acceptodds
Under review as a conference paper at ICLR 2027

HyCoPro: Hyperbolic Compositional Prototypes for Compositional Zero-Shot Learning

Abstract

Compositional Zero-Shot Learning (CZSL) aims to recognize unseen state–object compositions by learning from seen compositions. Recent CLIP-based methods have improved CZSL by turning each candidate pair into a text prompt, but the resulting composition prototypes are still often optimized as independent labels. This becomes fragile in open-world CZSL, where the model must score all possible compositions, many of which have no corresponding training images. To address this problem, we propose **HyCoPro**, a hyperbolic compositional closure framework for CZSL. HyCoPro organizes CLIP visual and textual features as role-specific state, object, and composition prototypes in a Poincaré ball, where Möbius addition closes primitive prototypes into composition prototypes. This closure is enforced from both visual and textual sides, strengthening unseen prototypes without requiring extra images. We further provide theoretical analysis explaining why HyCoPro is well suited to compositional recognition in CZSL. Experiments on three benchmarks demonstrate consistent gains, particularly under open-world evaluation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Compositional Zero-Shot Learning, Hyperbolic Representation Learning, Prototype Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_kcpCOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HyCoPro, which projects visual and textual features into role specific representations in the Poincaré ball that correspond respectively to states, objects, and compositions, and uses Möbius addition to construct compositional prototypes from primitive prototypes. The visual textual closure loss aligns the combined visual primitives with the text prototype obtained by encoding the complete prompt, while the bidirectional text text closure loss aligns, among candidate compositions, the prototype obtained by combining text primitives with the explicit compositional prototype. During training, the method uses Euclidean and hyperbolic classification losses, orthogonal regularization, and counterfactual regularization. At inference time, the model fuses explicit prototypes with prototypes formed by combining primitives, and primitive scores with composition scores, and further introduces Euclidean scores and feasibility priors depending on the dataset. The paper conducts closed world and open world evaluations on three standard datasets and presents three theoretical propositions. The method is plausible, and the reported MIT-States results are encouraging, but the submission has not yet sufficiently established its unique geometric contribution.

Critical strengths
  1. The paper addresses a relevant and valuable limitation in composition-level recognition in the CZSL field. Explicitly establishing a connection between primitive representations and compositional prototypes is a reasonable route for recognizing compositions without training images. Emphasizing open-world CZSL also fits the characteristics of the problem, since the expansion of the candidate set makes prototype quality and seen-unseen calibration more important.

  2. The two closure objectives provide a coherent mechanism for connecting primitives and compositions. The visual-textual and textual-textual objectives in Equations (3) to (6) connect complementary sources of supervision. In particular, the textual-side closure can regularize candidate compositions without introducing additional images. Retaining both the full-prompt prototype and the primitive-combination prototype also allows the model to combine holistic semantics with reusable semantic information.

  3. The reported improvement on MIT-States is relatively large. In Table 1, open-world HM increases from 22.4 for the strongest baseline in the table to 27.3, and AUC increases from 8.7 to 12.5. These improvements are commendable. Conducting experiments on three datasets under two evaluation settings also provides valuable evaluation breadth, although baseline coverage still needs to be corrected.

Critical weaknesses
  1. The contribution of the method is not clear. Troika [R1] adopts three-branch modeling for states, objects, and compositions. FlowComposer [R2] has studied explicit composition in embedding space. Pal et al. [R3] have studied compositional vision-language learning in hyperbolic space. These methods differ from this paper, but they make the contribution of this paper more focused on the specific closure form and the advantages verified for it. The introduction uses spatial packing capacity to argue for hyperbolic geometry and uses non-commutativity to argue for Möbius addition. However, neither argument proves that benchmark performance is limited by the packing capacity of Euclidean space, nor does either prove that this particular asymmetry can capture state-object semantics better than a learnable, role-aware Euclidean operator. The role-specific projection head itself already distinguishes the two types of input. The results in Figure 1 are noteworthy, but the definition of the operator and the control conditions in it are insufficiently explained and cannot yet resolve this issue. The paper needs to explain more clearly what contribution the proposed geometric structure brings beyond structured multi-branch learning and compositional consistency.

  2. The theoretical analysis does not substantially support the core claim about unseen composition recognition. Propositions 1-3 provide background intuition and conditional score relationships, rather than an explanation or guarantee of compositional generalization for this method.

  3. The ablation experiments do not sufficiently isolate the sources of improvement, nor do they explicitly verify the role of closure under the candidate scale emphasized in the paper. Table 2 removes inference scores, and removing hyperbolic scores is not equivalent to evaluating a Euclidean model retrained under matched adaptation, primitive supervision, closure losses, and regularization. The authors should conduct a controlled comparison at the training level. Likewise, the closure loss ablation is conducted only on UT-Zappos. In Figure 1, the full candidate set for this dataset contains only 192 compositions, whereas the larger datasets use sampled textual-side closure. On UT-Zappos, after removing text-to-text closure, open-world HM/AUC changes only from 53.4/37.8 to 53.0/37.1. These results support that closure yields gains in this setting, but they do not establish that it can explain the larger open-world improvements. The authors should conduct a matched Euclidean training comparison, closure ablations on larger benchmarks, and a control with and without the feasibility prior to resolve these attribution issues.

  4. The results reported in the paper are attractive on MIT-States; however, important baselines are missing from the compared methods, including ClusPro (ICLR25) [R4], FlowComposer (CVPR26) [R2], and EVA (ICLR2026) [R5]. After including these baselines, the authors do not appear to be highly competitive on the two benchmarks UT-Zappos and C-GQA. It is recommended that the authors include these baselines and re-evaluate the results.

  5. The writing of the entire paper should undergo a complete revision, especially Figure 1 and the writing and layout of the Method section. The presentation in Figure 1 is relatively unclear.

[R1] Huang, Siteng, et al. "Troika: Multi-path cross-modal traction for compositional zero-shot learning." 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2024.

[R2] He, Zhenqi, Lin Li, and Long Chen. "FlowComposer: Composable Flows for Compositional Zero-Shot Learning." 2026 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2026.

[R3] Pal, Avik, et al. "Compositional entailment learning for hyperbolic vision-language models." International Conference on Learning Representations. Vol. 2025. 2025.

[R4] Qu, Hongyu, et al. "Learning clustering-based prototypes for compositional zero-shot learning." International Conference on Learning Representations. Vol. 2025. 2025.

[R5] Zhang, Xiao, et al. "Decoupling primitive with experts: Dynamic feature alignment for compositional zero-shot learning." The Fourteenth International Conference on Learning Representations. 2026.

Questions for authors

Please address the above issues.

3 comments
AuthorsOpenReview ↗

Rebuttal by Authors

We thank the reviewer for recognizing our problem formulation, dual closure objectives, and open-world gains. Below we address Q1–Q5.

Q1: Unique contribution beyond [R1, R2, R3], operator comparison, and Fig. 1 protocol.

A1: Troika [R1] uses three-branch alignment, FlowComposer [R2] learns Euclidean velocity transport, and Pal et al. [R3] model entailment cones. Complementing them, HyCoPro formulates non-commutative algebraic closure (tsh⊕ctoh≈ts,oht_s^h \oplus_c t_o^h \approx t_{s,o}^h) in the Poincare ball via visual-text and text-text supervision.

  • Learnable role-aware Euclidean control: Please see A2 to Reviewer 6X9h for a matched training-level comparison on all three datasets. Parameter-free Mobius addition (Variant A) outperforms both a learnable tangent-space ordered [state, object] composer in hyperbolic space (Variant B, +0.53M params) and a full ordered Euclidean model retrained from scratch (Variant C), confirming gains beyond role-specific heads and learnable Euclidean composition.
  • UT-Zappos protocols (Supplementary Material A): Fig. 1(a) is a geometry-only diagnostic: two composition-only models (Poincare geodesic distance vs. L2-normalized cosine similarity) are trained with identical CLIP backbones, LoRA, and 512-d heads, without primitive branches, composition operators, closures, or priors. At test time, candidate size K∈{10,25,50,100,150,192}K \in \{10, 25, 50, 100, 150, 192\}; each nested set includes the ground-truth pair, prioritizes the 26 hard negatives sharing its state/object, and fills remaining slots with other pairs (5 sampling seeds), showing stronger hyperbolic discrimination as confusable candidates grow. Fig. 1(b) trains four operators from scratch in the same multi-branch framework (retaining role-specific heads, primitive supervision, and both closures over all 192 pairs; γ=0.1,α=0.5,β=1\gamma=0.1, \alpha=0.5, \beta=1, no prior), yielding three-seed mean HM/AUC: Ord. Mobius x⊕cyx \oplus_c y (51.8/35.9), Sym. Mobius exp⁡0c((log⁡0c(x⊕cy)+log⁡0c(y⊕cx))/2)\exp_0^c((\log_0^c(x \oplus_c y) + \log_0^c(y \oplus_c x))/2) (45.2/28.7), Euclidean N(x+y)\text{N}(x+y) (47.9/32.9), and Ord. Euclidean N(LN(W[x;y]+b))\text{N}(\text{LN}(W[x;y]+b)) (49.4/34.4).

Q2: Theoretical connection to unseen composition recognition.

A2: Please see A3 to Reviewer 6X9h, where we connect Mobius closure to unseen generalization via the hyperbolic triangle inequality. Smoothness and closure of ⊕c\oplus_c ensure that accurate primitive alignment and small closure residuals bound an unseen image feature's distance to its explicit prototype, supporting Propositions 2–3 without unseen training images.

Q3: Matched Euclidean retraining, closure ablations on large benchmarks, and feasibility prior control.

A3: All three controls are in our responses to Reviewers wQAb and 6X9h:

  1. Matched Euclidean retraining: Please see A2 to Reviewer 6X9h (Variants A, B, C), where HyCoPro outperforms matched Euclidean retraining on all three datasets.
  2. Closure ablations on large benchmarks: Please see A2 to Reviewer wQAb for ablations on MIT-States (28,175 pairs) and C-GQA (278,362 pairs), where removing Text-Text Closure drops OW HM by 0.82%/0.76%, and removing both drops OW HM by 2.07%/1.79% (and C-GQA CW HM by 3.28%).
  3. Feasibility prior control: Please see A3 to Reviewer wQAb, where HyCoPro without the prior (26.22/11.78 on MIT-States; 15.32/4.69 on C-GQA) remains above all baselines in the original Table 1.

Q4: Comparison with ClusPro [R4], FlowComposer [R2], and EVA [R5].

A4: We compare with [R4, R2, R5] below (HM / AUC):

MethodMIT-States CWMIT-States OWUT-Zappos CWUT-Zappos OWC-GQA CWC-GQA OW
ClusPro [R4]40.7 / 23.823.0 / 9.358.5 / 46.654.1 / 39.532.8 / 14.911.6 / 3.0
FlowComposer [R2]40.2 / 23.520.3 / 7.558.6 / 46.851.0 / 35.534.0 / 15.912.6 / 3.5
EVA [R5]41.0 / 24.022.8 / 9.460.2 / 50.254.2 / 40.236.9 / 18.817.9 / 5.6
HyCoPro (Ours)41.9 / 25.327.3 / 12.560.2 / 47.853.4 / 37.833.7 / 15.416.2 / 4.9
  • Vs. ClusPro [R4] & FlowComposer [R2]: HyCoPro achieves higher HM/AUC in 5 of 6 settings against each method, especially in OW (27.3/12.5 vs. 23.0/9.3 and 20.3/7.5 on MIT-States; 16.2/4.9 vs. 11.6/3.0 and 12.6/3.5 on C-GQA).
  • Vs. EVA [R5]: HyCoPro scores higher on MIT-States (CW 41.9/25.3 vs. 41.0/24.0; OW 27.3/12.5 vs. 22.8/9.4) and matches UT-Zappos CW HM (60.2). While EVA scores higher on UT-Zappos OW and C-GQA, it inserts 9 MoE experts (1 shared + 8 routed MLPs) into every layer of both CLIP image and text encoders, requiring 36.7M trainable parameters (Table 8 in [R5]). HyCoPro freezes the text encoder and uses visual LoRA with role-specific projection heads.

Q5: Revision of Figure 1 and Method section.

A5: We clarified the Fig. 1 caption and revised Sections 3.3–3.6.

Reviewer_kcpCOpenReview ↗

Response to the Authors

I thank the authors for their response. I acknowledge that the authors have added fair comparison experiments with Euclidean methods, expanded the ablation studies, and updated the performance comparisons with recent methods. These results partially address my previous concerns about the effectiveness of the individual components of the method. Therefore, I am willing to raise my score from 1 to 2.

However, I still have some concerns. First, the authors still do not provide a sufficiently detailed analysis explaining why HyCoPro achieves particularly strong performance on MIT-States but fails to demonstrate comparable advantages on UT-Zappos and C-GQA. Moreover, the rebuttal does not provide convincing evidence of a consistent performance advantage over recent methods, such as EVA (ICLR'26) and VP-CMJL (ICCV'25). Considering that MIT-States is a dataset with relatively high label noise, the current results are not yet sufficient to fully demonstrate that the method has a stable and consistent advantage across different compositional recognition scenarios. Given that this paper emphasizes the advantages of hyperbolic geometry in large-scale compositional spaces, I suggest that the authors further provide experimental results on additional benchmarks such as VAW-CZSL, or conduct a more comprehensive analysis of candidate composition scale on C-GQA. Second, although the authors addressed the relevant issues in their response, I still believe that the overall writing and presentation of the current paper are not sufficiently polished. In particular, Figure 2 contains too many complex modules and connections, making the overall framework difficult to understand intuitively. In addition, the main text spends considerable space discussing theoretical propositions, but these analyses provide relatively limited help in explaining the actual generalization behavior of the method on unseen compositions. I suggest that the authors further streamline the related content and devote more space to clearly explaining the core design of the method or to supplementing relevant experimental content.

AuthorsOpenReview ↗

Rebuttal by Authors

We thank Reviewer kcpC for raising the score and for the constructive follow-up feedback. Below, we provide the requested candidate-scale analysis on C-GQA and outline our plan to further streamline the presentation.

1. Candidate-scale analysis on C-GQA and dataset-dependent gains

Following your suggestion, we evaluated full HyCoPro (A) and its matched Euclidean counterpart (C) on the same 5,098 C-GQA test images. We used fixed epoch-3 checkpoints from the seed-42 sensitivity run for A and the seed-42 Euclidean control run for C. Starting from the 6,515-pair closed-world vocabulary, we progressively added distractor compositions up to the full 278,362-pair open-world space. Throughout the sweep, both models used identical nested candidate pools, fixed inference weights, and the same open-world feasibility prior.

CandidatesA HM / AUCC HM / AUCA−C HMA−C AUC
6,51533.26 / 15.0732.52 / 14.80+0.74+0.27
10,00031.95 / 13.8831.39 / 13.74+0.56+0.14
30,00027.59 / 10.6227.23 / 10.56+0.36+0.06
100,00021.65 / 7.4421.41 / 7.25+0.25+0.19
278,36215.93 / 4.8615.58 / 4.62+0.35+0.23

Differences are computed before rounding.

As the candidate pool grows, HM and AUC decline for both models. HyCoPro achieves higher mean HM and AUC at every tested scale, although the gaps are modest and do not widen consistently. At the full C-GQA scale, HyCoPro retains an advantage of 0.35 HM points and 0.23 AUC points over the matched Euclidean model. This experiment shows how the two models respond to a larger candidate pool within C-GQA. The extent to which candidate count explains the larger gains on MIT-States remains open. We will include this analysis and a concise discussion of dataset-dependent behavior in the revised manuscript.

2. Simplifying Figure 2 and streamlining the theory section

We appreciate your specific suggestions on the paper's presentation. In the revision, we will simplify Figure 2 by reducing visual clutter and explicitly highlighting the core visual–text and text–text closure paths so that the overall architecture is easier to follow at a glance. We will also condense the theoretical propositions in the main text, deferring extended details to the supplementary material, and reallocate the saved space to a clearer explanation of our core design and the additional experimental analyses.

Reviewer_6X9hOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HyCoPro, which addresses the compositional zero-shot learning problem by mapping CLIP-based state, object, and compositional representations into hyperbolic space. Möbius addition is used to construct compositional prototypes, while visual-text and text-text losses align compositional representations with full prompt prototypes. The inference stage combines compositional, original, and Euclidean scores. Experiments on three benchmarks demonstrate competitive performance, particularly outstanding results on the MIT-States dataset under open-world evaluation.

Critical strengths
  1. The motivation is clear: two objectives are explicitly connected through a consistent formulation, linking original representations with compositional ones.
  2. Text supervision is effectively utilized: candidate compositional texts provide structural constraints without requiring additional unseen compositional images.
  3. Experimental results are promising. Evaluation covers three benchmarks and two settings. On MIT-States, HM/AUC under open-world setting improves by 4.9/3.8 percentage points compared to the best baselines listed in Table 1.
Critical weaknesses
  1. The novelty is incomplete. The closely related H²EM method was not mentioned, resulting in the failure to fully establish the incremental contribution compared to the existing hyperbolic CZSL method.
  2. The attribution of contributions is not clear. The influence of components such as hyperbolic geometry, Möbius combination, LoRA, and score fusion could not be fully separated, and the construction of feasibility priors and their contributions are still unclear.
  3. The specificity of theoretical analysis is insufficient. The proposed propositions mainly describe general capabilities and scoring characteristics, but do not directly clarify how Möbius improves the generalization performance for unseen combinations.
Questions for authors
  1. Under the condition that the training and inference settings are consistent, how much independent performance improvement can the Möbius combination bring compared to the Euclidean combination operator that maintains the same state-object roles?
1 comment
AuthorsOpenReview ↗

Rebuttal by Authors

We thank the reviewer for recognizing our clear motivation, effective text supervision, and strong open-world results. We address each concern below.

Q1: Relationship and incremental contribution compared to H2EM.

A1: We thank the reviewer for pointing out H2EM. Although both use hyperbolic geometry for CZSL, they differ fundamentally in formulation, external dependencies, and compositional mechanisms:

  1. Hierarchical Entailment vs. Algebraic Closure: H2EM uses Lorentz entailment cones for taxonomic inclusion (sliced apple ⊂\subset apple ⊂\subset fruit). HyCoPro operates in the Poincare ball and uses non-commutative Mobius addition (⊕c\oplus_c) to synthesize composition prototypes directly from primitives (tsh⊕ctoh≈ts,oht_s^h \oplus_c t_o^h \approx t_{s,o}^h) via visual-text and text-text closure.
  2. Self-Contained vs. External LLM Taxonomies: H2EM prompts an external LLM (Gemini-2.5 Pro) to build hypernym trees, whereas HyCoPro is self-contained without external knowledge graphs or LLM queries.
  3. Role of Mobius Addition: H2EM uses Mobius addition only as an attention residual connection while encoding composition prompts independently; HyCoPro uses it to explicitly close state and object prototypes into composition prototypes across modalities.

Thus, H2EM models taxonomic hierarchies, whereas HyCoPro enforces algebraic primitive-to-composition closure.

Q2: Isolating Mobius composition from geometry, LoRA, score fusion, and feasibility priors.

A2: First, for LoRA and score fusion, Table 2 (w/o Hyperbolic Score) evaluates the same visual LoRA using only the Euclidean branch, where OW HM drops from 27.3 to 25.6 (MIT-States), 53.4 to 48.0 (UT-Zappos), and 16.2 to 14.6 (C-GQA), showing gains beyond LoRA alone.

Second, to isolate the Mobius operator from geometry under identical settings (fixing LoRA, losses, weights, and priors), we compare three variants:

  • Variant A (HyCoPro): Non-parametric Mobius addition (⊕c\oplus_c) in the Poincare ball with geodesic distance and hyperbolic interpolation.
  • Variant B (Ordered Euclidean + Hyperbolic): Replaces ⊕c\oplus_c with a learnable tangent-space Euclidean composer (concatenating log-mapped [state, object] features + Linear + LayerNorm + normalization + exp⁡0c\exp_0^c), retaining hyperbolic distance and interpolation.
  • Variant C (Full Ordered Euclidean): Uses the same learnable composer in normalized Euclidean space with cosine distance and Euclidean interpolation.
VariantMIT-States OW (HM / AUC)UT-Zappos OW (HM / AUC)C-GQA OW (HM / AUC)
A: Mobius + Hyperbolic (Ours)27.30 / 12.5053.40 / 37.8016.20 / 4.90
B: Ordered Euclidean + Hyperbolic26.22 / 11.5949.86 / 35.0414.92 / 4.41
C: Ordered Euclidean + Euclidean26.02 / 12.2349.44 / 34.5215.58 / 4.62

We have added this comparison to Table 4 in Section 5.4.

Two conclusions follow: (1) Mobius operator (A vs. B): Within the same hyperbolic space, non-parametric Mobius addition outperforms the tangent-space composer on all datasets (+1.08/+0.91 on MIT-States, +3.54/+2.76 on UT-Zappos, +1.28/+0.49 on C-GQA), while B adds 0.53M parameters. (2) Geometric compatibility (A vs. B vs. C): Combining a tangent-space linear composer with hyperbolic distance (B) underperforms the Euclidean pipeline (C) on C-GQA (14.92 vs. 15.58 HM), whereas intrinsic Mobius addition (A) performs best across all benchmarks (+1.28%, +3.96%, and +0.62% OW HM over C).

Third, for feasibility priors (offline primitive co-occurrence statistics from seen pairs following CSP/PLID, min-max normalized and added as a log prior on MIT-States and C-GQA), please see A3 to Reviewer wQAb, where HyCoPro exceeds all Table 1 baselines without the prior.

Q3: Specificity of theoretical analysis regarding Mobius generalization on unseen pairs.

A3: Propositions 1–3 characterize capacity and conditional margins. We connect Mobius closure to unseen generalization via the hyperbolic triangle inequality: for an unseen pair (s,o)(s,o), let δ=dc(tsh⊕ctoh,ts,oh)\delta = d_c(t_s^h \oplus_c t_o^h, t_{s,o}^h) be the text-closure residual (minimized by Text-Text Closure) and ϵ=dc(vx,ch,vx,sh⊕cvx,oh)\epsilon = d_c(v_{x,c}^h, v_{x,s}^h \oplus_c v_{x,o}^h) be the visual discrepancy (regularized by Visual-Text and Counterfactual Closure). By the geodesic triangle inequality, an unseen image feature's distance to its explicit prototype satisfies dc(vx,ch,ts,oh)≤ϵ+dc(vx,sh⊕cvx,oh,tsh⊕ctoh)+δd_c(v_{x,c}^h, t_{s,o}^h) \le \epsilon + d_c(v_{x,s}^h \oplus_c v_{x,o}^h, t_s^h \oplus_c t_o^h) + \delta. Since Mobius addition is smooth and closed in the Poincare ball, accurate primitive recognition (vx,sh→tshv_{x,s}^h \rightarrow t_s^h, vx,oh→tohv_{x,o}^h \rightarrow t_o^h) with small residuals (ϵ,δ)(\epsilon, \delta) bounds the unseen positive distance, supporting the margin conditions in Propositions 2–3 without unseen training images. We have added this bound to Section 4 as Eq. (20).

Reviewer_wQAbOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes HyCoPro, a hyperbolic prototype framework for compositional zero-shot learning. It represents state, object, and composition semantics in a Poincaré ball and uses Möbius addition to compose primitive prototypes, together with visual-text and text-text closure objectives. Experiments on MIT-States, UT-Zappos, and C-GQA show competitive results, particularly in the open-world setting, with ablations supporting the contributions of the proposed components.

Critical strengths
  1. Clear and coherent methodological contribution. HyCoPro provides a principled way to connect state and object primitives with composition prototypes through hyperbolic geometry and Möbius addition, rather than treating each composition as an independent label. The dual closure objectives make this compositional structure explicit.

  2. Strong empirical performance, especially in open-world CZSL. The method achieves the best reported open-world HM and AUC among the compared methods on all three benchmarks, and the ablations show that the main scoring and closure components contribute to the final performance.

Critical weaknesses
  1. The supplementary material states that the experimental setup uses a random seed of 42, while Table 1 reports only single numerical results without providing variance or standard deviation over multiple runs. Given that some of the performance improvements are relatively small, such as those on open-world UT-Zappos, the results are not sufficiently convincing.

  2. Table 3 evaluates the effects of Text-Text Closure and Visual-Text Closure only on UT-Zappos, whereas the main experiments cover three CZSL benchmarks: MIT-States, UT-Zappos, and C-GQA. Do these components show consistent effects on the other datasets as well?

  3. The final inference procedure introduces several dataset-specific hyperparameters, but the paper provides only limited explanation of how these parameters were selected. In the supplementary material, different values of (γ,α,β)(\gamma, \alpha, \beta) are used for the three datasets. In the open-world setting, feasibility priors are additionally applied to MIT-States and C-GQA. How were these parameters determined, and how sensitive is the performance to their choices?

Questions for authors
  1. Are the main results in Table 1 based on a single random seed? If multiple runs are available, please report mean and standard deviation, especially for open-world UT-Zappos.

  2. Table 3 evaluates the two closure components only on UT-Zappos. Are their effects consistent on MIT-States and C-GQA?

  3. How were the dataset-specific inference parameters (γ,α,β)(\gamma, \alpha, \beta) and feasibility priors selected? Were they chosen using validation data, and how sensitive are the results to these choices?

1 comment
AuthorsOpenReview ↗

Rebuttal by Authors

We thank the reviewer for recognizing our coherent methodological contribution and strong empirical performance, especially in open-world CZSL. We address each concern below.

Q1: Multiple random seeds and variance on open-world UT-Zappos.

A1: Table 1 reports single-run results under seed 42. Below we report the mean and sample standard deviation across three consecutive random seeds (41, 42, and 43) under Closed-World (CW) and Open-World (OW) settings.

DatasetCW S↑CW U↑CW HM↑CW AUC↑OW S↑OW U↑OW HM↑OW AUC↑
MIT-States57.46±0.0651.01±0.2641.90±0.0825.21±0.0957.46±0.0626.39±0.2627.32±0.1812.48±0.12
UT-Zappos68.32±0.7376.15±0.1459.49±0.6547.04±0.6668.32±0.7362.26±0.5352.22±1.0336.84±0.89
C-GQA45.52±0.1337.40±0.6233.15±0.5515.15±0.3145.52±0.1312.22±0.5016.04±0.244.77±0.14

On MIT-States and C-GQA, the OW HM standard deviations are 0.18 and 0.24 percentage points. Mean OW HM exceeds LOGICZSL by 4.92 points on MIT-States and I2CD by 1.14 points on C-GQA. UT-Zappos shows higher variability. Our 3-seed mean surpasses all Table 1 baselines in CW HM/AUC (59.49 / 47.04 vs. 59.2 HM by MSCI and 46.4 AUC by I2CD). In OW, our mean HM/AUC (52.22 / 36.84) exceeds LOGICZSL (50.8 / 36.2) and I2CD (50.2 / 35.6). Both MSCI scores (53.2 / 37.3) fall within one sample standard deviation of our respective means. We have added these multi-seed statistics to Table 6.

Q2: Consistency of closure ablations on MIT-States and C-GQA.

A2: We have extended the UT-Zappos closure ablations (Table 3) to MIT-States and C-GQA (Table 7).

VariantMIT-States CW (HM / AUC)MIT-States OW (HM / AUC)C-GQA CW (HM / AUC)C-GQA OW (HM / AUC)
Full Model41.90 / 25.3027.30 / 12.5033.70 / 15.4016.20 / 4.90
w/o Text-Text Closure40.96 / 24.4426.48 / 11.7932.37 / 14.4915.44 / 4.44
w/o Visual-Text Closure39.95 / 23.1626.27 / 11.4731.82 / 14.3014.93 / 4.28
w/o Both Closures40.00 / 23.1125.23 / 11.3830.42 / 12.9314.41 / 4.02

The closure effects are consistent across all three benchmarks. On MIT-States and C-GQA, both single-closure ablations reduce CW and OW HM and AUC. Removing both closures yields the largest drop in OW HM, by 2.07 and 1.79 percentage points, respectively. On both datasets, removing Visual-Text Closure reduces OW HM more than removing Text-Text Closure, supporting the benefit of enforcing Möbius compositionality on visual primitives.

Q3: Selection and sensitivity of inference hyperparameters (γ,α,β)(\gamma, \alpha, \beta) and feasibility priors.

A3: We used (γ,α,β)=(0.10,0.55,0.50)(\gamma, \alpha, \beta)=(0.10, 0.55, 0.50) for MIT-States, (0.10,0.50,1.00)(0.10, 0.50, 1.00) for UT-Zappos, and (0.20,0.30,0.60)(0.20, 0.30, 0.60) for C-GQA. Section 5.2 now clarifies that we selected these weights by grid search on the validation split and kept them fixed for the main test results.

To evaluate sensitivity, we conduct a post-hoc sweep on the test split. Fixing the trained model and the other two weights, we vary γ∈[0.10,0.50]\gamma \in [0.10, 0.50] (step 0.100.10), α∈[0.30,0.60]\alpha \in [0.30, 0.60] (step 0.050.05), and β∈[0.50,1.00]\beta \in [0.50, 1.00] (step 0.100.10). Figure 3 in Supplementary Material C.3 shows the full sensitivity curves for all three datasets. The table below summarizes the peak-to-peak variation (max⁡−min⁡\max - \min) in open-world HM and AUC.

ParameterSweep RangeMIT-States (OW HM / AUC)UT-Zappos (OW HM / AUC)C-GQA (OW HM / AUC)
γ\gamma[0.10,0.50][0.10, 0.50]0.34 / 0.440.81 / 0.310.29 / 0.12
α\alpha[0.30,0.60][0.30, 0.60]0.40 / 0.310.62 / 1.140.57 / 0.24
β\beta[0.50,1.00][0.50, 1.00]1.55 / 1.022.22 / 2.240.36 / 0.21

OW HM varies by less than one percentage point across the sweeps of γ\gamma and α\alpha on all three datasets. Varying β\beta changes OW HM by 1.55 points on MIT-States and 2.22 points on UT-Zappos, consistent with the complementary hyperbolic and Euclidean scores in Table 2.

Finally, we apply feasibility calibration on the large open-world candidate spaces of MIT-States and C-GQA and omit it on compact UT-Zappos, following prior open-world CZSL practice (e.g., CSP, PLID, Troika, MSCI). We compare open-world test performance with and without the additive log-feasibility prior at fixed inference weights.

Datasetw/o Feasibility Prior (OW HM / AUC)w/ Feasibility Prior (OW HM / AUC)
MIT-States26.22 / 11.7827.12 / 12.42
C-GQA15.32 / 4.6915.93 / 4.86

Without the prior, HyCoPro exceeds the best OW HM and AUC in Table 1 on MIT-States (26.22 / 11.78 vs. 22.4 / 8.7) and C-GQA (15.32 / 4.69 vs. 14.9 / 4.3) among the compared baselines. We have added this comparison to Table 8 in Supplementary Material C.3.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.