{"id":"98d55b7f-283c-4ec9-9226-88d0fac2a653","arxiv_id":"2507.01711","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"AdaGCD applies adaptive slot attention to decompose DINO image features into semantic components and pools them with global features, reporting SOTA accuracy on six GCD benchmarks.","lead":"AdaGCD is a new method for Generalized Category Discovery that combines adaptive slot attention with contrastive learning, using both global and local image features. The authors report state-of-the-art accuracy on six image classification benchmarks, with the largest gains on fine-grained datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed benefit of slot-based local features is not isolated: no ablation separates MSE reconstruction regularization from slot pooling, so the SOTA gain may not stem from category-aligned slots.","rationale":"The reader's weakest_assumption correctly identifies that the slots may not capture category-discriminative local semantics, and that no analysis connects reconstruction quality or slot content to downstream clustering. My concern is more specific: the experimental design does not isolate the slot-attention mechanism from the other changes introduced relative to CMS. A simple spatial-pooling baseline would settle whether the reported gains require the slot machinery or just an additional feature stream / reconstruction regularizer. This is load-bearing because the paper's central contribution is the component clusterer; if it is not the source of the empirical improvement, the SOTA claim would rest on an unvalidated mechanism. The reader's verdict of CONDITIONAL remains appropriate: the paper should be accepted only if the authors add the missing control and release code/error bars. I therefore do not change the verdict, but the condition should explicitly include this ablation. The R3D inconsistency in Sec. III-B is minor but worth fixing; it underscores that the architecture description is not yet precise enough to reproduce without the supplemental material.","tokens_in":10206,"tokens_out":5484,"duration_ms":63737,"concrete_test":"Run an ablation on CUB and Herbarium 19 in which the component clusterer Sθ is replaced by a non-parametric average pooling of the DINO local feature map h over spatial positions, keeping the same concatenation and the same MSE reconstruction loss on the pooled feature (or omitting the reconstruction loss entirely). If the accuracy stays within 1 point of AdaGCD over 3 seeds, the slot-attention decomposition is not the source of the improvement and the claimed mechanism is unsupported; if accuracy drops substantially, the slots matter. Also report the mean and standard deviation to assess whether the observed differences are significant.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that AdaGCD's gains over CMS come from 'clustered features' learned by Adaptive Slot Attention (Sec. III-A). Relative to CMS, AdaGCD changes multiple components at once: it adds a masked-slot decoder with MSE reconstruction loss (Eq. 1), a slot-pooling operation, and a concatenation of pooled slot features with the DINO global feature (Sec. III-B). The reported improvements (e.g., CUB 71.4 vs 68.2, Herbarium 19 50.1 vs 36.4, Table I) are attributed to 'spatial local information,' but Sec. IV-D ablates only loss ratios and Kmax; it never removes the slot branch or replaces it with a simple spatial pooling. Consequently, the improvements could plausibly arise from the reconstruction loss acting as an auxiliary regularizer, or from simply appending a second D-dimensional feature stream, rather than from slots capturing category-discriminative local semantics. The paper also states gall ∈ R3D after concatenating two D-dimensional vectors (Sec. III-B), which is an internal inconsistency that further suggests the feature-dimension details are not fully pinned down. Without an ablation isolating the slot-attention mechanism, the paper's central claim that clustered local features drive the performance gain is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AdaGCD, a generalized category discovery method that augments the CMS contrastive objective with Adaptive Slot Attention (AdaSlot). DINO ViT local feature maps are decomposed into a variable number of slots, the slots are pooled into an image-level vector, this vector is concatenated with the DINO global feature, and the combined representation is trained with supervised and unsupervised contrastive losses plus an MSE reconstruction loss. Experiments on six GCD benchmarks report state-of-the-art or competitive accuracy, with the largest gain on Herbarium 19 (50.1 vs. 45.1 for LegoGCD). Ablations vary the loss weights and the maximum slot number Kmax.","tokens_in":10442,"tokens_out":3770,"duration_ms":47133,"significance":"If the reported gains are reproducible and can be attributed to the slot-based local features, AdaGCD would be a useful contribution: it brings object-centric and spatially localized representations into GCD, and its adaptive slot-count mechanism addresses a practical limitation of fixed-capacity slot models. The paper provides full comparison tables on six datasets, loss-ratio ablations, a Kmax sensitivity study, and a t-SNE visualization. However, the central attribution of the gains to the slot mechanism is not isolated by any ablation, and there is a dimensional inconsistency in the feature concatenation, so the significance is conditional on a clarifying revision.","major_comments":[{"comment":"The central claim that clustered spatial local features drive the performance gains is not established, because no ablation removes the slot branch or replaces it with a simpler pooling operation. Table III varies only the loss weights while keeping the AdaSlot component, and Table IV varies Kmax but all rows still include AdaSlot plus the reconstruction loss. The single fixed-slot row compares only Kmax=5. To isolate the effect of Adaptive Slot Attention, the authors should add ablations such as (a) CMS plus the reconstruction loss without slots, and (b) CMS with mean pooling of DINO patch features concatenated to the global feature. Without these, the reported improvements over CMS could arise from the auxiliary MSE regularization or from simply appending an extra feature stream rather than from slots that capture category-discriminative local semantics.","section":"Sec. IV-D, Tables III and IV"},{"comment":"The text states that the pooled clustered features are concatenated with the global feature gdino ∈ RD and that this \"results in a unified vector gall ∈ R3D.\" Since both the global feature and the pooled slot features are D-dimensional, the concatenation should be in R2D, not R3D. This dimensional mismatch is unexplained and is important for reproducibility; the authors should specify exactly which vectors are concatenated (including whether the two augmented views are combined) and correct the stated dimension.","section":"Sec. III-B"},{"comment":"The claim that the method \"removes the need for predefined slot counts\" is overstated. Adaptive Slot Attention still requires the maximum slot number Kmax as a predefined hyperparameter, and Table IV shows that accuracy varies with Kmax (e.g., All accuracy ranges from 69.2 to 71.4). The manuscript should describe the contribution as adaptively selecting the number of slots up to a maximum, state how the number of active slots is determined (threshold or selection criterion), and discuss the dependence on Kmax in relation to the claimed adaptivity.","section":"Abstract and Sec. III-A"},{"comment":"Several reported margins are small (e.g., ImageNet100 Old: 95.2 vs. 95.6 for CMS; FGVC Aircraft New: 49.4 vs. 52.3 for CMS) and all results appear to come from a single run without standard deviations or multiple seeds. Since the overall SOTA claims depend on these close comparisons, the authors should report mean and standard deviation over at least three seeds, or release code so that the results can be reproduced and variance assessed.","section":"Table I and Sec. IV-C"}],"minor_comments":[{"comment":"The loss is called an MSE loss but is written as ||h − ĥ|| without a square; the exact form (L2 norm vs. squared L2 norm, and whether it is averaged over pixels) should be stated precisely.","section":"Eq. (1)"},{"comment":"The table header is confusing: \"Classes labeled unlabeled labeled unlabeled\" does not clearly indicate that the columns refer to known classes, unknown classes, labeled images, and unlabeled images. Please rename the columns.","section":"Table II"},{"comment":"The notation \"slot-scale features si_out\" is used without first defining the index i; the slot indexing and the relationship of si_out to sout ∈ R^{Kmax×D} should be made explicit.","section":"Sec. III-A"},{"comment":"Reference [41] for Gumbel-Softmax is cited as ICLR 2022, but the original Gumbel-Softmax paper appeared at ICLR 2017; please correct the citation or use the appropriate version.","section":"References"},{"comment":"The caption says \"Visualization of ten instances\" but the figure shows t-SNE embeddings of learned representations; the caption should state that the plot is a t-SNE projection of the learned features for sampled instances.","section":"Fig. 3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a plausible engineering contribution, and the reported gains on several benchmarks are interesting. The main weakness is the lack of an ablation isolating the slot-attention mechanism, which is essential to the paper's central claim. The dimensional inconsistency in Sec. III-B also needs to be fixed before the method can be reproduced. I would look favorably on a revision that adds the suggested ablations, reports multiple seeds, and clarifies the adaptive slot-count mechanism and feature dimensions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: AdaGCD is a plausible incremental contribution to Generalized Category Discovery. It takes AdaSlot, a published module for adaptive slot attention, and grafts it onto CMS-style contrastive learning, with slot features pooled and concatenated to DINO global features. The integration is new, and the reported numbers on six benchmarks, especially Herbarium 19 (50.1 vs 45.1 for LegoGCD) and CUB (71.4 vs 68.2), are solid, though single-run and without code.\n\nWhat the paper does well: the framing is clear, the method section is readable, and the ablations on loss weights and Kmax are more than many GCD papers bother with. The Kmax study at least includes a fixed-slot baseline, and performance is stable across slot capacities. The fine-grained gains make sense if slots are capturing local part information.\n\nWhere it falls short: the central attribution of the gain to 'clustered features' is not actually isolated. Compared to CMS, AdaGCD adds three things at once: the slot attention module, an MSE reconstruction loss on DINO local features, and a second feature stream appended to the global representation. There is no ablation that removes the slot branch or replaces it with plain spatial pooling. The improvement could therefore come from the reconstruction loss acting as an auxiliary regularizer, or simply from the extra feature stream, rather than from slots aligning with category-discriminative semantics. The paper claims the latter but does not demonstrate it.\n\nThere is also a small internal inconsistency: concatenating a D-dimensional global feature with a D-dimensional pooled slot feature gives R2D, not R3D as written. Minor, but the kind of thing that makes you wonder about dimension bookkeeping.\n\nThe abstract and conclusion overstate the 'no predefined slot count' point. The method still requires Kmax; it removes the need to choose the exact number, not the upper bound. And the fixed-slot comparison is only at Kmax=5, a weak test of adaptivity.\n\nWho this is for: anyone working on GCD or object-centric representation learning will find this a useful data point. It is not a reorganization of the field. It deserves a serious referee: ask for an ablation that isolates the slot mechanism, error bars or multiple seeds, and a fix to the dimension typo. I would engage with it.","headline":"AdaGCD is a plausible incremental GCD method that combines AdaSlot with CMS-style contrastive learning, but the key ablation is missing, so the slot-based mechanism remains unproven.","tokens_in":10976,"tokens_out":2616,"would_cite":true,"duration_ms":30247,"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":"AdaGCD combines adaptive slot attention over DINO local features with global contrastive learning and reports state-of-the-art results on six GCD benchmarks.","keywords":["Generalized Category Discovery","Adaptive Slot Attention","contrastive learning","DINO local features","spatial local information","fine-grained classification","open-world category discovery","slot attention"],"falsifier":"Ablate the component clusterer by replacing AdaSlot with simple mean pooling of the DINO local feature map (or with fixed-slot Slot Attention) and rerun the six benchmarks; if the all-class accuracies do not fall materially, the claim that adaptive slot components drive the improvement is falsified.","tokens_in":9991,"feed_emoji":"🧩","tokens_out":9735,"duration_ms":95486,"temperature":0.7,"pith_summary":"The paper proposes AdaGCD, a contrastive learning framework for Generalized Category Discovery (GCD), where a model must label images from both labeled and unseen classes without knowing how many unseen classes there are. The central claim is that the 2D spatial feature maps produced by the DINO backbone carry category-discriminative local structure that global features alone miss, and that decomposing this structure with an Adaptive Slot Attention module improves clustering. AdaSlot chooses how many slots to use for each image through a Gumbel-Softmax selection module, so the number of components is not fixed in advance. The pooled slot features are concatenated with DINO's global feature and trained with contrastive objectives adapted from a mean-shift baseline. On six benchmarks the method reports the highest overall clustering accuracy in its comparison table, with the largest gain on Herbarium 19 (50.1 versus 45.1).","feed_headline":"Adaptive slots lift category discovery accuracy on six benchmarks","feed_subtitle":"Local slot features plus DINO global features set six records, topped by +5.0 on Herbarium 19.","key_machinery":"The load-bearing object is the component clusterer $S_\\theta$, built on Adaptive Slot Attention (AdaSlot). Given the DINO local feature map $h \\in \\mathbb{R}^{H \\times W \\times D}$ and $K_{\\max}$ Gaussian-initialized slots $s_{\\mathrm{init}}$, AdaSlot uses a Gumbel-Softmax selection module to choose a per-image number of slots and a masked slot decoder $d_\\theta$ to reconstruct $h$ under the loss $\\mathcal{L}^{\\mathrm{rec}}_\\theta = \\| h - \\hat{h} \\|$. The selected output slots are average-pooled and concatenated with the DINO global feature $g_{\\mathrm{dino}}$ to form the image-scale vector in $\\mathbb{R}^{3D}$ that drives the supervised and unsupervised contrastive losses. The adaptive slot count is what removes the need to pre-specify how many components each image contains.","core_discovery":"Concretely, the paper reports all-class clustering accuracy of 83.4 on CIFAR100, 87.0 on ImageNet100, 71.4 on CUB, 59.6 on Stanford Cars, 56.4 on FGVC Aircraft, and 50.1 on Herbarium 19, each the highest All score in Table I. The method treats an image as an assembly of semantic components: DINO's final-layer local features are passed through AdaSlot, which selects a per-image subset from a maximum of $K_{\\max}$ slots and reconstructs the local feature map through a masked decoder under an MSE loss. Average-pooling the selected slot features and concatenating them with the DINO global representation gives an image-scale vector in $\\mathbb{R}^{3D}$ that is trained with the supervised and unsupervised contrastive losses borrowed from CMS. The paper's message is that representing images through components rather than a single global summary is what lets the model separate novel classes, especially in fine-grained domains.","pith_inferences":["Editorial inference: the same component-clustering recipe should transfer to other self-supervised vision transformers whose final-layer features keep spatial structure, provided their slots align with semantic parts; this is testable by swapping the DINO backbone.","Editorial inference: because AdaSlot picks a per-image slot count, the framework is naturally suited to multi-object or multi-label images where a single global feature cannot represent all categories, a regime the paper does not evaluate.","Editorial inference: the reconstruction term operates in DINO feature space, so the method inherits DINO's biases; with a less part-aligned backbone the MSE objective would likely need reweighting or replacement, and comparing that behavior would clarify how much of the gain is due to the slot mechanism versus the backbone."],"forward_implications":["The reported All accuracy on all six benchmarks in Table I is higher than every comparison method, so pooling adaptive slot features with the DINO global feature is a competitive strategy for GCD.","The dynamic slot count removes the need to predefine how many components each image contains, although the method still requires a maximum capacity $K_{\\max}$ and the evaluation still assumes a known class count via Hungarian matching.","The largest gains appear on fine-grained and specialist datasets (CUB, Stanford Cars, FGVC Aircraft, Herbarium 19), where category differences rest on subtle local structure rather than global appearance.","New-class accuracy improves while old-class accuracy stays strong, indicating the local-feature branch does not sacrifice known categories for novel ones.","Loss-weight tuning trades old-class retention against new-class adaptation; with weights 0.6 unsupervised, 0.3 supervised, and 0.1 reconstruction the paper reports its best balance."],"supporting_citations":[{"why":"Defines the GCD task, the labeled/unlabeled split, the Hungarian-matching evaluation protocol, and the loss-weight settings that AdaGCD inherits.","marker":"[1]"},{"why":"Supplies the DINO ViT-B/16 backbone whose global representation and 2D local feature map feed the component clusterer.","marker":"[19]"},{"why":"Provides the Adaptive Slot Attention module, including Gumbel-Softmax slot selection and the masked slot decoder used for reconstruction.","marker":"[20]"},{"why":"Supplies the contrastive mean-shift objective and its supervised/unsupervised loss composition, which AdaGCD reuses at image scale.","marker":"[26]"},{"why":"LegoGCD is the strongest recent baseline; on Herbarium 19 the paper reports 50.1 against LegoGCD's 45.1.","marker":"[27]"},{"why":"SimGCD is the parametric-classification baseline that sets the prior state of the art against which several AdaGCD results are compared.","marker":"[24]"},{"why":"Slot Attention provides the iterative attention mechanism that AdaSlot adapts from a fixed to an adaptive number of slots.","marker":"[39]"}],"fun_headline_variants":["Adaptive slots remove class count guesswork in category discovery","Dynamic slot count adapts clustering for known and novel classes","AdaSlot picks cluster numbers on the fly, boosting fine-grained discovery","Adaptive slots set six new records top gain +5.0 on Herbarium"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's gain rests on the assumption that the MSE reconstruction of DINO local features forces the slots to capture category-discriminative components rather than generic texture; if the slots instead encode low-level patterns, the pooled slot features would add noise and the reported improvements would not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive slots remove class count guesswork in category discovery","Dynamic slot count adapts clustering for known and novel classes","AdaSlot picks cluster numbers on the fly, boosting fine-grained discovery","Adaptive slots set six new records top gain +5.0 on Herbarium"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000861,"raw_usage":{"total_tokens":3725,"prompt_tokens":924,"completion_tokens":2801,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":540,"completion_tokens_details":{"reasoning_tokens":2727}},"tokens_in":540,"tokens_out":2801,"duration_ms":20277,"temperature":1.0,"reasoning_tokens":2727,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:44:17.839207+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Ablate the component clusterer by replacing AdaSlot with simple mean pooling of the DINO local feature map (or with fixed-slot Slot Attention) and rerun the six benchmarks; if the all-class accuracies do not fall materially, the claim that adaptive slot components drive the improvement is falsified.","supporting_citations":[{"cited_title":"Generalized category discovery,","cited_arxiv_id":null,"evidence_quote":"Defines the GCD task, the labeled/unlabeled split, the Hungarian-matching evaluation protocol, and the loss-weight settings that AdaGCD inherits."},{"cited_title":"Emerging properties in self-supervised vision transformers,","cited_arxiv_id":null,"evidence_quote":"Supplies the DINO ViT-B/16 backbone whose global representation and 2D local feature map feed the component clusterer."},{"cited_title":"Adaptive slot attention: Object discovery with dynamic slot number,","cited_arxiv_id":null,"evidence_quote":"Provides the Adaptive Slot Attention module, including Gumbel-Softmax slot selection and the masked slot decoder used for reconstruction."},{"cited_title":"Contrastive mean-shift learning for generalized category discovery,","cited_arxiv_id":null,"evidence_quote":"Supplies the contrastive mean-shift objective and its supervised/unsupervised loss composition, which AdaGCD reuses at image scale."},{"cited_title":"Solving the catastrophic forgetting problem in generalized category discovery,","cited_arxiv_id":null,"evidence_quote":"LegoGCD is the strongest recent baseline; on Herbarium 19 the paper reports 50.1 against LegoGCD's 45.1."},{"cited_title":"Parametric classification for generalized category discovery: A baseline study,","cited_arxiv_id":null,"evidence_quote":"SimGCD is the parametric-classification baseline that sets the prior state of the art against which several AdaGCD results are compared."},{"cited_title":"Object-centric learning with slot attention,","cited_arxiv_id":null,"evidence_quote":"Slot Attention provides the iterative attention mechanism that AdaSlot adapts from a fixed to an adaptive number of slots."}],"review_version":1}