validation in all quantitative and ablation experiments.","marker":"[14]"},{"why":"Supplies the learnable-prompt concept that motivates replacing manual prompts with trainable tokens.","marker":"[7]"}],"fun_headline_variants":["Learnable prompts outperform manual prompts in aerial detection","One-step prompt training boosts aerial object detection","LEAP:D uses learnable prompts to generalize aerial detection","Aerial detection gains from learnable prompt tuning","Domain-generalized aerial detection via learnable prompts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the similarity between the aligned detection feature and the image embedding truly captures domain-invariant information, while dissimilarity to the learned prompt embedding captures domain-specific information, and that this separation transfers to unseen drone domains; the paper itself notes that its experiments were limited to a single dataset, the VisDrone validation split, which comes from the same distribution as its training set.","fun_headline_variants_meta":{"raw":{"variants":["Learnable prompts outperform manual prompts in aerial detection","One-step prompt training boosts aerial object detection","LEAP:D uses learnable prompts to generalize aerial detection","Aerial detection gains from learnable prompt tuning","Domain-generalized aerial detection via learnable prompts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000168,"raw_usage":{"total_tokens":1227,"prompt_tokens":880,"completion_tokens":347,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":496,"completion_tokens_details":{"reasoning_tokens":275}},"tokens_in":496,"tokens_out":347,"duration_ms":3762,"temperature":1.0,"reasoning_tokens":275,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:55:41.585444+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train LEAP:D and the manual-prompt baseline on VisDrone and evaluate both on an unseen aerial dataset collected under different cities, seasons, altitudes, or weather; if LEAP:D does not exceed the baseline's mAP50 under that domain shift, the claimed domain-invariant separation is not doing the work. A second check is to remove the domain-invariant and domain-specific losses from the total loss and re-run: if mAP50 stays at 42.1, the improvement comes from the prompt update alone rather than the feature-separation mechanism.","supporting_citations":[{"cited_title":"Domain feature decomposition for efficient object de- tection in aerial images,","cited_arxiv_id":null,"evidence_quote":"Supplies the baseline architecture and the two-step manual-prompt training that LEAP:D replaces."},{"cited_title":"Learning to prompt for vision-language models,","cited_arxiv_id":null,"evidence_quote":"Provides the detector whose loss contributes to the total objective."},{"cited_title":"Delving into robust object detection from unmanned aerial ve- hicles: A deep nuisance disentanglement approach,","cited_arxiv_id":null,"evidence_quote":"Supplies the visual and textual encoders used to define the image embedding, prompt embedding, and similarity losses."},{"cited_title":"Conditional prompt learning for vision- language models,","cited_arxiv_id":null,"evidence_quote":"Provides the VisDrone dataset used for training and validation in all quantitative and ablation experiments."},{"cited_title":"Enhanced detection of small objects in aerial imagery: A high-resolution neural network ap- proach with amplified feature pyramid and sigmoid re- weighting,","cited_arxiv_id":null,"evidence_quote":"Supplies the learnable-prompt concept that motivates replacing manual prompts with trainable tokens."}],"review_version":1}