{"id":"3000fb8d-f297-4f6f-9227-6ad6b5217296","arxiv_id":"2507.06928","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"APL improves fine-grained Generalized Category Discovery by learning shared, correspondable object-part features with an all-min contrastive loss, replacing the CLS token and gaining 2 to 6 accuracy points over SimGCD, SPTNet, and CMS.","lead":"A plug-and-play module called APL replaces the single global CLS feature used by Generalized Category Discovery models with object-part features learned from shared queries and DINO attention priors, plus a new all-min contrastive loss. On fine-grained benchmarks it reports consistent gains of about 2 to 6 accuracy points over three published GCD frameworks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The all-min contrastive loss (Eq. 5) is underspecified for unlabeled negative sampling; if negative pairs are sampled so that no cross-class pair is in the denominator, the reported gains could come from the host GCD loss alone and the central claim is unverified.","rationale":"The reader's weakest_assumption concerns DINO attention-to-part coherence and epsilon/T sensitivity. I agree that is a real fragility, but the paper itself builds in a corrective mechanism (learnable queries seeded by DINO priors, hard assignment via Gumbel-Softmax) and the empirical ablations in Tables 3 and 4 give some evidence that the queries and hard assignment improve over raw DINO priors and k-means. The deep unresolved point is the exact definition of the all-min loss for unlabeled data, because Eq. 5 is the mathematical core of the claimed novelty and the paper's stated negative-sampling rule is a verbal heuristic with no formal guarantee. If the sampling rule leaves the denominator empty or excludes most informative negatives, the loss collapses to a positive-pair-only term, and the ablation that credits row 3 over row 2 in Table 4 would be measuring something other than the proposed all-min mechanism. This is not a style complaint; it is a precise, testable specification gap. It is also more load-bearing than the generic reproducibility complaints (no code, no seeds), because it directly attacks whether the method's central contribution, as opposed to the host GCD loss, is responsible for the gain. I therefore recommend keeping the reader's CONDITIONAL verdict, upgraded with this more specific condition: the authors must supply the exact negative-sampling algorithm, including the threshold rule and batch-level guarantees, and verify via ablation that the all-min term contributes as claimed. My disagreement with the reader is narrow: the reader's weakest_assumption focuses on part-prior noise, whereas I identify the underspecified loss as the single most load-bearing concern, since Eq. 5 is the paper's stated methodological novelty and its current description admits of a vacuous implementation.","tokens_in":14526,"tokens_out":2604,"duration_ms":27000,"concrete_test":"Request the exact sampling procedure for S-(Ia) for unlabeled Ia and re-run Table 4 on CUB with a deterministic, explicitly enumerated negative set: (a) use all other labeled images of different classes; (b) use all other unlabeled images in the batch; (c) use only negatives whose min part similarity is below the batch positive threshold as the text implies. If the all-min loss term's gradient is identical across (a), (b), and (c), or if variant (c) leaves a fraction of batches with an empty denominator, then the ablation row 5 in Table 4 cannot be attributed to the all-min contrastive mechanism as stated, and the paper must either provide the missing rule or weaken the claim.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that APL shows significant enhancements on fine-grained GCD benchmarks via adaptive part learning, with the paper's stated novelty resting heavily on the all-min contrastive loss (Sec 3.3, Eq. 5). The loss specification for unlabeled data is incomplete: for unlabeled images, the negative set is described as 'randomly sample some images from B as negative samples, ensuring that not all parts of these images are similar to each corresponding part of Ia, and use the similarity of all positive pairs in the batch as the threshold for sampling negative images' (Sec 3.3). This is not a precise rule: there is no pseudocode, no formula for the threshold, and no statement of how many negatives are sampled or whether the pruning removes all cross-class negatives. Without this rule, the denominator in Eq. 5 can be empty in expectation when positives are highly similar, making Lall-min vacuous and leaving the reported gains attributable to Ldiv plus the host GCD loss rather than to the novel all-min objective. This is load-bearing because the ablation in Table 4 attributes +2.5 to +2.7 points to switching from a min-max loss to the all-min loss; if the negative-sampling rule makes the all-min loss reduce to a positive-only term, the mechanism claimed (adaptively constraining only the least similar negative part) is untested by the paper's experiments.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes APL, a plug-and-play part-learning module for Generalized Category Discovery (GCD). It uses DINO's final-layer attention heads as a part prior, a set of shared learnable part queries to obtain consistent image-specific part features, and a hard-assignment mechanism with Gumbel-Softmax to decompose each image into parts. The part features replace the CLS token representation in existing GCD frameworks. A new \"all-min\" contrastive loss is introduced to constrain only the least similar part of negative pairs while keeping all corresponding parts of positive pairs similar, and a diversity loss prevents part-query collapse. Experiments on four fine-grained benchmarks and three generic datasets report consistent improvements over SimGCD, SPTNet, and CMS, with ablations attributing the gains to learnable queries, hard assignment, and the all-min loss.","tokens_in":14621,"tokens_out":6138,"duration_ms":73118,"significance":"If the method is correctly specified, APL addresses a real limitation of global-CLS-token GCD methods and offers a practically valuable plug-and-play enhancement. The empirical evidence is coherent and consistent: Table 1 shows gains across all four fine-grained datasets for three different host frameworks, and Table 4 attributes the improvement to the proposed all-min loss rather than to the host GCD loss alone. The idea of sharing generalizable parts while highlighting discriminative parts is well motivated. However, the implementation details of the central equations and the unlabeled negative-sampling rule are not yet precise enough to validate the claimed mechanism or to reproduce the experiments.","major_comments":[{"comment":"Equation (3) is dimensionally inconsistent as written. Fprior is defined as [ϵ(A1)*Fpatch, ..., ϵ(AM)*Fpatch], so it has shape M×C, and Q is T×C. The expression ProjQ(Q) * ProjK(Fpart)^⊤ / √C then yields a T×M matrix if Fpart=Fprior, or a T×N matrix if Fpart=Fpatch, not the claimed T×C image-specific query QI. In addition, the symbol Fpart is undefined. Since this equation is the foundation of the part-query construction, the authors should give the correct formula with explicit tensor shapes.","section":"§3.2, Eq. (3)"},{"comment":"The hard-assignment mechanism in Eq. (4) selects a single patch for each query: one_hot(arg max_N(...)) returns exactly one nonzero entry per query in the forward pass, so each part feature Ppart equals the feature of one patch token. This contradicts the text's description that \"patches within the same query constitute a potential object part\" and the visualization in Figure 3c, which shows multi-patch regions. Unless a soft weighted aggregation is intended, the method does not discover multi-patch object parts as claimed. Please clarify whether each part is a single patch or a region, and if a region, specify how the binary assignment is formed.","section":"§3.2, Eq. (4)"},{"comment":"The negative-sampling rule for unlabeled images is underspecified. The text states that for unlabeled Ia the authors \"randomly sample some images from B as negative samples, ensuring that not all parts of these images are similar to each corresponding part of Ia, and use the similarity of all positive pairs in the batch as the threshold for sampling negative images.\" This is not a precise algorithm: it does not define the threshold mathematically, does not state how many negatives are sampled, does not specify how to handle cases where no image passes the threshold, and does not state whether augmentations of the same image are excluded. Because Table 4 attributes a gain of about 2.4–2.7 points to the all-min loss over its alternatives, this missing specification is load-bearing for the paper's central claim and must be provided.","section":"§3.3, Eq. (5)"},{"comment":"The paper reports no sensitivity analysis over the key hyperparameters of the proposed module: the number of part queries T, the threshold filter ϵ in Eq. (3), the Gumbel temperature τ in Eq. (4), and the negative-sampling threshold in Eq. (5). The authors acknowledge that DINO's part perception \"is too sensitive to variations among different images\" and claim that the learnable queries stabilize it, but no quantitative consistency or correspondence metric is reported, and no ablation shows how performance varies with T or ϵ. Without these, the central mechanism of adaptive part discovery is asserted rather than directly demonstrated.","section":"§3.2–§3.3, Tables 3–4"}],"minor_comments":[{"comment":"The main results are reported for a single run; please add standard deviations or at least a multi-seed evaluation to assess the significance of the reported improvements.","section":"§4.2, Table 1"},{"comment":"The comparison method \"min-max contrastive loss\" in row 3 is never defined in the main text; provide its equation or a precise reference so the ablation is interpretable.","section":"§4.3, Table 4"},{"comment":"In the diversity loss, the use of max(0, sim(Pt, P̂t)) means that negative cosine similarities do not contribute; please clarify whether this is intended and whether the part features are normalized before computing the cosine similarity.","section":"§3.3, Eq. (6)"},{"comment":"References [10] and [11] are the same paper (Dosovitskiy et al.); merge them into one entry to avoid duplication.","section":"References"},{"comment":"There are several typographical errors, including \"unlabled\" in §3.1, \"created auklet\" in §1, and \"gneralized\" in §4.2; these should be corrected.","section":"Throughout"},{"comment":"The manuscript refers to an Appendix for dataset details and additional ablations, but no appendix is included in the submitted text; please ensure the appendix is present and that the main text is self-contained regarding the negative-sampling algorithm.","section":"Appendix"}],"recommendation":"major_revision","confidential_remarks":"The central empirical claim is plausible and the direction is interesting, but the main text currently lacks the algorithmic precision needed to verify the mechanism. In particular, Eq. (3) has a tensor-shape problem, Eq. (4) appears to select only a single patch per query, and the unlabeled negative-sampling rule in Eq. (5) is not a well-defined procedure. These points are fixable within the scope of a revision, so I recommend major revision rather than rejection. If the appendix already contains the missing details, the authors should make those details explicit in the main text."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does something genuinely new: it brings DINO-prior part discovery into GCD via shared learnable queries, then trains the parts with an all-min contrastive loss that only constrains the least-similar part of negative pairs. That combination is not in the cited GCD literature, and the all-min objective is distinct from a standard min-max variant. The empirical core is a real strength: consistent 2-to-6 point gains over SimGCD, SPTNet, and CMS across four fine-grained datasets, with the largest gains often on novel classes. The ablations in Tables 3 and 4 are internally coherent and correctly attribute the gains to the queries, hard assignment, and all-min loss. The k-means and DINO-only baselines are sensible checks, and the plug-and-play framing is honest.\n\nThe soft spots are real but not load-bearing. First, no code and no standard deviations are shipped, so the magnitude of the deltas could be seed noise. Second, the text claims consistent outperformance on generic datasets, but Table 2 shows SelEx beating the method on CIFAR-100 All (82.3 vs 80.9). That's a minor inconsistency, but sloppy. Third, the CMS comparison relies on an author reproduction (CMS dagger) that underperforms the published CMS on CUB and Aircraft, which inflates the reported average gain. The authors need to justify the reproduction or report against published numbers. Fourth, the negative-sampling rule in Eq. 5 for unlabeled images is genuinely underspecified: \"randomly sample some images... ensuring that not all parts are similar... use positive-pair similarity as threshold\" is not a precise rule. The stress-test worry about an empty denominator is possible in principle, but the paper does sample negatives, so the loss is not vacuous by construction; the real problem is reproducibility. There's also no ablation over the number of queries T or the threshold epsilon, leaving the DINO-prior assumption untested for sensitivity.\n\nNone of this kills the central claim. The method is plausible, the experiments as reported support it, and the limitations are addressable. Who benefits? Anyone working on GCD or fine-grained open-world recognition will want to know this part-level angle exists. I'd bring it to a reading group to discuss the all-min loss and the part-discovery mechanism.\n\nRecommendation: send to peer review. Require code or variance estimates, fix the generic-dataset claim, clarify the negative-sampling rule with a formula or pseudocode, and add a sensitivity analysis over T and epsilon. With those, the paper would be solid.","headline":"A solid plug-and-play part-level enhancement for GCD with plausible gains, but reproducibility gaps and a vague negative-sampling rule keep the central claim from being fully verified.","tokens_in":15398,"tokens_out":1784,"would_cite":true,"duration_ms":20666,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that replacing the global CLS token with pooled, adaptively learned part features resolves the discriminability-versus-generalization trade-off in fine-grained generalized category discovery, and supports this with…","keywords":["generalized category discovery","fine-grained image recognition","object part discovery","all-min contrastive learning","part queries","DINO part priors","plug-and-play representation learning","self-supervised vision transformers"],"falsifier":"Take a set of fine-grained images, apply strong random cropping or occlusion, and compare APL's hard assignment $\\mathbf{H}_{\\text{part}}$ of the same image across augmentations: if a large fraction of patches switch part queries, the claimed consistency and correspondence are not present. The same test can be run as an ablation over the threshold $\\epsilon$ in Eq. (3) and the query count $T$; if no setting both stabilizes assignments and preserves accuracy, the part prior is not carrying the gain.","tokens_in":14092,"feed_emoji":"🧩","tokens_out":10886,"duration_ms":106609,"temperature":0.7,"pith_summary":"Generalized Category Discovery (GCD) asks a model to classify images from both known and novel classes using only a small labeled subset, and existing methods lean on a self-supervised DINO transformer whose global CLS token feature must balance telling similar classes apart against transferring knowledge to new classes. This paper claims that the CLS token forces a trade-off: pushing global features apart to separate similar species also pushes their shared local parts apart, hurting transfer, while superclass-level sharing blurs novel classes into similar known ones. The proposed Adaptive Part Learning (APL) discovers object parts without annotations, using shared learnable part queries steered by DINO attention priors, and learns part features with an all-min contrastive loss that pulls all corresponding parts of positive pairs together but pushes apart only the least similar corresponding part of negative pairs. Pooled part features then replace the CLS token inside existing GCD frameworks. If the claim holds, a plug-in part-level representation improves both known/novel separation and knowledge transfer on fine-grained benchmarks without retraining or re-architecting the host method.","feed_headline":"Shared part queries add up to 6.3 accuracy points in class discovery","feed_subtitle":"A plug-in that swaps the CLS token for pooled part features sharpens fine-grained known/novel separation.","key_machinery":"The load-bearing object is the set of shared learnable part queries $Q \\in \\mathbb{R}^{T \\times C}$, which act as bridges that give the same index to corresponding parts in different images. The DINO prior $F_{\\text{prior}}$ from thresholded attention-head features (Eq. 3) steers these queries; the Gumbel-Softmax hard assignment $\\mathbf{H}_{\\text{part}}$ (Eq. 4) converts patch-query similarities into binary part masks while preserving gradients; the all-min contrastive loss (Eq. 5) performs the adaptive selection of discriminative parts; and the diversity loss (Eq. 6) prevents query collapse. Together these turn an image into $T$ corresponding parts whose pooled features replace the CLS token.","core_discovery":"The paper's central claim is that object parts, rather than global image semantics, are the right unit for GCD: discriminative parts such as the head separate similar species, while shared parts such as legs keep common structure transferable. Concretely, the method extracts DINO part priors by thresholding the final-layer attention maps, transforms a set of $T$ shared learnable part queries into image-specific queries by cross-attention, hard-assigns image patches to queries through a differentiable Gumbel-Softmax step, and treats the patches assigned to each query as one consistent part. An all-min contrastive loss (Eq. 5) makes all corresponding parts of positive pairs similar and only the least-similar corresponding part of negative pairs dissimilar, so discriminative parts are sharpened while the rest stay shareable; a diversity loss (Eq. 6) keeps queries from collapsing. The pooled part feature replaces the CLS token in SimGCD, SPTNet, and CMS, yielding average gains of 3.7, 2.9, and 3.7 accuracy points across four fine-grained datasets, with larger gains generally on novel classes.","pith_inferences":["If the all-min mechanism is the real driver, the same 'hardest corresponding part only' negative constraint could transfer to other part- or prototype-based fine-grained recognition methods, which typically push all negative parts apart.","The paper does not measure part consistency quantitatively; a natural extension is to test APL's hard assignments across augmented or occluded views and check whether accuracy tracks the assignment-stability rate.","Because DINO part attention is treated as a frozen prior, a domain shift away from natural images could break the part prior; per-domain adaptation of the queries or the threshold $\\epsilon$ in Eq. (3) is a concrete testable extension.","The interpretability claim could be validated directly by evaluating the discovered part masks against human part annotations, such as CUB part keypoints, even though training is unsupervised; the paper's t-SNE visualizations suggest semantic parts but do not quantify alignment."],"forward_implications":["Plugging APL into SimGCD raises average accuracy by 3.7 points across CUB-200, Stanford Cars, FGVC-Aircraft, and Herbarium-19, with the largest single gain of 6.3 points on Stanford Cars.","Novel-class accuracy improves more than known-class accuracy in most fine-grained settings, for example +6.2 novel versus +5.7 known on Stanford Cars with SimGCD, which is the signature of better knowledge transfer from known to novel classes.","Because only the pooled part feature is swapped in, the same APL module attaches to parametric frameworks (SimGCD, SPTNet) and to the mean-shift framework (CMS) without changing their host losses beyond adding the all-min and diversity losses.","On generic datasets the gains shrink, for example +0.7 novel points on ImageNet-100, consistent with the claim that part sharing matters most when classes are visually similar and share components."],"supporting_citations":[{"why":"Supplies the DINO backbone: its CLS/patch features are the representation APL replaces, and its attention heads provide the part priors.","marker":"[4]"},{"why":"Establishes that DINO features perceive object parts, the prior APL exploits and stabilizes.","marker":"[29]"},{"why":"Defines the GCD task, the known/novel split, and the clustering-accuracy metric APL is evaluated on.","marker":"[40]"},{"why":"SimGCD, the parametric-classification baseline that APL plugs into for a 3.7-point average gain.","marker":"[43]"},{"why":"SPTNet, a spatial-prompt-tuning GCD baseline that APL plugs into for a 2.9-point average gain.","marker":"[42]"},{"why":"CMS, the mean-shift clustering baseline whose reproduced version hosts APL and gains 3.7 points on average.","marker":"[7]"},{"why":"Gumbel noise distribution used in the differentiable hard assignment of patches to part queries.","marker":"[13]"},{"why":"Gumbel-Softmax trick used alongside [13] to keep hard assignment differentiable.","marker":"[26]"}],"fun_headline_variants":["Plug-and-play part learning adds up to 6.3 accuracy points to GCD","Part-aware features lift fine-grained class discovery by up to 6.3 points","Swap CLS token for part features: up to 6.3 accuracy gain in GCD","Adaptive part queries sharpen known/novel separation, up to 6.3 points"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method depends on DINO's attention maps pointing at coherent, reusable object parts, and on the shared queries keeping those parts aligned across different images even when parts are occluded or cropped out; if that alignment fails, the hard part assignments are noise and the all-min loss is constraining random clusters.","fun_headline_variants_meta":{"raw":{"variants":["Plug-and-play part learning adds up to 6.3 accuracy points to GCD","Part-aware features lift fine-grained class discovery by up to 6.3 points","Swap CLS token for part features: up to 6.3 accuracy gain in GCD","Adaptive part queries sharpen known/novel separation, up to 6.3 points"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000448,"raw_usage":{"total_tokens":2277,"prompt_tokens":975,"completion_tokens":1302,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":591,"completion_tokens_details":{"reasoning_tokens":1211}},"tokens_in":591,"tokens_out":1302,"duration_ms":13876,"temperature":1.0,"reasoning_tokens":1211,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:52:23.336956+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of fine-grained images, apply strong random cropping or occlusion, and compare APL's hard assignment $\\mathbf{H}_{\\text{part}}$ of the same image across augmentations: if a large fraction of patches switch part queries, the claimed consistency and correspondence are not present. The same test can be run as an ablation over the threshold $\\epsilon$ in Eq. (3) and the query count $T$; if no setting both stabilizes assignments and preserves accuracy, the part prior is not carrying the gain.","supporting_citations":[{"cited_title":"Emerg- ing properties in self-supervised vision transformers","cited_arxiv_id":null,"evidence_quote":"Supplies the DINO backbone: its CLS/patch features are the representation APL replaces, and its attention heads provide the part priors."},{"cited_title":"Generalized category discovery","cited_arxiv_id":null,"evidence_quote":"Defines the GCD task, the known/novel split, and the clustering-accuracy metric APL is evaluated on."},{"cited_title":"Parametric classification for generalized category discovery: A baseline study","cited_arxiv_id":null,"evidence_quote":"SimGCD, the parametric-classification baseline that APL plugs into for a 3.7-point average gain."},{"cited_title":"Contrastive Mean-Shift Learning for Generalized Category Discovery","cited_arxiv_id":"2404.09451","evidence_quote":"CMS, the mean-shift clustering baseline whose reproduced version hosts APL and gains 3.7 points on average."},{"cited_title":"Statistical theory of extreme values and some practical applications: a series of lectures","cited_arxiv_id":null,"evidence_quote":"Gumbel noise distribution used in the differentiable hard assignment of patches to part queries."},{"cited_title":"A* sam- pling","cited_arxiv_id":null,"evidence_quote":"Gumbel-Softmax trick used alongside [13] to keep hard assignment differentiable."}],"review_version":1}