SEAL: Semantic-Aware Hierarchical Learning for Generalized Category Discovery

We sincerely appreciate the reviewer's thoughtful and positive feedback. We greatly appreciate the recognition that the paper is well-organized from motivation to methodology and experiments, the proposed modules are clearly explained, and SEAL demonstrates strong performance across fine-grained datasets. Our detailed responses to the concerns are provided below. We‘ll also further strengthen the final manuscript based on the reviewers' feedback. A well-documented code together with all trained models will be made public.

Q1 Hierarchical Data Availability

Thanks for the valuable question. We would like to highlight that the hierarchical information used in our framework is easily obtainable and naturally aligned with real-world structures, rather than being rigid or manually imposed. As detailed in our supplementary material, the hierarchical labels used in our framework are either constructed based on biological taxonomies (e.g., for CUB and Herbarium), or derived from existing internal dataset structures (e.g., Aircraft, where models are already organized in the dataset by manufacturer (e.g., Boeing or Douglas) and further grouped into families such as 767 or 777.), or can be generated using large language models (LLMs) in cases where external hierarchies are unavailable.

We highlight that these hierarchies are not artificially constructed, while they naturally exist in the corresponding domains and align with how taxonomic knowledge is organized in biological and real-world settings. For instance, in botanical research, taxonomists commonly use labelled specimens of known species to classify newly collected, unlabelled samples into existing taxonomic hierarchies or to identify unseen species [a,b]. Similarly, our method assumes access to hierarchical labels only for seen categories, which is both practical and realistic, and reflects real-world conditions where semantic structure is available for known data but not for novel categories.

[a] The role of taxonomy in species conservation - 2004

[b] Species delimitation 4.0: integrative taxonomy meets artificial intelligence - 2024

Q2 Distinction from [41]

Thanks for the insightful question. While Infosieve [41] introduces multi-level learning by encoding each sample using a binary path in an implicit decision tree, it inherently assumes the existence of a meaningful and fixed tree structure. However, such a structure may not accurately reflect the true semantic organization of categories, which is further supported by the observed performance gap in our experiments. For example, SEAL outperforms Infosieve by a significant margin on Scars (65.3 vs. 55.7) and on Aircraft (62.0 vs. 56.3).

In contrast, SEAL emphasizes the use of semantic-guided hierarchical supervision, which is both flexible and interpretable, and does not rely on predefined or artificially imposed tree structures. Our framework is grounded in a solid information-theoretic analysis, which motivates the design and ensures that the introduced hierarchy enhances rather than misleads the learning of target-level categories. This makes SEAL more robust and better aligned with real-world semantic structures compared to tree-based approximations.

Q3 Discussion of performance on generic datasets

Thanks for the thoughtful question. The relatively smaller improvements observed on CIFAR-100 and ImageNet-100 can be attributed to the nature of their semantic structure. These datasets are coarse-grained, and the hierarchical relationships between classes are often broad and loosely defined (e.g., grouping semantically distant categories like " computer keyboard" and "clock" under a general class as "household electrical devices"), resulting in weaker and less informative supervision signals from the hierarchy. In contrast, fine-grained datasets such as CUB, Aircraft, and Scars feature tighter semantic groupings (e.g., species within the same genus or models from the same manufacturer), which allow our framework to more effectively exploit semantic hierarchies for representation learning and category separation. This difference naturally leads to more pronounced gains in fine-grained settings. We appreciate the reviewer for pointing this out, and we will further explore strategies to enhance the utility of hierarchical information in coarse-grained scenarios as part of our future work.

Q4 Ablation Studies

Thanks for the insightful question. As shown in Tab. 4 (Line 1), directly incorporating hierarchical supervision leads to performance degradation on CUB and Aircraft. While our theoretical motivation (Section 4.1) suggests that hierarchical information can be beneficial, naïvely applying it in the GCD setting can cause the training to converge to suboptimal local minima due to inconsistent or noisy supervision. To address this, we introduce Cross-Granularity Consistency (CGC) to ensure alignment between hierarchical levels, and Semantic-Guided Hierarchical Soft Contrastive Learning (HSCL) to better model inter-class relationships by relaxing the equally-negative assumption used in prior work. The improvements observed after applying these modules (Tab.4, Ours) further support their effectiveness in robustly leveraging hierarchical signals under the challenges of GCD.

Regarding the Semantic-Guided HSCL, we clarify that it does not hurt performance on CUB. In fact, comparing Tab.4 (line 2 without HSCL) with the final Ours row shows that incorporating this module improves CUB accuracy from 57.8 to 66.2, demonstrating its effectiveness.

W1 Performance on CUB

Thanks for pointing this out. We will revise the abstract to avoid overstatement. We attribute the relatively lower performance on CUB not to a limitation of our method, but rather to the unique data distribution of the CUB dataset, which presents specific challenges that affect certain types of label assignment strategies.

Specifically, we observe that CUB exhibits a higher degree of intra-class multi-modality—that is, individual classes tend to contain multiple distinct sub-clusters—compared to other datasets such as Scars and Aircraft. To empirically validate this, we conduct an analysis using Gaussian Mixture Models (GMMs) on PCA-reduced, L2-normalized features, selecting the optimal number of components M within each class via BIC (Such operation is standard for mixture selection [c], used in [d] to estimate the multimodality of a given distribution and used in class‑number estimation [e]. Our results show that 145 out of 200 classes in CUB require $M\geq2$, indicating multi-modal structure, whereas only 7 out of 196 in SCARS and 20 out of 100 in Aircraft show such behavior. This confirms that CUB classes are often fragmented into multiple sub-modes. We report the average number of components per class and the percentage of classes with $M\geq2$.

Under such conditions, non-parametric classifiers like k-means—as used in Infosieve [41]—are better suited to capture these sub-clusters. In contrast, our method uses a parametric classifier head trained with cross-entropy, following the widely adopted SimGCD [57] framework. While this design is effective in many settings, it may be structurally under-specified for modeling multi-modal distributions, leading to suboptimal label assignment on CUB.

To further examine this, we re-evaluate our trained model using the same semi-supervised k-means in [41] clustering (averaged over 3 runs) on extracted features. Under this semi-supervised k-means setting, our method outperforms Infosieve [41] on CUB, especially when using a stronger DINOv2 backbone, which reduces initialization noise and yields more structured features. Additionally, applying semi-supervised k-means to the SimGCD baseline yields a similar performance boost, further confirming that the lower performance is not due to the learned representations, but rather due to a mismatch between classifier type and the underlying data structure.

We will clarify this in the final version and include the k-means evaluation results to present a more complete picture of our method’s effectiveness.

[c] Bishop, Pattern Recognition and Machine Learning, Sec. 9 (2006).

[d] Wasserstein Distances, Neuronal Entanglement, and Sparsity, ICLR(2025)

[e] X‑means: Extending k‑means with efficient estimation of the number of clusters, ICML(2000).

W2 Code

Thanks for the suggestion. We will release our code upon acceptance.

W3 LegoGCD

Thanks for the suggestion. We would like to clarify that we have cited LegoGCD[6] and compared our method with LegoGCD[6] in Tab.2, where our method achieves stronger results. LegoGCD mitigates catastrophic forgetting via entropy regularization and dual-view KL divergence, whereas our method emphasizes the role of semantic hierarchies in generalized category discovery. These complementary approaches address GCD from distinct perspectives, jointly highlighting the value of structural guidance. We will include a discussion of LegoGCD in the Related Work section in the revised version.

We sincerely appreciate your thoughtful follow-up and your acknowledgment of our clarifications regarding related work, dataset differences, and our ablation studies.

Regarding hierarchical information, we further provide clarafications from two perspectives, and hope it could help.

1. Our method assumes hierarchical labels only for the labelled data, and no hierarchy is presumed for the unlabelled portion. Such hierarchies are not manually predefined but derived directly from the labelled data, such as existing taxonomies in biological datasets. Assuming such structure avaible for labelled data is common in related areas such as image classification[9, 15], and Open-Set Recognition[31, 58], and mirrors real-world practice: in botanical research, taxonomists commonly use labelled specimens of known species to classify newly collected, unlabelled samples into existing taxonomic hierarchies or to identify unseen species [a,b]. In this sense, our assumption reflects a practical scenario rather than an idealized one.

2. To evaluate more realistic setting, Appendix C (Table 4) reports results where neither the hierarchy nor the number of classes is known; we employ LLM to generate hierarchical labels from labelled data, and SEAL remains strong, outperforming baselines without ground-truth category numbers. We further provide an additional study using LLM-generated hierarchies (ChatGPT) that capture alternative semantic dimensions. For example, on Scars, we complement the vehicle type hierarchy (e.g., SUV/Van/Coupe) with a brand-based hierarchy (e.g., Audi/BMW). Under hierarchies from various semantic dimensions, our approach remains strong and achieves state-of-the-art performance, indicating that SEAL is not tied to a single predefined hierarchy. In this paper, we highlight that semantic-guided hierarchies—rather than abstract, hand-crafted ones—are key for GCD, whether sourced from curated taxonomies, generated by LLMs, or defined along different semantic dimensions. For reproducibility, we report results with the most readily accessible hierarchy in the main paper, while our framework readily accommodates other valid hierarchies.

Once again, we sincerely thank you for your thoughtful reply and for acknowledging our earlier clarifications, and we hope the additional analyses further address your concerns. If you have any further concerns, we would be happy to discuss and improve the manuscript accordingly.

We sincerely thank the reviewer for the thoughtful and positive feedback. We greatly appreciate the reviewer's recognition of the value of incorporating hierarchical information into GCD and acknowledgment of our sound experiment performance across three datasets. Our detailed responses to mentioned concerns are provided below. We‘ll also further strengthen the final manuscript based on the reviewers' feedback. A well-documented code will be made public.

'Naturally occurring' & Difference with [41, 55]:

Thanks for raising this point. In Lines 54–55, by "naturally occurring semantic hierarchies", we refer to domain-inherent structures that are readily available without manual design. For instance, taxonomic hierarchies in species classification (e.g., family, genus, species) are well-established and widely used in scientific databases. These reflect real-world semantic relationships and require no additional annotation. Our method leverages such domain-native hierarchies as semantically meaningful supervision. We will revise the wording to avoid ambiguity in the final version.

Moreover, our approach differs from prior works-Infosieve [41] and TIDA [55] in both the source and use of hierarchical information. Infosieve [41] encodes category structure via binary paths in an implicit decision tree, assuming a fixed latent hierarchy that may not match true semantics—leading to suboptimal performance (e.g., SEAL outperforms Infosieve: Avg. 64.5 vs. 60.5). TIDA [55], meanwhile, uses manually defined hierarchies for prototype-level supervision, which risks introducing noisy supervision and lacks scalability. In contrast, SEAL employs readily available, semantic-guided hierarchies and integrates them through an information-theoretic framework to provide consistent, structured supervision across multiple levels, ultimately improving both representation quality and category discovery.

Justification of motivation:

Thanks for highlighting this point. We acknowledge that the statement in Lines 41-43 may come across as strong, and we will revise the final version to smooth our tone.

To support our motivation, we provide two forms of empirical evidence illustrating the potential risks of handcrafted or semantically misaligned hierarchies:

1. TIDA [55] adopts manually defined abstract hierarchies, which, while aiming to introduce structure, may not reflect true semantic relationships. This can lead to suboptimal supervision signals. In our experiments, SEAL significantly outperforms TIDA on multiple datasets, for example, CUB (66.2 vs. 54.7) and Aircraft (62.0 vs. 54.6), suggesting that misaligned or artificial hierarchies can adversely affect GCD performance.

2. To further validate this, we conduct an additional experiment using randomly generated coarse-level labels as a substitute for semantically meaningful hierarchies. In this setting, performance drops substantially compared to using true semantic hierarchies, indicating that the effectiveness of our method is not simply due to hierarchy, but rather the semantic alignment of the hierarchical information.

Assumption validity:

Thanks for the thoughtful comment. We assume ONLY the availability of hierarchical labels for labelled data, which is fully aligned with the Generalized Category Discovery task and reflects realistic conditions in many domains.

In Sec. 4.1, we leverage available hierarchical labels to decompose the mutual information via the chain rule (Line 135), showing that incorporating coarse-grained labels provides more informative and semantically structured supervision.

For unlabelled data, we make no assumptions about label availability. Instead, we apply the variational lower bound: $I(X, Y)\geq\mathbb{E}_X \left[ \mathrm{KL}(q(Y \mid X) \parallel p(Y)) \right]$ where $q(Y \mid X)$ is approximated by the model's predictive distribution.

This leads to a tractable lower bound expressed as$\mathcal{H}_\theta(\hat{Y}_u \mid X_u) - \mathcal{H}(\hat{Y}_u)$ where $\hat{Y}_u$ refers to the model's predicted label distribution for the unlabelled data, as detailed in Appendix, Lines 43–50.

We will revise the manuscript to clarify this assumption and make the notation between labelled and unlabelled data more explicit in our theoretical motivation.

Number of hierarchies:

Thanks for the insightful question. The number of hierarchical levels $H$ is known, given the ground-truth hierarchies of labelled categories. In the main paper, we assume the number of classes at the target level is known, following common practice in the GCD community [23, 24, 33, 41, 52, 55], which ensures fair comparison under controlled settings. For the category numbers of other levels, we estimate them using the estimation method from GCD [23]. Actually, they are consistent with the ground-truth numbers (see Appendix Tab. 3). We will offer a more detailed clarification in the final version.

To evaluate more realistic cases where the number of classes at the target level is also unknown, we also provide results in Appendix C (Line 110 onward), where class numbers are estimated using the method from GCD [23] with only labeled hierarchical data. As shown in Appendix Tab. 4, SEAL maintains strong performance, outperforming other baselines even without ground-truth category numbers.

Transition Matrix Update Mechanism:

Thanks for pointing this out. We refine the transition matrices by averaging the coarse-level model-predicted distribution ($p_h$) of samples predicted to the corresponding fine-grained class, enabling bottom-up propagation of semantic structure.

Given model predictions at two granularity levels—$p_H$ (fine-grained) and $p_h$ (coarse)—we assign pseudo-labels at the fine level via $y_H = \arg\max(p_H)$ (Algo.1, line 6). For each novel class $k$ (i.e., not matching any base category), we collect samples predicted as class $k$ where $y_H == k$ and compute their average predicted distribution at the coarse level: $\text{mean}(p_h[y_H = k])$.

This average captures how the fine-level class $k$ relates to higher-level categories, enabling bottom-up refinement of the hierarchy. The underlying intuition is that samples grouped under the same fine-level class should also exhibit consistent coarse-level semantics.

We will improve the presentation to clarify this refinement process and its motivation.

Fusion Mechanism & $\lambda_{s}$:

Thanks for the insightful comment. In our implementation, the fused similarity matrix is computed as$\tilde{S_h} =\frac{1}{h} \sum_{n=h}^1 S_n$where $S_n$ denotes the similarity at level n. This top-down fusion integrates coarse-level semantics into fine-grained supervision.

Unlike prior methods that treat all non-positive samples equally, we leverage hierarchical structure to impose semantic-aware soft supervision. For example, samples from semantically related categories (e.g., Siamese (Cat) vs. Egyptian Mau (Cat) are penalized less than those from distant ones (e.g., Siamese (Cat) vs. Beagle(Dog) ). This enables the model to learn more nuanced, structure-aware decision boundaries.

The variable $\lambda_s$ in line 191 is a hyperparameter used to control the smoothness of our proposed semantic-aware soft labels. We further provide an additional ablation study to demonstrate the effectiveness and impact of this component. We will define its meaning in the final version.

Curriculum Learning Schedule:

Thanks for the helpful comment. The weighting coefficient $\lambda_c$ follows a linear decay schedule: $\lambda_c = 1 - \frac{\text{epoch}}{\text{TotalEpochs}}$, gradually shifting focus from semantic alignment (via cosine similarity) to finer positional discrimination as training progresses.

An ablation study comparing fixed values ($\lambda_c = 0, 0.5, 1$) and exponential decay($1 - \exp\left(-\frac{\text{epoch}}{\text{TotalEpochs}} \right)$) confirms that decaying schedules consistently outperform fixed baselines, validating the effectiveness of this curriculum design.

We will provide more clarification regarding this strategy in the final version.

We sincerely thank the reviewer for the thoughtful feedback. We greatly appreciate the reviewer's recognition of the paper's clarity, the clear and compelling motivation for using semantic hierarchies in GCD, and the detailed presentation of the SEAL framework. We're also glad that the theoretical justification and computational analysis in the supplementary material were found to be valuable. Our detailed responses are provided below. We will also further strengthen the final manuscript based on the reviewers' concluding feedback. A well-documented code together with all trained models will be made public.

Q1&W1 Difference between DebGCD

Thanks for the valuable question. DebGCD addresses general label bias, while SEAL targets the core challenge of leveraging semantic hierarchy in GCD, leading to more scalable and lightweight solution, especially under strong representation.

DebGCD tackles the label bias issue by introducing an auxiliary debiased classifier trained on pseudo-labels, along with an OOD-aware regularization strategy. This approach is particularly effective under weaker backbones such as DINO, where label bias is more severe and novel-class separation is more difficult. By explicitly mitigating overfitting to known classes, DebGCD improves robustness in such less discriminative feature spaces.

In contrast, SEAL focuses on a space-agnostic challenge: how to effectively leverage inherent semantic hierarchies when only a subset of the data is labeled. Grounded in an information-theoretic framework, SEAL introduces two key components—Cross-Granularity Consistency and Semantic-Guided Hierarchical Soft Contrastive Learning—to provide structure-aware supervision across multiple levels of semantic granularity. These components enable the model to capture fine-grained semantic relationships and improve generalization to novel categories.

While SEAL and DebGCD achieve comparable performance under DINO (64.5 vs. 64.4 avg.), SEAL consistently outperforms DebGCD, and the advantage becomes more pronounced under the stronger DINOv2 backbone (76.3 vs. 74.9). This suggests that while DebGCD is highly effective at correcting label bias under weak initialization, SEAL excels in utilizing hierarchical structure, especially when the feature space is less noisy.

Furthermore, we perform a direct comparison of model complexity in terms of model parameters and FLOPs, and find that SEAL introduces significantly fewer parameters and lower computational overhead, making it a more lightweight solution.

Q2&W2 Performance on CUB

Thanks for this insightful question. We attribute the relatively lower performance gain on CUB not to a limitation of our method, but rather to the unique data distribution of the CUB dataset, which presents specific challenges that affect certain types of label assignment strategies.

Specifically, we observe that CUB exhibits a higher degree of intra-class multi-modality—that is, individual classes tend to contain multiple distinct sub-clusters—compared to other datasets such as Scars and Aircraft. To empirically validate this, we conduct an analysis using Gaussian Mixture Models (GMMs) on PCA-reduced, L2-normalized features, selecting the optimal number of components M within each class via BIC (Such operation is standard for mixture selection [a], used in [b] to estimate the multimodality of a given distribution and used in class‑number estimation [c]). Our results show that 145 out of 200 classes in CUB require $M\geq2$, indicating multi-modal structure, whereas only 7 out of 196 in SCARS and 20 out of 100 in Aircraft show such behavior. This confirms that CUB classes are often fragmented into multiple sub-modes. We report the average number of components per class and the percentage of classes with $M\geq2$.

Under such conditions, non-parametric clustering method like k-means—as used in Infosieve [41]—are better suited to capture these sub-clusters. In contrast, our method uses a parametric classifier trained with cross-entropy, following the widely adopted SimGCD [57] framework. While this design is effective in many settings, it may be structurally under-specified for modeling multi-modal distributions, leading to suboptimal label assignment on CUB.

To further examine this, we re-evaluate our model using the semi-supervised k-means in [41] clustering (averaged over 3 runs) on extracted features. Under this semi-supervised k-means setting, our method outperforms Infosieve [41] on CUB, especially when using the strong DINOv2 backbone, which reduces initialization noise and yields more structured features. Additionally, applying this to the SimGCD baseline yields a similar performance boost, further confirming that the lower performance is not due to the learned representations, but rather due to a mismatch between clustering method and the underlying data structure.

We will clarify this in the final version and include the k-means evaluation results to present a more complete picture of our method’s effectiveness.

[a] Bishop, Pattern Recognition and Machine Learning, Sec. 9 (2006).

[b] Wasserstein Distances, Neuronal Entanglement, and Sparsity, ICLR(2025)

[c] X‑means: Extending k‑means with efficient estimation of the number of clusters, ICML(2000).

Q3 Difference with related work

Thanks for the helpful suggestion. We will revise the presentation accordingly with a more detailed discussion.

Key Advantages of Our Hierarchical Semantic-Guided Approach :

1. Semantic-Guided Hierarchical Supervision
   Unlike Infosieve [41], which encodes category structure using binary paths in an implicit decision tree (thus assuming a fixed latent tree structure), and TIDA [55], which relies on handcrafted artificial hierarchies, which may introduce erroneous supervision as evidenced by the performance gap between TIDA and SEAL (e.g., 54.7 vs. 66.2 on CUB, 54.6 vs. 62.0 on Aircraft). SEAL is the first to incorporate semantically guided hierarchical labels into the learning process. These labels are readily available in many real-world scenarios and provide meaningful structural cues.
2. Semantic-Aware Soft Contrastive Learning
   Prior contrastive learning methods (e.g., GCD [48], SimGCD [57], DebGCD [33]) treat all negative samples equally. In contrast, SEAL introduces semantic-aware soft contrastive learning by leveraging the class hierarchy to impose graded penalties.
   For instance, samples from semantically similar categories—such as Siamese and Egyptian Mau (both domestic cats)—should be penalized less than distant ones like Beagle (a dog breed).
   This structure-aware contrastive learning improves representation quality and allows more nuanced inter-class separation.

Methods like DTC [23], RS [21, 22], UNO [18], NCL [65], DualRS [63], and JOINT [27] operate under the Novel Category Discovery setting, assuming all unlabelled samples are from unseen classes. These methods leverage clustering, rank statistics, and label balancing, but are not directly applicable to GCD.

GCD [48] generalizes NCD by assuming that unlabelled data contains both seen and novel classes. Under this setting, ORCA [5] introduces an uncertainty-adaptive margin to dynamically balance the learning speed between seen and novel categories. CiPR [24] constructs pseudo-labels from confident unlabelled samples through clustering and enforces consistency via a contrastive loss. DCCL [39] further improves pseudo-label quality by leveraging local structure mining in a teacher-student contrastive learning framework. SimGCD [57] establishes a strong end-to-end baseline by jointly training a feature extractor and a parametric classifier using self-distillation and entropy regularization, while SPTNet [52] enhances SimGCD with spatial prompt tuning to emphasize discriminative object regions. Infosieve [41] encodes each sample using binary paths in an implicit decision tree, which assumes a fixed latent hierarchical structure. TIDA [55], on the other hand, introduces handcrafted abstract hierarchies by constructing prototypes at manually defined levels.

W3 Backbone choose for Tab.3

Thanks for pointing this out, and we will fix the order typo of Tab.2 and Tab.3 in the final version. For Tab.3 (Pets and Herbarium), we report results using DINO, which is a common practice in the GCD community for these datasets in prior works [41,48,52,57]. To further address your concern, we have additionally included the results of our method using DINOv2 for reference, where SEAL achieves consistent improvements when switching from DINO to DINOv2, especially on the more challenging Herb19 dataset (+12.1 for All), highlighting the scalability of our method under stronger visual representations.

Limitations Section:

Thanks for pointing this out. We will revise the paper to move the limitations from Appendix to the main paper to include a concise summary of the most relevant limitations.

We sincerely thank the reviewer for the thoughtful and positive feedback. We greatly appreciate the reviewer's recognition of the value of first introducing biological classification systems into GCD, clear advantage over similar work DebGCD, and sufficient ablation studies. Our detailed responses to the listed concerns are provided below. We will also further strengthen the final manuscript based on the reviewers' concluding feedback. A well-documented code together with all trained models will be made public.

W1 Static Hierarchy Dependency

We would like to clarify that the semantic hierarchies leveraged in our framework are adaptable and well-aligned with real-world structures, rather than rigid or artificially imposed. In biological sciences, hierarchical classification arises naturally and serves as a foundational principle in taxonomy. For example, in botanical research, taxonomists commonly use labelled specimens of known species to classify newly collected, unlabelled samples into existing taxonomic hierarchies or to identify novel species [a,b].

Similarly, in our work, we assume that hierarchical labels are available ONLY for the seen categories, which is a widely adopted and realistic assumption. Such labels are commonly used and readily available in many prior studies[9,15,40]. In practice, they can be obtained directly from curated resources (e.g., the GBIF botanical database [Supp.9]) or automatically generated using large language models (LLMs), as shown in our supplementary material B.1. This design not only aligns with the natural structure of biological classification systems but also enhances the interpretability of open-world learning.

To further demonstrate the practicality of our setting, we test the use of LLM-generated hierarchical labels. Specifically, we apply ChatGPT to generate taxonomic structures, and observe that the outputs align with the ground-truth hierarchies. This shows that modern LLMs are capable of producing reliable semantic hierarchies with minimal effort, and our method still achieves strong performance under this setup, validating the effectiveness and accessibility of such automatically constructed hierarchies in real-world discovery scenarios.

[a] The role of taxonomy in species conservation - 2004

[b] Species delimitation 4.0: integrative taxonomy meets artificial intelligence - 2024

W2 Label Sensitivity

Thanks for pointing out label sensitivity. While we would like to clarify that we do not state that performance drops sharply when the real semantic structure conflicts with a preset hierarchy. Rather, our concern is that incorrect or arbitrarily defined hierarchies such as the abstract superclasses manually designed in prior work [55] may introduce misleading supervisory signals due to their lack of semantic grounding.

To further support this point, we provide an additional experiment using randomly generated hierarchical labels, which indeed resulted in performance degradation. This highlights the importance of semantically meaningful hierarchies rather than abstract self-defined hierarchies.

W3 Recommend beautifying the images in the article

Thanks for the suggestion. We will try to further beautify the images. If more detailed suggestions could be provided, we would be happy to adopt.