{"id":"4b161f2c-7c70-4880-a64d-41e8183eb866","arxiv_id":"2412.00433","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A top-K token selector added to a view-decoupled transformer improves aerial-ground person re-identification accuracy on CARGO and AG-ReID, but the paper's specifications are internally inconsistent.","lead":"This paper adds a token-selection module to an existing view-decoupled transformer for aerial-ground person re-identification. It reports modest accuracy gains (up to 1.18% mAP on CARGO) over the VDT baseline, but the method description contains mathematical errors and internal inconsistencies.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (2)'s Gumbel-Softmax is not a continuous relaxation of Top-K; it is a soft weighted average over all tokens, so the paper's central mechanism is not the one being evaluated and the reported gains cannot be attributed to dynamic token selection.","rationale":"The reader's weakest assumption identifies the same load-bearing issue: Eq. (2) is not a valid continuous relaxation of Top-K selection. My independent reading confirms this and adds a second concrete symptom: the same equation is a dense weighting, so it cannot drop tokens, which contradicts the paper's own claims about redundancy reduction and computational efficiency. The paper provides no code, no formal verification, and no independent support for the mechanism, so the empirical numbers cannot be traced to the described method. I considered whether other issues -- such as the ablation baseline inconsistent with Table II, or possible test-set-driven hyperparameter selection -- are more central, but they are secondary: even if all numbers are reproducible, the absence of a well-defined top-K selector means the method as described is not the claimed dynamic token selection method. Therefore the rejection verdict stands without change. I would encourage the authors to either clarify the actual differentiable selection procedure (e.g., straight-through hard top-K or a perturbed optimizer) or revise the claims to match a soft-attention reweighting module.","tokens_in":8672,"tokens_out":3114,"duration_ms":33819,"concrete_test":"Implement two controlled variants on the CARGO A<->G protocol with identical training settings: (A) exactly as Eq. (2), using Gumbel-Softmax to produce a dense weight vector over all tokens and no token dropping; (B) hard top-K=2 selection with straight-through gradients or a perturbed top-K optimizer. If variant (A) reproduces the reported 50.63% Rank-1 / 43.39% mAP / 29.46% mINP, then the published equation is doing all-token weighting, not selection. As an additional check, run variant (A) with K=2 and with K=M (i.e., retain all tokens); if the results are identical, the selection mechanism has no effect, confirming that Eq. (2) does not implement top-K.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that VTS selects the top-K tokens via a differentiable relaxation. In Section III-B, Eq. (1) defines si = softmax(t_i^T Wq Wk^T t_i / sqrt(d)). As written, the argument is a scalar, so the softmax has no normalization over the token index; even if intended as a softmax over i, the scores become a probability distribution over all tokens. Eq. (2) then computes \\hat s_i = exp((s_i+g_i)/tau) / sum_j exp((s_j+g_j)/tau), which is just a Gumbel-Softmax reweighting of that distribution. As tau -> 0, this converges to a one-hot argmax, i.e., selection of exactly one token, not K tokens. There is no mechanism in Eq. (2) -- no repeated sampling without replacement, no masking, no top-K perturbed optimizer as in Berthet et al. [17] -- that yields K selected tokens. Consequently, Eq. (2) is not a continuous relaxation of the TopK operation in Eq. (1). Moreover, since \\hat s_i is a dense weight vector over all M tokens, no tokens are dropped; this contradicts the stated benefits of reduced redundancy and computational efficiency. The implementation details say 'top two rated tokens' are selected, but no equation or procedure in the paper produces exactly two discrete tokens in a differentiable way. If the actual code uses a hard top-K with straight-through estimation, the paper does not say so, making the method unreproducible. The ablation in Table IV comparing 'with/without VTS' therefore tests an underspecified module; the observed gain could come from an extra attention layer or from soft reweighting rather than from token selection. Since every novelty and contribution claim depends on the correctness of this selection mechanism, this is the load-bearing weakness.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Dynamic Token Selective Transformer (DTST) for Aerial-Ground Person Re-identification (AGPReID), built on the View-Decoupled Transformer (VDT) baseline. The core novelty is a Visual Token Selector (VTS) that is claimed to rank image tokens by importance and select the top-K tokens in a differentiable manner via a Gumbel-Softmax relaxation. The authors report state-of-the-art results on the CARGO and AG-ReID datasets, with ablations over the number of selected tokens, attention heads, and token insertion positions.","tokens_in":9107,"tokens_out":4211,"duration_ms":40882,"significance":"The problem addressed is relevant and the authors choose appropriate public benchmarks (CARGO, AG-ReID) with standard evaluation metrics. The paper also provides a useful ablation perspective on token selection hyperparameters. However, the central methodological claim is not supported by the equations as written: the proposed differentiable top-K mechanism is not actually a top-K selector, and the baseline numbers are internally inconsistent across tables. If the method were correctly formulated and the experiments properly controlled, dynamic token selection for AGPReID could be a meaningful contribution; as presented, the contribution is not established.","major_comments":[{"comment":"The scalar softmax in Eq. (1) is degenerate: for a scalar input x, softmax(x) = exp(x)/exp(x) = 1, so every token receives the same importance score and the TopK ranking is undefined. If the intended meaning is a softmax over the token dimension, the equation must explicitly normalize over j, e.g., s_i = exp(t_i^T Wq Wk^T t_i / sqrt(d)) / sum_j exp(...). As written, the proposed scoring mechanism cannot produce a meaningful ranking.","section":"Section III-B, Eq. (1)"},{"comment":"The Gumbel-Softmax expression is not a continuous relaxation of the TopK operation. As tau -> 0, the normalized exponential over (s_i + g_i)/tau converges to a one-hot distribution on a single argmax token, not to a selection of K tokens. For finite tau, it is a dense weighting that combines all M tokens, so no token is dropped, contradicting the claimed redundancy reduction. The implementation detail in Section IV-B that 'top two rated tokens' are selected is not derivable from Eq. (2); no repeated sampling without replacement, masking, or perturbed top-K optimizer is specified. Consequently, the described VTS does not implement the advertised top-K selection mechanism, and the reported accuracy gains cannot be attributed to it.","section":"Section III-B, Eq. (2)"},{"comment":"The claimed improvement over the VDT baseline is internally inconsistent. In Table II, VDT achieves 47.50 Rank-1 and 42.21 mAP on the A↔G protocol of CARGO; in Table IV, model-a without the VTS achieves only 45.00 Rank-1 and 42.05 mAP on the same protocol. If model-a is intended to be the VDT baseline, the discrepancy changes the reported gains from 3.13% to 5.63% in Rank-1 and from 1.18% to 1.34% in mAP. The source of this discrepancy is unexplained, so the ablation in Table IV does not cleanly isolate the contribution of the VTS.","section":"Section IV-C vs Section IV-D, Tables II and IV"},{"comment":"The final model configuration (2 attention heads, top-2 selected tokens, last-layer insertion) is chosen by evaluating configurations on the CARGO A↔G protocol, which is the same protocol used for the state-of-the-art comparison in Table II. This means the reported SOTA result is the best result from a search over the test evaluation protocol, which inflates the expected performance and makes the comparison to prior methods optimistic. A validation split or repeated cross-validation should be used for hyperparameter selection before reporting SOTA numbers.","section":"Section IV-D, Table V"}],"minor_comments":[{"comment":"The table header contains typos: 'mMAP' should be 'mAP' and 'blod' should be 'bold'; also, the text says 'conducted on the on the CARGO dataset', with a duplicated 'on the'.","section":"Section IV-D, Table V"},{"comment":"The AG-ReID dataset is attributed to reference [12] in the Related Work but to reference [19] in Table I and Section IV-A; the authors should clarify which version of the dataset is used and cite consistently.","section":"Section II-A and Table I"},{"comment":"The sentence 'Building on the principles of differentiable top-K selection [18]' cites reference [18] (TS2-Net), but the differentiable perturbed-optimizer method described in the text is reference [17] (Berthet et al.); the citation appears incorrect.","section":"Section II-B"},{"comment":"The phrase 'insertion position of VST' should read 'insertion position of VTS' for consistency with the abbreviated module name.","section":"Section IV-D"},{"comment":"The paper does not report whether results are averaged over multiple runs or whether evaluation is performed on a single test protocol; reporting run-to-run variance or at least noting single-run results would improve reliability.","section":"Tables II and III"}],"recommendation":"reject","confidential_remarks":"The core mathematical description of the proposed token selection mechanism is not a top-K selector, and the baseline inconsistency between Tables II and IV makes the experimental attribution unreliable. Since these are load-bearing issues for the paper's central contribution, I do not see how they can be fixed within a standard revision cycle without re-deriving the method and re-running the experiments. The hyperparameter selection on the test protocol compounds the concern about the validity of the SOTA claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi [Name],\n\nQuick take: the advertised contribution — a differentiable top-K token selector — is not actually described by the equations in the paper. Eq. (1) runs a softmax on a scalar, so it returns 1 for every token; Eq. (2) is a Gumbel-Softmax reweighting that converges to a one-hot argmax (one token) as temperature drops, not a top-K selection. There is no repeated sampling, masking, or perturbed top-K machinery in the text. So the central mechanism is undefined.\n\nWhat's real and useful: applying token selection to aerial-ground person re-identification is a reasonable extension, and the authors are honest that they build on STTS, TS2-Net, and Berthet et al. The experiments span two datasets, several protocols, and a reasonable ablation of heads/tokens/positions. If the code actually does a hard top-2 with straight-through estimation, the reported gains are plausible and modest (~1-3%). But the manuscript doesn't say that.\n\nThe soft spots are load-bearing. The ablation baseline in Table IV (45.00 Rank-1, 42.05 mAP) doesn't match the VDT numbers in Table II (47.50 Rank-1, 42.21 mAP), so the claimed 5.63% Rank-1 improvement from adding VTS is not against the same baseline. Also, the final hyperparameters (2 tokens, 2 heads, last layer) were selected by scanning exactly those choices on the A↔G test protocol of CARGO; the headline result is the best of that search. That's test-set-driven model selection, even if the search is small.\n\nWho is this for? A reader tracking AGPReID might want to know that token selection has been tried here. But anyone wanting to reproduce or build on it will be stuck. I would not cite it as evidence that token selection works, because the method isn't pinned down.\n\nRecommendation: desk-reject in current form. If the authors fix the equations, clarify how exactly the discrete top-K is made differentiable, align the ablation baseline, and release code, it could be a modest but sound paper for a workshop or a second-tier venue. Right now the central claim isn't verifiable.\n\nBest,\n[Your name]","headline":"The paper's central token-selection mechanism is undefined by its own equations, so the reported gains can't be attributed to it.","tokens_in":9630,"tokens_out":3741,"would_cite":false,"duration_ms":34012,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Dynamic token selection lets an aerial-ground person re-identification transformer beat its full-token baseline on both CARGO and AG-ReID.","keywords":["aerial-ground person re-identification","token selection","vision transformer","Gumbel-Softmax","top-k selection","view-decoupled transformer","CARGO","AG-ReID"],"falsifier":"Take the trained DTST and, for a fixed input, compare the indices produced by the Gumbel-Softmax relaxation in equation (2) with the indices produced by the exact hard top-$K$ rule in equation (1) under the same token scores; if the two index sets disagree on a substantial fraction of inputs, the model is not performing the token selection the paper describes, and one could retrain with straight-through hard selection to see whether the accuracy gains persist.","tokens_in":8485,"feed_emoji":"🎯","tokens_out":7191,"duration_ms":65866,"temperature":0.7,"pith_summary":"This paper asks whether an aerial-ground person re-identification model can match people across cameras better by throwing away most of the image and attending only to the most informative patches. It proposes the Dynamic Token Selective Transformer (DTST), which scores each image token with a lightweight attention predictor, keeps the top two tokens, and feeds only those into a view-decoupled transformer. On the CARGO dataset's aerial-to-ground protocol the paper reports 50.63% Rank-1 and 43.39% mean average precision, beating the VDT baseline by 3.13 points in Rank-1 and 1.18 points in mAP; on AG-ReID it reports Rank-1 gains of 0.57 points (aerial to ground) and 1.04 points (ground to aerial). The paper's claim is that dynamically dropping redundant tokens is what drives these gains, rather than the baseline's view-decoupling machinery.","feed_headline":"Aerial-ground person re-ID improves when model keeps two tokens","feed_subtitle":"On CARGO's aerial-to-ground protocol, DTST's top-2 token selection beats the full-token baseline by 3.13 Rank-1 points.","key_machinery":"The load-bearing object is the Visual Token Selector (VTS) inserted after the final transformer block. It computes a relevance score per token, $s_i$, as a softmax over a query-key product, then picks the top-$K$ tokens; a Gumbel-Softmax reweighting is used as a continuous relaxation so gradients can flow through the discrete selection. The selector's two hyperparameters, $K=2$ and a two-head attention predictor, are the settings the ablations identify as best. The VTS sits inside the View-Decoupled Transformer's token stream, so the selected tokens inherit the meta/view token subtraction and orthogonal-loss supervision that the baseline uses.","core_discovery":"The paper's central claim is that a View-Decoupled Transformer for aerial-ground person re-identification becomes more accurate when a Visual Token Selector (VTS) picks the $K=2$ most informative image tokens and discards the rest before final feature aggregation. The selector scores tokens with a softmax attention predictor, ranks them, and uses a Gumbel-Softmax relaxation of hard top-$K$ selection so the choice is trainable end to end. On the CARGO benchmark this configuration improves the VDT baseline by 3.13% in Rank-1, 1.18% in mAP, and 0.43% in mINP under the aerial-to-ground protocol, and by 1.60%, 1.51%, and 2.00% under the ALL protocol. On AG-ReID, Rank-1 rises by 0.57% for aerial-to-ground and 1.04% for ground-to-aerial. The paper interprets these gains as evidence that selecting fewer, identity-relevant tokens reduces redundancy and sharpens identity features.","pith_inferences":["A testable extension the paper leaves implicit is whether the gains come from token pruning itself or from the extra parameters of the two-head selector; retraining with a fixed, non-selective two-token aggregation would isolate the mechanism.","The paper's efficiency claim is not backed by latency or FLOPs measurements, so a direct wall-clock comparison with the VDT baseline would determine whether the selector's overhead cancels the savings from discarding tokens.","Because the Visual Token Selector is architecture-agnostic, it could be grafted onto other view-heterogeneous retrieval pipelines or video models, but such transfer is an extension beyond the two benchmarks reported here.","One could test whether the Gumbel noise is necessary at all by replacing equation (2) with a straight-through hard top-$K$ estimator; equal accuracy would suggest the relaxation is incidental to the result."],"forward_implications":["On CARGO's aerial-to-ground protocol, DTST reaches 50.63% Rank-1, 43.39% mAP, and 29.46% mINP, which beats VDT by 3.13, 1.18, and 0.43 points respectively.","On CARGO's ALL protocol, DTST exceeds VDT by 1.60 points Rank-1, 1.51 points mAP, and 2.00 points mINP.","On AG-ReID, DTST improves Rank-1 by 0.57 points for aerial-to-ground and 1.04 points for ground-to-aerial compared with VDT.","Ablations show that the best configuration uses the selector after the last layer with two attention heads and two selected tokens; raising the head count or selecting more than three tokens lowers accuracy.","Because only the top-$K$ tokens continue to the final aggregation stage, the method is intended to cut redundant computation, although the paper does not report measured inference time."],"supporting_citations":[{"why":"Supplies the View-Decoupled Transformer baseline, the CARGO dataset, and the view-decoupling machinery that DTST extends.","marker":"[15]"},{"why":"Provides the differentiable perturbed top-K selection principle used to make the Visual Token Selector trainable.","marker":"[17]"},{"why":"Supplies the differentiable top-K token-selection technique that the VTS builds on.","marker":"[18]"},{"why":"Motivates token selection in vision transformers as a way to cut redundancy and computation.","marker":"[16]"},{"why":"Provides the AG-ReID benchmark and its aerial-to-ground and ground-to-aerial evaluation protocols.","marker":"[19]"},{"why":"Defines the aerial-ground person re-identification task and introduces the AG-ReID dataset.","marker":"[12]"}],"fun_headline_variants":["Selecting top-2 tokens boosts aerial-ground person re-ID","Fewer tokens, sharper identity: top-2 selection wins in aerial-ground re-ID","DTST: keeping only 2 image tokens improves re-ID accuracy","Top-2 token pruning lifts Rank-1 by 3.13% in aerial-ground re-ID","Two tokens beat full-token model in aerial-ground re-ID"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole accuracy story depends on equation (2) being a faithful differentiable stand-in for the hard top-$K$ selection in equation (1), because if the relaxation does not actually choose the top-$K$ tokens during training, the reported gains cannot be attributed to token selection.","fun_headline_variants_meta":{"raw":{"variants":["Selecting top-2 tokens boosts aerial-ground person re-ID","Fewer tokens, sharper identity: top-2 selection wins in aerial-ground re-ID","DTST: keeping only 2 image tokens improves re-ID accuracy","Top-2 token pruning lifts Rank-1 by 3.13% in aerial-ground re-ID","Two tokens beat full-token model in aerial-ground re-ID"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001056,"raw_usage":{"total_tokens":4445,"prompt_tokens":973,"completion_tokens":3472,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":589,"completion_tokens_details":{"reasoning_tokens":3372}},"tokens_in":589,"tokens_out":3472,"duration_ms":18559,"temperature":1.0,"reasoning_tokens":3372,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:23:19.762822+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the trained DTST and, for a fixed input, compare the indices produced by the Gumbel-Softmax relaxation in equation (2) with the indices produced by the exact hard top-$K$ rule in equation (1) under the same token scores; if the two index sets disagree on a substantial fraction of inputs, the model is not performing the token selection the paper describes, and one could retrain with straight-through hard selection to see whether the accuracy gains persist.","supporting_citations":[{"cited_title":"View-decoupled transformer for person re-identification under aerial-ground camera network,","cited_arxiv_id":null,"evidence_quote":"Supplies the View-Decoupled Transformer baseline, the CARGO dataset, and the view-decoupling machinery that DTST extends."},{"cited_title":"Learning with differentiable pertubed optimizers,","cited_arxiv_id":null,"evidence_quote":"Provides the differentiable perturbed top-K selection principle used to make the Visual Token Selector trainable."},{"cited_title":"Ts2- net: Token shift and selection transformer for text-video retrieval,","cited_arxiv_id":null,"evidence_quote":"Supplies the differentiable top-K token-selection technique that the VTS builds on."},{"cited_title":"Efficient video transformers with spatial-temporal token selection,","cited_arxiv_id":null,"evidence_quote":"Motivates token selection in vision transformers as a way to cut redundancy and computation."},{"cited_title":"Ag- reid.v2: Bridging aerial and ground views for person re-identification,","cited_arxiv_id":null,"evidence_quote":"Provides the AG-ReID benchmark and its aerial-to-ground and ground-to-aerial evaluation protocols."},{"cited_title":"Aerial-ground person re-id,","cited_arxiv_id":null,"evidence_quote":"Defines the aerial-ground person re-identification task and introduces the AG-ReID dataset."}],"review_version":1}