{"id":"9f2e4ffe-8f84-45df-8b1d-36b26549f2f5","arxiv_id":"2501.10695","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"HGRL clusters states and objects into homogeneous groups via a GloVe co-occurrence graph and learns group-specific visual and prompt representations, improving compositional zero-shot recognition on three benchmarks.","lead":"This paper proposes HGRL, a method for compositional zero-shot learning that groups similar states and objects into clusters, learns separate visual and text representations per group, and uses them to recognize unseen combinations. It reports state-of-the-art results on three standard benchmarks, with the largest gains on UT-Zappos and C-GQA.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The group construction rests on GloVe cosine similarity as a proxy for visual group membership (Eq. 4), but the paper never validates that proxy; if it is noisy, the claimed grouping benefit is not established.","rationale":"The reader's weakest-assumption analysis identifies the same load-bearing assumption: GloVe cosine similarity between category names is used as a proxy for membership in the same perceptual homogeneous group. I agree that this is the most scientifically central risk. The paper's central narrative is that HGRL improves transferability by adaptively discovering and aggregating categories with shared properties; Eq. (4) is the concrete instantiation of that discovery, and the w/o TCPG ablation is an indirect test at best. The concern is not that disagreement with consensus is itself a flaw; the issue is internal to the construction: if the text-based graph does not track visual similarity, then the claimed mechanism is not doing the work attributed to it. The paper does provide useful ablation evidence and visualizations, and the reported improvements are compatible with the group idea, so the appropriate response is not rejection but an explicit validation or replacement of the GloVe proxy. This reinforces the existing CONDITIONAL verdict rather than changing it; hence the verdict recommendation is UNCHANGED.","tokens_in":13102,"tokens_out":5536,"duration_ms":62581,"concrete_test":"Replace the GloVe adjacency in Eq. (4) with a visual k-NN adjacency computed from mean CLIP image features of the object/state categories, keeping the same threshold and training protocol, and rerun the closed-world UT-Zappos and C-GQA experiments. If the visual adjacency gives comparable or better AUC, the text proxy is not necessary and the claimed grouping mechanism is not validated; if it gives much worse AUC, the authors should demonstrate that GloVe co-occurrence, not arbitrary category proximity, is driving the gain. Additionally, report 3-seed means and standard deviations for the final Table 1 entries to assess whether the reported margins, especially the 0.8 AUC gain on MIT-States, are significant.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing condition for the central claim is that the groups formed by Eq. (4) are coherent in a visual, not merely textual, sense. A_ij = M^o_ij * I[M^o_ij >= 0.5] * I[s_i = s_j] uses cosine similarity of GloVe word embeddings as a stand-in for whether two categories belong to the same homogeneous perceptual group. This graph enters both GA VR, where visual features are aggregated across samples, and indirectly GAPE, where route confidences determine which samples may augment each other. If GloVe similarity is high for textually co-occurring but visually dissimilar categories, or low for visually similar ones, the aggregated features become noisy and the transferability benefit disappears. The paper's only evidence for the proxy is the w/o TCPG ablation (UT-Zappos closed-world AUC drops from 46.8 to 43.1; C-GQA from 16.3 to 14.4). That drop is consistent with the graph helping, but it does not show that the help comes from correct grouping rather than from any generic feature-aggregation regularization. The concern is compounded by an unspecified implementation detail: UT-Zappos object labels are multiword phrases such as 'Shoes.Sneakers.and.Athletic.Shoes', and the paper does not state how GloVe embeddings are obtained for these labels when computing M^o_ij.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Homogeneous Group Representation Learning (HGRL) for compositional zero-shot learning, motivated by the claim that existing all-to-one and one-to-one representation paradigms create a transferability-discriminability trade-off. HGRL has three components: Group-Aware Visual Representation (GA VR), which uses a mixture-of-experts network with a GloVe-based text co-occurrence graph to form homogeneous groups and aggregate visual features; Decoupled Group Prompt (DGP), which learns a separate prompt context for each group; and Group-Aware Pair Enhancement (GAPE), which augments state and object features using group-compatibility-weighted pairs. The method is evaluated in closed- and open-world settings on MIT-States, UT-Zappos, and C-GQA with a CLIP ViT-L/14 backbone, reporting state-of-the-art AUC and HM numbers and ablations over the three components.","tokens_in":13444,"tokens_out":9416,"duration_ms":90402,"significance":"If the reported results hold, HGRL is a useful contribution to CZSL: the grouping perspective is clearly different from the all-to-one/one-to-one dichotomy, the three modules are coherent and each is supported by the Table 2 ablations, and the gains on UT-Zappos and C-GQA are large enough to be meaningful even without formal tests. The paper also provides a fair comparison with recent CLIP-based methods on three standard benchmarks and includes visualization analyses. However, the absence of repeated-run statistics, the undocumented hyperparameter selection, the unvalidated GloVe-based grouping proxy, and the loose theoretical section mean the central claim of state-of-the-art performance is not yet fully established.","major_comments":[{"comment":"All results in Table 1 and Table 2 appear to be single-run point estimates: no standard deviations, seeds, or significance tests are reported, and the text does not state how many runs each number is averaged over. This is load-bearing because the closed-world MIT-States improvement over the second-best method is only 0.8 AUC (23.3 vs 22.5) and 0.9 HM (40.2 vs 39.3), a margin that can easily be within run-to-run noise for CLIP-based training. Please report mean and standard deviation over multiple seeds, or otherwise provide significance evidence for the claimed state-of-the-art results.","section":"Table 1 and Section 4.2"},{"comment":"Hyperparameter selection is not reproducible as described. In Section 4.1, λ is set to 1.0 for UT-Zappos and 0.1 for MIT-States and C-GQA, while ks and ko are 3 for UT-Zappos and 5 for the other datasets, and no validation split or selection criterion is given; additionally, the top-K expert filtering used in Eq. (2) is never specified. Because the test sets include unseen compositions, tuning these values on the test set would inflate the reported numbers. The paper should state the validation protocol and report results for the chosen values, ideally with sensitivity curves that include error bars.","section":"Section 4.1 and Eq. (2)"},{"comment":"Eq. (4) constructs the entire grouping mechanism from GloVe cosine similarity: A_ij = M^o_ij * I[M^o_ij >= 0.5] * I[s_i = s_j], with M^o_ij = cos(v_i^o, v_j^o). The paper asserts that word co-occurrence probability 'largely matches' membership in the same perceptual superclass, but the only empirical support is the w/o TCPG ablation in Table 2, which shows only that removing the graph hurts performance. It does not show that the GloVe groups are visually coherent; a generic feature-aggregation regularizer could produce a similar drop. Please validate the proxy directly (for example, by comparing GloVe-based group adjacency with clusters in the visual feature space or with group annotations), and also state how GloVe embeddings are computed for multiword UT-Zappos labels such as 'Shoes.Sneakers.and.Athletic.Shoes'.","section":"Section 3.2, Eq. (4)"},{"comment":"Section 3.6 is not a rigorous derivation. The notation d(S,U | p(...)) is not defined, the claimed decomposition of d(S,U) into conditional divergences is asserted rather than proved, and the conclusion d(S*,U) <= d(S,U) does not follow without assumptions on the domains and the group-conditional measures. Since this section is explicitly titled 'Theoretical Insights,' it should either be made into a proper theorem with stated assumptions and proof or be recast as motivation with the 'theoretical' claim softened.","section":"Section 3.6, Eq. (24)"},{"comment":"Eq. (21), the inference formula for pstate(s|xs), is inconsistent with the training loss in Eq. (12): the numerator in Eq. (12) is weighted by the route confidence p^{Sj}_r, but the numerator in Eq. (21) is unweighted and contains a stray dot. The object-branch inference likely has the same issue. Please correct the formulas and verify that the inference used for the reported numbers matches the trained objective.","section":"Section 3.5, Eq. (21)"}],"minor_comments":[{"comment":"The abstract and introduction contain several grammatical errors, such as 'Conditional dependency present one of the trickiest problems' and 'a new perspective formulates state (object) representation learning'; these should be copy-edited.","section":"Abstract and Introduction"},{"comment":"There are several typos: 'memebers' in Section 3.2, 'Visualizaiton' in the Fig. 1 caption, 'there datasets' in the Table 1 caption, and 'CGQ' instead of C-GQA in Section 4.2.","section":"Section 3.2, Fig. 1, Table 1, Section 4.2"},{"comment":"Eq. (5) calls Λ a 'unit diagonal matrix'; this should be the identity matrix.","section":"Eq. (5)"},{"comment":"In Eq. (2), the sum over the selected K experts is written without an index set; please make explicit that the sum runs over the top-K experts selected by the route scores.","section":"Eq. (2)"},{"comment":"The symmetric object-branch analogue of Eq. (4), which should use M^s_ij and I[o_i = o_j], is not written out; please state it explicitly.","section":"Section 3.2"},{"comment":"Fig. 3(a)-(b) report 'state accuracy' and 'object accuracy' without defining these metrics, and the λ-sensitivity plots in Fig. 3(c)-(d) have no error bars and no legend indicating which curve is which.","section":"Fig. 3"},{"comment":"The term 'text co-occurrence probability graph' is misleading because Eq. (4) uses a cosine similarity threshold, not an estimated co-occurrence probability; align the terminology with the actual computation.","section":"Section 3.2, Eq. (4)"},{"comment":"ProCC [Huo et al., 2024] appears in the reference list but is never cited in the main text.","section":"References"},{"comment":"The ablation study in Table 2 omits MIT-States; since the main claim covers all three datasets, it would be informative to include MIT-States ablations as well.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The paper's authors have prior CZSL work (DFSP and ProCC) with the same corresponding author; ProCC is in the reference list but not cited in the body, and DFSP is a direct baseline. This is not itself a flaw, but the editor may wish to check novelty disclosure and related-work completeness. I also note that no code release is mentioned; given the number of unspecified details (top-K, multiword GloVe embeddings, validation protocol), code release would materially improve verifiability."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read: this is a solid, benchmark-driven CZSL paper. The idea is that instead of all-to-one or one-to-one representations, you discover latent homogeneous groups and learn group-specific visual and prompt representations. That's a sensible middle ground. What's actually new is combining a GloVe co-occurrence graph as a pseudo-hierarchy with MoE routing and per-group prompts, and applying it to CZSL. The three components each do something identifiable, and the ablations show they all contribute.\n\nThings it does well: the method is clearly described, the loss equations are coherent, and the ablation table is honest. The gains on UT-Zappos and C-GQA are nontrivial. The text co-occurrence graph is a cheap substitute for missing hierarchical labels, and the paper does not oversell it as a labeled hierarchy. There's no obvious circularity: the benchmark numbers are measured, not derived from the claim.\n\nSoft spots: the GloVe cosine similarity is the load-bearing assumption, and the paper never validates that it tracks perceptual similarity. The ablation 'w/o TCPG' only shows that removing it hurts, not that the groups are correct. Also, UT-Zappos labels are multiword phrases; the paper doesn't say how GloVe embeddings are computed for those. Minor issues: single-run results with no error bars, hyperparameters set per dataset with no stated validation split, and the MIT-States closed-world margin is 0.8 AUC. The theory section is a post hoc domain-adaptation sketch, not a derivation.\n\nNone of this sinks the central idea. The method is reproducible in principle and the components make sense. But as it stands, the empirical evidence is not as strong as the 'significant margin' language in Section 4.2 suggests.\n\nBottom line: deserve a serious referee. If the authors release code with seeds, report standard deviations, and clarify the GloVe embedding for multiword labels, this is a publishable contribution.","headline":"HGRL is a credible middle-ground CZSL model with consistently strong benchmark numbers; the main weakness is that its GloVe-based grouping proxy is under-validated and the results are single-run.","tokens_in":13927,"tokens_out":1851,"would_cite":true,"duration_ms":17154,"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":"Grouping states and objects into homogeneous sub-groups, each with its own learned representation center, balances transferability and discriminability in compositional zero-shot learning and sets new state-of-the-art scores on three…","keywords":["Compositional Zero-Shot Learning","Homogeneous Group Representation Learning","Mixture-of-Experts","Vision-Language Models","Prompt Learning","GloVe word embeddings","Transferability and Discriminability","Domain Adaptation"],"falsifier":"One decisive experiment: replace the GloVe adjacency in Eq. (4) with an adjacency computed from visual similarity of held-out CLIP image features (or from human-annotated superclass labels) and retrain HGRL; if the visual- or label-guided version does not match or beat the GloVe version, then word-embedding similarity is not doing the work the paper claims. Alternatively, one can inspect the discovered groups for a dataset such as MIT-States and measure intra-group visual similarity of the learned representations: if the groups are not internally visually consistent, the grouping assumption fails.","tokens_in":12892,"feed_emoji":"🧩","tokens_out":10053,"duration_ms":82714,"temperature":0.7,"pith_summary":"The paper argues that the two dominant paradigms for compositional zero-shot learning—one shared representation per state or object ('all-to-one') versus a separate representation per state-object pair ('one-to-one')—each sacrifice either transferability or discriminability. It proposes Homogeneous Group Representation Learning (HGRL), which clusters states and objects into a small number of homogeneous sub-groups, learns a distributed representation center per group, and adaptively aggregates features through a mixture-of-experts router guided by word-embedding similarities. On the MIT-States, UT-Zappos, and C-GQA benchmarks, HGRL reports higher area-under-the-curve (AUC) and harmonic-mean (HM) accuracy than prior CLIP-based methods, with the largest closed-world gain on UT-Zappos (46.8 AUC versus 44.5 for the previous best). A sympathetic reader would care because the result suggests that the transferability–discriminability tradeoff in CZSL is not a zero-sum game: a few learned group centers can carry both shared and group-specific semantics.","feed_headline":"Homogeneous-group prototypes beat global and per-pair learning","feed_subtitle":"Reports state-of-the-art closed- and open-world AUC on all three benchmarks.","key_machinery":"The load-bearing mechanism is the adaptive grouping loop: a text co-occurrence probability graph (TCPG) derived from GloVe cosine similarity among category names, thresholded at 0.5, defines an adjacency matrix over a batch of samples (Eq. 4); the route network in the Group-Aware Visual Representation (GAVR) module assigns each sample to the K most confident experts, and group-specific features are a weighted sum of those experts' outputs fused with the original feature (Eqs. 1–3). The same route confidences are reused in Group-Aware Pair Enhancement (GAPE) to gate which samples may augment each other (Eq. 14), and the Decoupled Group Prompt (DGP) appends a learnable per-group prompt token so each group has its own text prototype rather than one shared prototype per category. The theoretical section frames the design as minimizing a domain-adaptation upper bound by increasing the probability that an unseen domain's homogeneous group matches the seen group that transfers knowledge to it.","core_discovery":"The central claim is that state and object representations in compositional zero-shot learning should be learned as multiple homogeneous sub-group representations rather than as a single global prototype or as per-instance dependencies. A route network assigns each sample to a small set of experts, and a text co-occurrence probability graph built from GloVe embeddings of category names provides soft supervision for which categories belong together by aggregating features of compatible samples. Each group receives its own learnable prompt context (Decoupled Group Prompt), and Group-Aware Pair Enhancement mixes state and object features only for pairs whose group-confidence vectors are compatible. The paper reports state-of-the-art closed-world AUC of 23.3 on MIT-States, 46.8 on UT-Zappos, and 16.3 on C-GQA, with corresponding harmonic means of 40.2, 59.0, and 34.6, and open-world AUC gains of up to 4.2 points over the previous best on UT-Zappos.","pith_inferences":["If the paper is right, replacing the GloVe-based adjacency with a vision-derived grouping (e.g., clustering CLIP visual features of seen images) could make the method work on datasets where word co-occurrence and perceptual similarity diverge; the paper does not run that comparison.","The theory suggests a principled way to choose the number of groups: stop increasing ks and ko when validation AUC saturates, rather than fixing them at 3 or 5; the sensitivity plots in Fig. 3 hint at this but do not test it.","The compatibility gate in GAPE could be applied symmetrically to object-state pairs in the other direction or extended to multi-hop neighbors in the adjacency graph, which the paper does not explore.","Because the method builds on frozen CLIP features plus GloVe, it is a drop-in module for other CZSL baselines; one could measure the marginal gain of adding HGRL to a non-CLIP method, which the paper does not do."],"forward_implications":["The transferability–discriminability tradeoff in CZSL is not a zero-sum game: a small number of learned group centers capture both shared and group-specific semantics, which is why HGRL beats both the all-to-one and one-to-one baselines.","Group-level prompts provide a supervision-free way to inject structured prior knowledge into vision-language models, since the grouping signal comes from word co-occurrence statistics rather than hierarchical labels.","The adjacency construction (GloVe similarity thresholded at 0.5) is reusable across datasets and requires no extra annotation, making the method easy to adapt to other attribute-object tasks.","The largest reported gains appear on UT-Zappos, a small dataset with strong visual grouping structure, suggesting the benefit scales with how cleanly categories cluster in feature space.","Fusing backbone, pair, and state × object branch scores at inference (Eq. 22) yields consistent improvements over any single branch, so the components are complementary."],"supporting_citations":[{"why":"Supplies the frozen CLIP visual and text encoders that all branches (backbone, state, object, pair) build on.","marker":"[Radford et al., 2021]"},{"why":"Provides GloVe word embeddings whose cosine similarity defines the text co-occurrence probability graph adjacency in Eq. (4).","marker":"[Pennington et al., 2014]"},{"why":"Gives the domain-adaptation upper bound that the paper decomposes to argue group-conditioned transfer reduces distributional divergence.","marker":"[Ben-David et al., 2006]"},{"why":"Introduces the mixture-of-experts architecture that underlies the route network and expert sub-networks in GAVR.","marker":"[Jacobs et al., 1991]"},{"why":"Formulates the task of inferring analogous attributes, the problem setup for compositional zero-shot learning.","marker":"[Chen and Grauman, 2014]"},{"why":"Provides the Troika baseline and evaluation protocol that HGRL compares against on the three datasets.","marker":"[Huang et al., 2024]"},{"why":"Provides the Retri-Aug baseline, the previous best reported numbers on the same benchmarks that HGRL claims to surpass.","marker":"[Jing et al., 2024]"},{"why":"Supplies the MIT-States dataset used for the main closed- and open-world results.","marker":"[Isola et al., 2015]"},{"why":"Supplies the UT-Zappos dataset used for the main closed- and open-world results.","marker":"[Yu and Grauman, 2014]"},{"why":"Supplies the C-GQA dataset used for the main closed- and open-world results.","marker":"[Naeem et al., 2021]"}],"fun_headline_variants":["Homogeneous groups beat global and per-pair for zero-shot composition","HGRL: subgroup prototypes outperform all-to-one and one-to-one in CZSL","Subgroup prototypes, not global or per-pair, hit new SOTA in zero-shot composition","Adaptive homogeneous groups yield SOTA in zero-shot composition"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the cosine similarity between GloVe word embeddings of two category names (for instance 'dog' and 'tiger') is a reliable guide to whether images of those categories in the same state look alike, so that aggregating visually compatible samples across groups transfers knowledge correctly.","fun_headline_variants_meta":{"raw":{"variants":["Homogeneous groups beat global and per-pair for zero-shot composition","HGRL: subgroup prototypes outperform all-to-one and one-to-one in CZSL","Subgroup prototypes, not global or per-pair, hit new SOTA in zero-shot composition","Adaptive homogeneous groups yield SOTA in zero-shot composition"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001547,"raw_usage":{"total_tokens":6176,"prompt_tokens":927,"completion_tokens":5249,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":543,"completion_tokens_details":{"reasoning_tokens":5165}},"tokens_in":543,"tokens_out":5249,"duration_ms":36748,"temperature":1.0,"reasoning_tokens":5165,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T19:01:49.368038+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"One decisive experiment: replace the GloVe adjacency in Eq. (4) with an adjacency computed from visual similarity of held-out CLIP image features (or from human-annotated superclass labels) and retrain HGRL; if the visual- or label-guided version does not match or beat the GloVe version, then word-embedding similarity is not doing the work the paper claims. Alternatively, one can inspect the discovered groups for a dataset such as MIT-States and measure intra-group visual similarity of the learned representations: if the groups are not internally visually consistent, the grouping assumption fails.","supporting_citations":[{"cited_title":"Learning graph embeddings for compositional zero-shot learning","cited_arxiv_id":null,"evidence_quote":"Supplies the C-GQA dataset used for the main closed- and open-world results."},{"cited_title":"Glove: Global vec- tors for word representation","cited_arxiv_id":null,"evidence_quote":"Provides GloVe word embeddings whose cosine similarity defines the text co-occurrence probability graph adjacency in Eq. (4)."},{"cited_title":"Analysis of representations for domain adaptation","cited_arxiv_id":null,"evidence_quote":"Gives the domain-adaptation upper bound that the paper decomposes to argue group-conditioned transfer reduces distributional divergence."},{"cited_title":"Inferring analogous attributes","cited_arxiv_id":null,"evidence_quote":"Formulates the task of inferring analogous attributes, the problem setup for compositional zero-shot learning."},{"cited_title":"Troika: Multi-path cross-modal traction for compositional zero- shot learning","cited_arxiv_id":null,"evidence_quote":"Provides the Troika baseline and evaluation protocol that HGRL compares against on the three datasets."},{"cited_title":"Retrieval-augmented primitive representa- tions for compositional zero-shot learning","cited_arxiv_id":null,"evidence_quote":"Provides the Retri-Aug baseline, the previous best reported numbers on the same benchmarks that HGRL claims to surpass."},{"cited_title":"Discovering states and transformations in im- age collections","cited_arxiv_id":null,"evidence_quote":"Supplies the MIT-States dataset used for the main closed- and open-world results."}],"review_version":1}