{"id":"241f9938-aecf-4ac2-95f4-19e90ecf16be","arxiv_id":"2608.08840","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"SLAP uses partial optimal transport to selectively align local image patches with identity-aware prompts, improving fish re-identification accuracy on the Melops dataset and two other marine benchmarks.","lead":"This paper introduces a method for fish re-identification that aligns local image patches with identity-specific text prompts using partial optimal transport, focusing on informative body regions instead of the whole image. The authors report small but consistent accuracy improvements on a longitudinal wild fish dataset and two other marine animal benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"POT cost matrix as written compares unequal-dimension vectors (768-d patch tokens vs 512-d text embeddings); the shared-space premise is unstated and unverified.","rationale":"The reader identified the shared-space assumption as the weakest point; stress-testing sharpens it into a concrete dimensional and architectural inconsistency. The paper's central empirical claim is otherwise supported by controlled five-seed comparisons and a clean ablation, so the concern is not that the results are fabricated. Rather, the mechanism claimed to explain the gains is underspecified in the exact place where it matters: the POT cost matrix. Without either a stated projection of patch tokens or evidence that local patch embeddings are aligned with text embeddings, the selective-locality interpretation is not verifiable from the manuscript. This does not warrant rejection, because the issue is addressable: an explicit projection plus a transport-plan visualization or region-level analysis could settle it. The verdict remains CONDITIONAL, consistent with the reader's assessment, with the condition made more precise.","tokens_in":13415,"tokens_out":5948,"duration_ms":67076,"concrete_test":"Re-implement the Section 3 POT loss with a stock pretrained CLIP ViT-B/16 and insert a shape assertion before computing C: assert V.shape[-1] == T_y.shape[-1]. In the public model these are 768 and 512, so the assertion fails unless an unstated projection is applied. If it fails, the paper must specify how patch tokens are mapped into the text embedding space; if it passes only after adding W_proj to each patch, run the same ablation (No CM vs Global vs OT vs POT) with and without that projection to determine whether the reported gains actually depend on the projected patch-text cosine geometry.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim attributes SLAP's gains to POT-based selective local alignment. For that mechanism to operate, the cost matrix C_ik = 1 - v_i^T t_k^y / (||v_i|| ||t_k^y||) must be a meaningful cosine distance between local visual patches and identity prompts. As described in Section 3, this is not well-defined: CLIP ViT-B/16 produces local patch embeddings V of dimension d_v = 768, while frozen text encoder prompt embeddings T_y have dimension d_t = 512. The paper explicitly projects only the global representation, f_proj_v = f_v W_proj, and never states that patch tokens are projected. Even if an unstated projection is applied, CLIP's shared vision-language space is learned only for the global [CLS] representation after the visual projection; local patch tokens are not contrastively aligned with text embeddings. Thus the assertion that V and T_y 'naturally reside in the shared vision-language embedding space' is not secured by CLIP pretraining. The POT transport plan may therefore be computed on incomparable or poorly calibrated vectors, so the claimed selective suppression of background and weakly discriminative regions is not established by the reported mAP gains alone.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SLAP, a CLIP-based framework for fish re-identification that replaces global image-text alignment with selective local alignment between visual patch embeddings and multiple identity-aware prompt embeddings, computed via Partial Optimal Transport (POT). The visual encoder is a frozen CLIP ViT-B/16 with LoRA adapters; learnable prompt context tokens adapt the frozen text encoder. The POT loss is trained jointly with triplet and classification losses, and at inference only the adapted image encoder is retained. Experiments on the Melops dataset under closed-set and open-set protocols report consistent improvements over several CLIP-based ReID baselines, with ablations showing a monotone progression from no cross-modal supervision to global alignment to full OT to partial OT. Additional experiments on SeaStarReID2023 and SeaTurtleID2022 report higher mAP than the same baselines.","tokens_in":13685,"tokens_out":5037,"duration_ms":57397,"significance":"If the reported results hold, the paper makes a useful contribution to animal re-identification by introducing a principled selective-alignment mechanism into vision-language ReID and by demonstrating its practical value on a challenging longitudinal fish dataset. The experimental methodology is a strength: five random seeds are used, all methods share the same backbone, data splits, optimizer, and preprocessing, and the central ablation in Table 2 isolates the alignment strategy. The parameter-efficient design (LoRA + prompt tuning, visual-only inference) is also well motivated. However, the core computational step as written is not well-defined because the cost matrix compares vectors of different dimensions, and the paper does not provide direct evidence that the transport plan actually selects identity-discriminative body regions rather than acting as an unstructured regularizer. These issues must be resolved before the central claims can be accepted.","major_comments":[{"comment":"The cost matrix C_ik = 1 - v_i^T t_k^y / (||v_i|| ||t_k^y||) is not well-defined as written. For CLIP ViT-B/16, the local patch embeddings V have dimension d_v = 768, while the prompt embeddings T_y have dimension d_t = 512. The paper explicitly projects only the global representation f_v via W_proj in R^{d_v x d_t} and never defines a projection for patch tokens. The sentence claiming that V and T_y 'naturally reside in the shared vision-language embedding space' is not sufficient, because CLIP's shared space is learned for the global [CLS] representation after the visual projection, not for raw patch tokens. This is a load-bearing issue: if the implementation actually projects patch tokens to d_t, that projection must be described (including its parameters and whether it is trainable), and if it does not, the reported POT-based training cannot be reproduced from the paper. Please clarify and, if necessary, re-run the experiments with the corrected formulation.","section":"Section 3, Partial Optimal Transport; Visual Feature and Prompt Encoding"},{"comment":"The paper attributes the gains of POT over OT and global alignment to selective suppression of background and weakly discriminative regions, but no direct evidence of selectivity is provided. The ablation in Table 2 shows monotone improvements, yet those improvements could also arise from a different effective loss landscape or from partial mass truncation acting as a regularizer. To support the mechanistic claim, the authors should provide an analysis of the learned transport plan, for example by visualizing average transport mass per image patch, comparing mass allocated to foreground versus background regions, or showing that the selected patches correspond to biologically plausible features such as the operculum, eye, and snout. Without such evidence, the title and central interpretation of the method overstate what the current experiments establish.","section":"Section 3 and Section 4, Table 2"},{"comment":"The open-set protocol description is underspecified in a way that affects interpretation of the results. The text states that approximately 20% of test identities are excluded from training while query and gallery contain both seen and unseen identities, but it does not report how many unseen identities are used, how their query and gallery images are balanced, or how mAP and Rank-1 are computed when an unseen identity may appear only once. Please provide these details, because the claimed generalization benefit of POT under the open-set protocol depends on the precise composition of the query and gallery sets.","section":"Section 4, Evaluation Protocols"}],"minor_comments":[{"comment":"The additional dataset results report only mAP without standard deviations or the number of seeds, and the text states improvements of +1.7 and +1.9 percentage points without statistical comparison; please report variability or clearly label these as single-run results.","section":"Section 4, Table 4"},{"comment":"The identity token is described both as 'fixed identity-specific token embedding' and as 'shared across all samples belonging to the same identity, remains fixed throughout training,' but the initialization and construction of this token are not specified; please clarify how the identity token is obtained and why it is not optimized.","section":"Section 3, Learnable Prompt-Based Text Representation"},{"comment":"The text states that the transported mass ratio was set to rho = 0.8 'based on an ablation study,' but no table or figure for that ablation appears in the main text or supplementary material; please include the full rho sensitivity results.","section":"Section 4, Implementation Details"},{"comment":"Several references are incomplete: reference [1] lacks author names, and references [4], [23], and [26] use 'et al.' without listing the full author list; please complete the bibliographic entries.","section":"References"},{"comment":"There are minor typographical and formatting issues, including a line break in the abstract ('demon-stratethegeneralizationcapability') and irregular spacing in a few places; a careful proofread would improve readability.","section":"Abstract and Section 1"}],"recommendation":"major_revision","confidential_remarks":"The closest baseline, MetaPrompt-ReID, is from the same research group, and all baselines appear to be re-implemented by the authors. This is not itself a flaw, but it increases the importance of providing code or detailed implementation instructions so that independent researchers can verify the comparisons. The dimensional inconsistency in the cost matrix is the most serious technical concern and should be addressed before any acceptance decision. If the authors can clarify that the patch tokens are projected to the text embedding dimension and provide transport-plan visualizations, the paper could become a solid contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Kirsten,\n\nShort version: the paper has a nicely controlled empirical comparison and a clean ablation, but the central technical object—the cosine cost between patch embeddings and prompt embeddings—is not well-defined as written. The patch tokens are 768-d, the text prompts are 512-d, and only the global representation is projected with W_proj. So v_i^T t_k is undefined unless there is an unstated projection of the patches. That is a load-bearing problem, not a cosmetic one. The claim that both sets 'naturally reside in the shared vision-language embedding space' is also not true for raw patch tokens in CLIP; the shared space is established after the visual projection head for the [CLS] token.\n\nWhat is actually new: applying partial optimal transport to local vision-language alignment for animal ReID, replacing global alignment. That is a sensible extension, and the empirical gains are consistent: +2.2 mAP / +1.8 Rank-1 on CS, +2.4 / +2.1 on OS over MetaPrompt-ReID, with five seeds and identical backbone/splits/optimizer. The ablation in Table 2 is clean: no cross-modal, global, OT, POT gives a monotone improvement. So the experimental discipline is solid.\n\nThe soft spots: first, the dimension issue. If the authors actually project patch tokens, they forgot to say so; that is a quick fix but essential. Second, the loss weights λ_tri and λ_pot are never given—couldn't reproduce without them. Third, ρ=0.8 is chosen via ablation on the same benchmark, so there is some risk of overfitting to the test set, though the effect seems moderate. Fourth, the extra datasets show gains of only ~1.7–1.9 mAP without significance tests; minor. Fifth, the 'selective suppression of background' claim is not directly verified—no transport plan visualization or patch-level analysis. That is a minor omission but would strengthen the paper.\n\nOverall: the empirical package is credible, but the mechanism as described is not. This is a 'major revision and re-review' situation, not a desk reject—the idea is worth pursuing and the experiments are carefully done. If the authors add the missing projection (or correct the dimensions) and report the loss weights, I'd be comfortable with it going forward.\n\nRecommendation: send to peer review, but flag the dimension issue prominently to the reviewers.","headline":"Clean experiments wrapped around a cost matrix that is currently undefined; the POT mechanism is not established without a patch projection or dimension fix.","tokens_in":14205,"tokens_out":3288,"would_cite":false,"duration_ms":31163,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Selective local alignment, not global image-text matching, lifts fish re-ID accuracy.","keywords":["fish re-identification","partial optimal transport","vision-language alignment","CLIP adaptation","LoRA","prompt learning","marine animal re-identification","open-set retrieval"],"falsifier":"Inspect the learned transport plan on Melops test images: if the transported mass does not concentrate on identity-relevant regions such as the operculum, eye, and snout, or if replacing POT with uniformly random patch selection of equal mass produces no drop in mAP, then the selective-alignment explanation is not supported.","tokens_in":13245,"feed_emoji":"🐟","tokens_out":3484,"duration_ms":38637,"temperature":0.7,"pith_summary":"The paper is trying to establish that replacing global image-text alignment with selective local vision-language alignment improves fine-grained fish re-identification. On the longitudinal Melops dataset of wild corkwing wrasse, the proposed SLAP method raises mAP from 53.0% to 55.2% and Rank-1 from 44.8% to 46.6% under the closed-set protocol, and from 37.9% to 40.3% mAP and 31.6% to 33.7% Rank-1 under the open-set protocol. These gains come from letting a Partial Optimal Transport plan match only a subset of visual patches to identity-aware prompts, emphasizing discriminative body regions while suppressing background. The paper also shows consistent gains on two additional marine ReID datasets without dataset-specific tuning.","feed_headline":"Selective alignment lifts fish re-ID mAP by 2.2 points","feed_subtitle":"Partial optimal transport lets identity-relevant body regions dominate cross-modal supervision, beating global CLIP alignment.","key_machinery":"The load-bearing mechanism is the Partial Optimal Transport alignment loss. Visual patch embeddings and identity-aware prompt embeddings define a cosine-distance cost matrix; the transport plan is optimized with relaxed marginal constraints, an entropy regularizer, and a transported mass ratio that controls how much of the visual area can be matched. This plan gives soft prompt-to-patch correspondences that concentrate supervision on the strongest cross-modal matches. It is trained jointly with batch-hard triplet losses and auxiliary identity classification, while LoRA adapts the visual encoder and learnable prompt tokens adapt the text encoder. At inference only the LoRA-adapted visual encoder remains, so the selective alignment adds no deployment compute.","core_discovery":"The paper's central claim is that conventional global alignment between a whole image embedding and a text embedding is a bottleneck for fine-grained animal re-identification, because it lets background and weakly discriminative regions share in cross-modal supervision. SLAP instead computes a transport plan between local visual patch embeddings and multiple identity-aware prompt embeddings, with partially relaxed marginal constraints so that only a fraction of patches need participate. On the Melops dataset this selective local alignment outperforms all compared CLIP-based methods on every reported metric, with the largest relative gains under the open-set protocol where test identities are partly unseen. Ablations show the POT formulation beats no cross-modal supervision, global alignment, and full optimal transport, indicating that both locality and selectivity contribute to the improvement.","pith_inferences":["If the proposed mechanism is genuinely driving the gains, a testable extension is to visualize the learned transport plan and check whether mass concentrates on the operculum, eye, and snout regions the paper cites as identity-relevant; that would confirm the selectivity interpretation rather than a generic regularization effect.","A stronger stress test would compare POT against random or attention-based patch selection with the same transported mass; if random selection performs equally, the paper's selectivity story would need revision.","The same selective alignment idea may transfer to other fine-grained recognition tasks where discriminative information is localized, such as person re-identification or vehicle re-identification, though the paper does not test those settings.","The paper does not directly verify that LoRA-adapted patch embeddings and prompt embeddings reside in a shared space where cosine distance marks identity-relevant locality; that premise is inferred only from downstream retrieval gains."],"forward_implications":["If the central claim holds, CLIP-based ReID methods should move away from single global image-text alignment when identity cues are localized.","The open-set improvements suggest selective local alignment transfers better to previously unseen identities than global alignment, which matters for long-term wildlife monitoring.","The method retains a purely visual inference pipeline, so practitioners can adopt it without additional text or transport computation at deployment time.","The ablation pattern indicates that triplet supervision, POT-based alignment, and auxiliary identity loss each play distinct roles, meaning future methods should treat selective cross-modal alignment as complementary to metric learning rather than a replacement.","Consistent gains on sea star and sea turtle datasets suggest the benefit is not specific to fish or to the Melops collection protocol."],"supporting_citations":[{"why":"Supplies the Melops longitudinal corkwing wrasse dataset on which the main comparisons and ablations are run.","marker":"[27,28]"},{"why":"MetaPrompt-ReID is the strongest baseline and the global-alignment method that SLAP is directly compared against and built upon.","marker":"[30]"},{"why":"Provides the pretrained CLIP image and text encoders whose shared embedding space is used for both visual patches and prompt embeddings.","marker":"[24]"},{"why":"LoRA is the parameter-efficient visual adaptation method that keeps most of the visual backbone frozen.","marker":"[9]"},{"why":"Provides the optimal transport framework and notation that the partial transport formulation extends.","marker":"[31]"},{"why":"Supplies the partial optimal transport problem and its relaxed marginal constraints used to select a subset of patch-prompt correspondences.","marker":"[23]"},{"why":"Prior work introducing OT-based local prompt learning for vision-language models, which motivates prompt-to-patch alignment.","marker":"[4]"},{"why":"Provides the learnable prompt-tuning method for adapting frozen vision-language encoders to new tasks.","marker":"[37]"}],"fun_headline_variants":["Local selection beats global CLIP for fish re-ID","Partial optimal transport picks fish ID cues from patches","Selective local alignment sharpens fish re-ID","Avoid forced matches: POT boosts fish ID accuracy","Partial transport highlights fish's telltale patches"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that after LoRA adaptation, visual patch embeddings and identity-aware prompt embeddings lie in a shared space where cosine distance reliably marks identity-relevant body regions, so that the transport plan selects informative patches and suppresses background.","fun_headline_variants_meta":{"raw":{"variants":["Local selection beats global CLIP for fish re-ID","Partial optimal transport picks fish ID cues from patches","Selective local alignment sharpens fish re-ID","Avoid forced matches: POT boosts fish ID accuracy","Partial transport highlights fish's telltale patches"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00059,"raw_usage":{"total_tokens":2738,"prompt_tokens":884,"completion_tokens":1854,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":500,"completion_tokens_details":{"reasoning_tokens":1782}},"tokens_in":500,"tokens_out":1854,"duration_ms":15948,"temperature":1.0,"reasoning_tokens":1782,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:22:17.904753+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the learned transport plan on Melops test images: if the transported mass does not concentrate on identity-relevant regions such as the operculum, eye, and snout, or if replacing POT with uniformly random patch selection of equal mass produces no drop in mAP, then the selective-alignment explanation is not supported.","supporting_citations":[{"cited_title":"Expert Systems with Applications332, 133618 (2026)","cited_arxiv_id":null,"evidence_quote":"MetaPrompt-ReID is the strongest baseline and the global-alignment method that SLAP is directly compared against and built upon."},{"cited_title":"In: ICML","cited_arxiv_id":null,"evidence_quote":"Provides the pretrained CLIP image and text encoders whose shared embedding space is used for both visual patches and prompt embeddings."},{"cited_title":"In: ICLR (2023)","cited_arxiv_id":null,"evidence_quote":"Supplies the partial optimal transport problem and its relaxed marginal constraints used to select a subset of patch-prompt correspondences."},{"cited_title":"In: CVPR (2022) SLAP for Fish Re-Identification 19","cited_arxiv_id":null,"evidence_quote":"Provides the learnable prompt-tuning method for adapting frozen vision-language encoders to new tasks."}],"review_version":1}