acceptodds
Under review as a conference paper at ICLR 2027

MGME: Manifold-Guided Multi-Prototype Expansion for Open World Object Detection

Abstract

In Open World Object Detection, a detector must recognize known classes while simultaneously identifying instances of unknown categories. Existing approaches typically represent unknown objects using the arithmetic mean of known class embeddings or a single generic text prompt. However, a single mean proxy compresses diverse semantic directions into one representation and may fall into semantically ambiguous regions, a limitation we term mean collapse, thereby restricting its ability to capture multimodal unknown semantics. To overcome this geometric flaw, we propose the Manifold-Guided Multi-Prototype Expansion (MGME) framework. Instead of compressing unknowns into a singular proxy, MGME models them as a diverse set of unknown prototypes distributed along the valid boundaries of known classes via orthogonal geodesic traversal. Specifically, we extract a focused orthogonal residual subspace to serve as a collision-free exploration space for novelty generation. Furthermore, we design a dual-constraint selection mechanism that combines cross-class boundary pruning with large language model-driven semantic steering to ensure the generated prototypes are both topologically safe and semantically aligned. Finally, we introduce Discriminative Prototype Orthogonalization during the training phase to push known text prototypes apart, establishing highly discriminative decision boundaries. Extensive experiments across five real-world object detection datasets demonstrate that MGME effectively prevents manifold collisions and generates novel unknown prototypes with strong visual objectness, achieving state-of-the-art performance and yielding significant improvements in unknown object discovery.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Open World Object Detection, Unknown Object Discovery, Riemannian Geometry, Multi-Prototype Learning, Vision-Language Models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_jdXmOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This submission proposes MGME for open-world object detection, which replaces a single unknown-class representation with multiple synthetic unknown prototypes. The method estimates a principal subspace from known visual features, samples directions from a truncated residual subspace, and projects these directions onto the tangent space of each known-class center. Candidate prototypes are generated through spherical geodesic traversal, filtered by their similarity to other known classes, and then ranked using a domain-level text embedding derived from LLM-generated concepts. The method also add a decorrelation loss on known-class text prototypes during adaptation. Experiment on five RWD datasets report improvements in average unknown and known detection performance over the evaluated baselines.

Critical strengths
  1. The motivation for using multiple unknown prototypes is reasonable.

I also agree that a single representation is a restrictive model for heterogeneous unknown objects. Generating several prototypes around known categories provides a practical alternative, without requiring annotated unknown instances.

  1. The reported aggregate gains are substantial.

On ViT-B/16, average U-mAP increases from 6.8 for PASS to 11.8. And on ViT-L/14, it increases from 15.9 for UMB to 19.2. Average known-class performance also improves. Evaluation across five domains and two backbone sizes provide useful evidence that the gains are not restricted to one setting.

  1. The comparison is reasonable.

The submission report using common splits, support samples, attribute pools, and evaluation code for the reproduced methods. Sharing the generated vocabulary with BASE-ZS+LLM is also helpful, although this comparison does not fully isolate the geometric component. Component ablations and failure cases also provide additional insight into the method.

Critical weaknesses
  1. The geodesic ablation does not adequately establish the claimed geometric advantage.

For unit vectors cc and vv with c⊤v=0c^\top v=0,

\frac{c+\alpha v}{\||c+\alpha v\||} =\cos\theta c+\sin\theta v, \qquad \alpha=\tan\theta, \$$ for the angles used in this submission. Thus, normalized additive perturbation with a matched step size produces exactly the same candidate as Eq. 3. Sec. 4.3 appropriately limit its conclusion to the particular implemented Euclidean variant, but that variant is not specified. The large performance gap in Tab. 2 therefore cannot be interpreted without knowing whether normalization, directions, and angular displacement were matched. This does not invalidate the overall method, but it weakens the broader contrast between Euclidean and manifold-based construction. 2. **The connection between generated prototypes and the detection head is missing.** Sec. 3.2– Sec. 3.4 produce vectors in $\mathbb{R}^d$, whereas Eq. 7 requires additional columns in the attribute assignment matrix $W$. The submission does not explain how each generated vector is converted into an $M$-dimensional weight column. If this involves reconstruction through the attribute embeddings, the effective classifier may differ from the generated prototype, which may potentially change the geometric constraints established during selection. I think this is a necessary detail for understanding and reproducing the method. 3. **Several geometric guarantees are stronger than the construction supports.** Unit norm does not establish that a prototype has visual objectness, or lies in a populated region of the visual feature distribution. Similarly, the threshold in Eq. 4 constrains similarity to non-parent class centers, not the actual decision boundaries or full class distributions. Moreover, projecting a residual vector onto a class-specific tangent space do not generally preserve orthogonality to the principal subspace: $$\ q^\top\left[v_{\mathrm{res}}-(v_{\mathrm{res}}^\top c_k)c_k\right] =-(v_{\mathrm{res}}^\top c_k)(q^\top c_k) \$$ for a principal direction $q$. Claims of collision-free exploration and preserved foundational attributes should therefore be qualified or supported by direct measurements. 4. **The experiments do not sufficiently separate the benefit of multiple prototypes from the proposed generation mechanism.** I think it's better to include known-class centers used as unknown prototypes, random tangent-space prototypes, and direct LLM text prototypes, all evaluated with the same adapted model and comparable prototype counts. BASE-ZS+LLM differs in adaptation as well as prototype construction, thus its gap to MGME does not isolate geometric alignment. Also, the one-prototype-per-class setting in Tab. 4 is not a global single-prototype baseline.
Questions for authors
  1. What is the formula for the Euclidean-addition ablation? Are the perturbation directions, normalization, angular distances, sampling budgets, and selection procedures matched to MGME?

  2. How are generated unknown prototypes converted into the additional columns of WW in Eq. 7?

  3. How much principal-subspace alignment is reintroduced by Eq. 2 in practice? It's better to also report candidate rejection rates and prototype similarities to known instances to support the claims about boundary safety.

  4. Can the authors compare alternative prototype constructions under a fixed adapted checkpoint and matched prototype counts? In particular, direct LLM text prototypes and random tangent-space prototypes would help to establish what the proposed residual-space construction add.

5 comments
AuthorsOpenReview ↗

Response to Reviewer jdXm (Part 1/4)

We sincerely thank the reviewer for the careful and constructive assessment of our work, particularly for recognizing the motivation of multi-prototype unknown modeling, the substantial empirical improvements, and the fairness of our experimental comparisons. We appreciate your detailed comments on the geometric formulation and implementation, which have helped us clarify several important aspects of MGME.

Q1 (W1): Clarification of the Euclidean-addition ablation, including its construction formula, normalization, angular displacement, and comparison settings.

A1. Mathematical equivalence. We agree that, for a unit known-class center ckc_k and a unit tangent direction v⊥v^\perp satisfying ck⊤v⊥=0c_k^\top v^\perp=0, the spherical exponential map in Eq. (3) is mathematically equivalent to normalized Euclidean addition with a matched angular displacement:

=\cos\theta\\,c_k+\sin\theta\\,v^\perp =\operatorname{Normalize}(c_k+\tan\theta\\,v^\perp).$$ Therefore, when the exploration direction and angular displacement are matched, these two formulations produce identical prototypes and cannot exhibit an intrinsic performance difference. **This ablation was intended to evaluate an alternative prototype-construction strategy against MGME's residual-guided geometric construction, rather than to compare two mathematically equivalent formulations. We apologize that this intended comparison was not sufficiently clarified in the original manuscript.** **A2. Implementation of the Euclidean-addition variant.** We apologize for the ambiguity introduced by the term "Euclidean Add." in Table 2. Specifically, this variant refers to an alternative prototype construction based on **Euclidean addition along the difference between neighboring known-class centers**, rather than replacing Eq. (3) with its equivalent normalized additive form. For each known-class center $c_k$, we identify its nearest other known-class center $c_{j^*}$ according to cosine similarity. We then compute $$\begin{aligned} d_k &= c_{j^*}-c_k,\\\\ p_k^{\mathrm{Euc}} &=\operatorname{Normalize}\left( c_k+0.5[d_k+0.1\\|d_k\\|_2\xi] \right), \end{aligned}$$ where $\xi$ is a randomly sampled unit vector. The resulting prototype is **L2-normalized**. The coefficient 0.5 controls interpolation along the inter-class difference and is not chosen to match the angular displacement $\theta$ in Eq. (3). Thus, the variant differs from MGME in both its exploration directions and angular displacements. Therefore, Table 2 compares different prototype-construction strategies rather than two equivalent parameterizations of the same traversal. **A3. Interpretation of the results.** The purpose of this variant is to evaluate whether simple inter-class Euclidean addition can provide an effective alternative to MGME's residual-guided, angularly controlled prototype construction. As reported in Table 2, the variant achieves U/K-mAP of **5.6/39.9 on Aerial** and **1.8/28.2 on Medical**, compared with **12.6/42.0** and **15.1/28.3** for MGME. These results support the effectiveness of the complete MGME prototype-construction strategy relative to this particular alternative, rather than demonstrating an intrinsic superiority of the exponential map over mathematically equivalent normalized addition. **Manuscript revisions.** We have explicitly clarified the equivalence between geodesic traversal and normalized Euclidean addition in **Sec. 3.3**, renamed the ablation **"w/o Geodesic Escape (Inter-class Euclidean Add.)"** in **Table 2**, and added its implementation formula and corresponding interpretation in **Sec. 4.3**. We appreciate the reviewer's observation, which has helped us distinguish the geometric parameterization from the actual unknown-prototype construction strategy and present a more precise interpretation of our results.
AuthorsOpenReview ↗

Response to Reviewer jdXm (Part 2/4)

Q2 (W2): How are the generated unknown prototypes incorporated into the attribute-based detection head in Eq. (7)?

A1. Direct integration into the detection head. We thank the reviewer for highlighting this missing implementation detail. We agree that the original description did not sufficiently explain how the generated dd-dimensional unknown prototypes are integrated into the attribute-based classifier.

Importantly, the generated prototypes are directly appended to the embedding matrix rather than reconstructed through the original attribute embeddings. Therefore, their geometric representations are preserved when incorporated into the detection head.

Specifically, the original attribute-based classifier contains an attribute embedding matrix Eatt∈RM×dE_{\mathrm{att}}\in\mathbb R^{M\times d} and an attribute-to-class assignment matrix W∈RM×KW\in\mathbb R^{M\times K}, where MM is the number of attributes and KK is the number of known classes.

Let Punk∈RU×dP_{\mathrm{unk}}\in\mathbb R^{U\times d} denote the generated unknown-prototype matrix, with each row corresponding to one prototype pip_i. We extend the two matrices as

E_{\mathrm{att}}^+ &=[E_{\mathrm{att}};P_{\mathrm{unk}}] \in\mathbb R^{(M+U)\times d},\\\\ W^+ &=\operatorname{blockdiag}(W,I_U) \in\mathbb R^{(M+U)\times(K+U)}. \end{aligned}$$ Here, $I_U$ is an identity matrix that establishes a direct one-to-one mapping between the generated unknown prototypes and the additional unknown-response channels. **A2. Preservation of the generated prototype representations.** For a visual embedding $v$, the expanded raw-logit computation in Eq. (7) becomes $$\begin{aligned} Logit_{\mathrm{raw}} &=\cos(v,E_{\mathrm{att}}^+)\\,W^+\\\\ &=[Logit_K,Logit_{U_1},\ldots,Logit_{U_U}]. \end{aligned}$$ Because the additional block in $W^+$ is $I_U$, each generated unknown prototype directly contributes its own cosine-similarity response. No additional attribute reconstruction or learned projection is applied to the generated prototypes at this stage. Consequently, the effective unknown classifier directions remain those obtained through the geometric construction and selection process, addressing the reviewer's concern that reconstructing the prototypes through the original attributes might alter their geometric properties. The individual unknown responses are then aggregated into a single unknown score: $$Logit_U^*=\max_{i=1,\ldots,U}Logit_{U_i}.$$ Following the inherited inference framework, this response is combined with the existing attribute-relevance and outlier-suppression factors: $$P_U(v)=\sigma(Logit_U^*)\\,p_{\mathrm{ID}}(v)\\,p_{\mathrm{OOD}}(v).$$ Thus, MGME expands the internal unknown representation while retaining a unified unknown prediction and the baseline scoring framework. **Manuscript revisions.** We have added the explicit definitions and dimensions of $E_{\mathrm{att}}^+$ and $W^+$ in **Sec. 3.5 (Multi-Prototype Inference)** and revised **Eq. (7)** to show how the generated prototypes are directly incorporated into the detection head. We appreciate the reviewer's suggestion, which has helped us clarify an important implementation detail and improve the reproducibility of our method.
AuthorsOpenReview ↗

Response to Reviewer jdXm (Part 3/4)

Q3 (W3): Geometric Constraints and Boundary Safety

We thank the reviewer for the insightful suggestions regarding the geometric properties of our construction, particularly visual objectness, residual-subspace orthogonality, and boundary safety. We agree that these properties should be distinguished from strict mathematical guarantees. To better substantiate our claims, we have conducted additional geometric analyses covering principal-subspace alignment, candidate rejection rates, and prototype similarities to known instances.

A1. Residual-subspace and tangent-space orthogonality. We agree that the tangent-space projection in Eq. (2) does not necessarily preserve strict orthogonality to the global principal subspace.

Let QPQ_P denote the principal-subspace basis and VresV_{\mathrm{res}} the focused residual-subspace basis. Since both are extracted from the same eigendecomposition, QP⊤Vres=0Q_P^\top V_{\mathrm{res}}=0. Thus, the initial direction vres=Vreszv_{\mathrm{res}}=V_{\mathrm{res}}z is strictly orthogonal to the principal subspace.

However, projecting this direction onto the local tangent space of the known-class center ckc_k gives

\left(v_{\mathrm{res}}-\langle v_{\mathrm{res}},c_k\rangle c_k\right).$$ As correctly noted by the reviewer, this operation may reintroduce principal-subspace components: $$Q_P^\top\left(v_{\mathrm{res}}-\langle v_{\mathrm{res}},c_k\rangle c_k\right) =-\langle v_{\mathrm{res}},c_k\rangle Q_P^\top c_k.$$ **Nevertheless, our additional geometric analysis shows that the practical magnitude of this deviation is small.** Across all five RWD datasets and both ViT-B/16 and ViT-L/14 backbones, the mean principal-subspace energy ratios, measured as $E_{\mathcal P}=\\|Q_P^\top v^\perp\\|_2^2$, range from $4.162\times10^{-4}$ to $7.345\times10^{-4}$. These results indicate that although tangent-space projection can reintroduce principal-space components, the resulting directions retain low principal-subspace alignment on average. We have corrected the strict inclusion in Sec. 3.1 (Eq. 1), clarified the projection in Sec. 3.2, and added the complete quantitative results to the Appendix under *Principal-Subspace Energy After Tangent Projection*. **A2. Boundary-pruning effectiveness.** We agree that the similarity threshold in Eq. (4) constrains candidate alignment with non-parent known-class centers, rather than guaranteeing separation from complete known-class distributions or decision boundaries. To further examine the practical effect of this constraint, we measured candidate rejection rates and prototype-to-known-instance similarities before and after pruning, prior to semantic selection. Specifically, for each candidate prototype, we compute its maximum cosine similarity to known instances excluding its parent class, and then average these maxima over the candidate set. The ViT-L/14 results for domains with nonzero rejection rates are summarized below: | Dataset | Rejection Rate | Before | After | |---|---:|---:|---:| | Aerial | 4.52% | 0.82 | 0.81 | | Medical | 31.19% | 0.84 | 0.83 | | Surgery | 40.00% | 0.82 | 0.81 | The pruning mechanism rejects candidates with excessive similarity to other known-class centers, with rejection rates varying across the evaluated domains. These differences may reflect variations in inter-class feature proximity. Importantly, the average maximum similarity to non-parent known instances also decreases after pruning, providing additional evidence that the center-based constraint helps reduce potential interference with known instances. These observations are further supported by the existing ablation in Table 2: removing Boundary Pruning decreases Aerial U-mAP from **12.6 to 10.7** and Medical K-mAP from **28.3 to 25.3**, demonstrating its contribution to unknown discovery and known-class preservation. We have also clarified that unit normalization guarantees a representation on the hypersphere, but does not by itself establish visual objectness. Accordingly, we characterize the proposed constraints as mechanisms for reducing potential cross-class interference and guiding prototype construction, rather than guaranteeing collision-free exploration. **Manuscript revisions.** We have clarified the residual-subspace and tangent-space constraints in Secs. 3.1–3.2 and included the complete principal-subspace energy analysis in the Appendix. We have also added the boundary-pruning diagnostics to the Appendix and refined the corresponding statements concerning visual objectness and boundary safety in Secs. 3.2–3.4. We appreciate the reviewer's suggestions, which have helped us strengthen the quantitative evidence and provide a more precise characterization of MGME's geometric constraints.
Reviewer_1HFCOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Manifold-Guided Multi-Prototype Expansion (MGME) for open-world object detection. The authors argue that representing unknown objects with the mean of known-class embeddings or a single text prompt causes “mean collapse” and cannot capture diverse unknown semantics. MGME constructs a residual subspace from known visual features, generates multiple unknown prototypes from known-class centers through spherical traversal, and selects candidates using cross-class boundary pruning and LLM-based semantic guidance. It also introduces an orthogonalization loss for known-class text prototypes and aggregates the responses of multiple unknown prototypes during inference. Experiments on five domains of the RWD benchmark show improvements over FOMO, UMB, and PASS, especially in overall unknown-object detection performance.

Critical strengths
  1. The paper addresses an important problem in open-world object detection. Representing the unknown space with multiple prototypes is intuitively more suitable than using a single mean vector, since unknown objects may have diverse and multimodal semantics.

  2. The proposed framework is reasonably comprehensive. It combines prototype generation, boundary pruning, semantic selection, training regularization, and inference-time aggregation in a unified pipeline.

  3. The empirical results are promising. MGME improves the overall U-mAP from 6.8 to 11.8 with ViT-B/16 and from 15.9 to 19.2 with ViT-L/14 compared with the strongest reported baselines. The improvements are particularly noticeable on the Game, Medical, and Aerial datasets.

  4. The experimental section covers five distinct application domains and two OWL-ViT backbones. The appendix also provides pseudocode, hyperparameters, sensitivity studies, qualitative results, and failure cases, which improves the clarity of the presentation.

Critical weaknesses
  1. The central geometric claim is not adequately supported. For orthogonal unit vectors, the geodesic construction in Eq. (3) is equivalent to normalized Euclidean addition when the Euclidean step size is chosen to produce the same angular displacement. Therefore, the large performance difference between geodesic escape and “Euclidean Add.” in Table 2 may result from unmatched step sizes or missing normalization rather than an inherent advantage of geodesic traversal. A strictly controlled comparison is needed to support the main geometric claim.

  2. The residual-subspace argument is mathematically incomplete. The method first samples a direction from the residual subspace and then projects it onto the local tangent space using Eq. (2). If the class center contains components in the principal subspace, this operation can reintroduce principal-space components into the final direction. Therefore, the resulting direction is not necessarily contained in the orthogonal residual subspace claimed in Eq. (1).

Questions for authors

What is the exact implementation of “Euclidean Add.” in Table 2? Was the resulting vector L2-normalized, and was its step size selected to match the angular displacement in Eq. (3)? Please include a controlled comparison between Eq. (3) and normalized Euclidean addition with the same angular displacement.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 1HFC (Part 1/3)

We sincerely thank the reviewer for the careful and constructive assessment of our work and for recognizing the motivation, comprehensive framework, and promising empirical improvements of MGME. We particularly appreciate your mathematically precise concerns regarding our geometric formulation. We address the two main weaknesses separately and provide additional controlled comparisons.

Q1(W1): Mathematical equivalence between geodesic traversal and normalized Euclidean addition, and clarification of the Euclidean-addition ablation.

A1. Mathematical equivalence. We agree that, for a unit class center ckc_k and a unit tangent direction v⊥v^\perp satisfying ck⊤v⊥=0c_k^\top v^\perp=0, the spherical exponential map in Eq. (3) is mathematically equivalent to normalized Euclidean addition with a matched angular displacement:

=\cos\theta\\,c_k+\sin\theta\\,v^\perp =\operatorname{Normalize}(c_k+\tan\theta\\,v^\perp).$$ Thus, when the class center, tangent direction, angular displacement, and subsequent processing are identical, these formulations produce the same prototypes and cannot yield an intrinsic performance difference. We acknowledge that the exponential-map parameterization itself should not be regarded as superior to its mathematically equivalent normalized additive form. **A2. Clarification of the Euclidean-addition variant.** We apologize for the ambiguity caused by the term "Euclidean Add." in Table 2 and the insufficient description of its implementation. This ablation was intended to evaluate an alternative prototype-construction strategy against MGME's residual-guided geometric construction, rather than compare two mathematically equivalent formulations. Specifically, the evaluated variant employs Euclidean addition along the difference between neighboring known-class centers. For each center $c_k$, we select its nearest other known-class center $c_{j^*}$ according to cosine similarity and construct $$\begin{aligned} d_k &= c_{j^*}-c_k,\\\\ p_k^{\mathrm{Euc}} &=\operatorname{Normalize}\left( c_k+0.5[d_k+0.1\\|d_k\\|_2\xi] \right), \end{aligned}$$ where $\xi$ is a randomly sampled unit vector. **The resulting prototype is L2-normalized. However, the coefficient 0.5 was not selected to match the angular displacement $\theta$ in Eq. (3).** The variant also differs in exploration directions, candidate-sampling budget, prototype count, and selection procedure. Therefore, Table 2 evaluates different prototype-construction strategies rather than a strictly matched comparison of equivalent traversal formulations. **A3. Interpretation of the results.** As shown in Table 2, this variant achieves U-mAP of **5.6 on Aerial and 1.8 on Medical**, compared with **12.6 and 15.1** for MGME. These results support the effectiveness of MGME's overall prototype-construction strategy relative to this particular inter-class Euclidean-addition alternative, rather than demonstrating an intrinsic advantage of the exponential map over equivalent normalized addition. To directly address the reviewer's request for a controlled comparison and further isolate the contribution of our construction strategy, **we provide two additional sets of comparisons in Part 2/3**. **Manuscript revisions.** We have clarified the mathematical equivalence in **Sec. 3.3**, renamed the ablation **"w/o Geodesic Escape (Inter-class Euclidean Add.)"** in **Table 2**, and added its construction formula and interpretation in **Sec. 4.3**.
AuthorsOpenReview ↗

Response to Reviewer 1HFC (Part 2/3)

Q1 (W1) (continued): Controlled comparisons of equivalent formulations and alternative prototype constructions.

Following the reviewer's suggestion, we conducted two complementary experiments: (1) a strictly matched comparison between Eq. (3) and normalized Euclidean addition, evaluating both prototype-level equivalence and end-to-end detection performance; and (2) additional comparisons with alternative prototype-construction strategies under a fixed adapted checkpoint, providing more controlled comparisons than the original inter-class Euclidean-addition variant.

A4. Matched geodesic vs. normalized Euclidean addition. To directly address the reviewer's request, we compared

p_{\mathrm{geo}} &=\operatorname{Normalize} (\cos\theta\\,c_k+\sin\theta\\,v^\perp),\\\\ p_{\mathrm{add}} &=\operatorname{Normalize} (c_k+\tan\theta\\,v^\perp), \end{aligned}$$ using **identical class centers, unit tangent directions, and angular displacements**, with L2 normalization applied to both formulations. The experiment covered **43,008 sampled tangent directions** and six angles ($20^\circ,24^\circ,28^\circ,32^\circ,36^\circ,40^\circ$), yielding **258,048 matched prototype pairs**. | Comparison | Maximum Absolute Difference | |---|---:| | Prototype element-wise difference | $1.79\times10^{-7}$ | | End-to-end U-AP difference | 0.040587 | | End-to-end K-AP difference | 0.019161 | The two formulations generate virtually identical prototypes up to floating-point precision. More importantly, **their end-to-end detection performance also remains nearly unchanged**, with maximum absolute U-AP and K-AP differences of only 0.040587 and 0.019161, respectively. These results provide direct numerical and detection-level evidence for the reviewer's observation: **when the exploration direction and angular displacement are matched, geodesic traversal and normalized Euclidean addition are effectively equivalent.** Therefore, MGME's performance advantage should not be attributed to the exponential-map parameterization itself, but to its structured unknown-prototype construction. **A5. More controlled comparisons with alternative prototype constructions.** As clarified in Part 1/3, the original inter-class Euclidean-addition variant differs from MGME in exploration directions, angular displacements, and prototype-generation settings. To provide a more consistent comparison and further isolate the contribution of our construction strategy, we conducted additional experiments using the **same fixed adapted ViT-L/14 checkpoint, detection head, and evaluation protocol**. We considered three alternative constructions: - **Known Centers:** Directly reuse one known-class center per unknown-prototype query. - **Random Tangent:** Generate normalized prototypes using randomly sampled tangent-space directions rather than residual-guided directions. - **Direct LLM:** Directly use LLM-generated domain concept embeddings as unknown prototypes. Random Tangent and Direct LLM use prototype counts matched to MGME, whereas Known Centers uses one prototype per known-class center. Unlike the original Table 2 variant, these comparisons eliminate differences arising from model adaptation and provide better-controlled evaluations of alternative unknown representations. The Task-1 U/K-mAP results on ViT-L/14 are summarized below: | Method | Aquatic | Aerial | Game | Medical | Surgery | Overall | |---|---|---|---|---|---|---| | Known Centers | 15.5/53.8 | 11.4/42.9 | 27.9/20.4 | 9.0/22.6 | 12.8/39.5 | 15.3/35.9 | | Random Tangent | 16.4/54.1 | 11.8/42.9 | 28.2/21.1 | 9.7/23.0 | 13.0/39.8 | 15.8/36.2 | | Direct LLM | 5.5/54.1 | 3.0/42.9 | 9.1/21.8 | 0.0/23.8 | 3.0/39.9 | 4.1/36.5 | | **MGME** | 20.6/52.8 | 12.6/42.0 | 34.4/22.0 | 15.1/28.3 | 13.5/45.5 | **19.2/38.1** | MGME improves overall U/K-mAP by **+3.44/+1.92** over Random Tangent and **+3.92/+2.26** over Known Centers. In particular, its U-mAP improvements over Random Tangent reach **+6.22 on Game** and **+5.42 on Medical**. Direct LLM achieves only 4.13 overall U-mAP, suggesting that directly using domain-level text concepts is insufficient to reproduce the effectiveness of the proposed geometric construction. **Overall interpretation.** The first experiment confirms that the exponential map and its matched normalized-addition form are equivalent, both numerically and in end-to-end detection performance. The second experiment provides additional evidence that MGME's performance gains arise from its structured, residual-guided prototype construction rather than the choice of an equivalent geometric parameterization. **Manuscript revisions.** We have clarified the mathematical equivalence in Sec. 3.3 and the original Euclidean-addition variant in Sec. 4.3. We have also added the alternative prototype-construction comparisons and analysis in **Sec. 4.4 (Table 3)**.
AuthorsOpenReview ↗

Response to Reviewer 1HFC (Part 3/3)

W2: Does tangent-space projection preserve orthogonality to the global principal subspace?

A1. Clarification of the two orthogonality constraints. We thank the reviewer for identifying this mathematical distinction. We agree that the tangent-space projection in Eq. (2) does not necessarily preserve strict orthogonality to the global principal subspace.

Let QP=[q1,…,qr]Q_P=[q_1,\ldots,q_r] denote the principal-subspace basis and Vres=[qr+1,…,qr+Dres]V_{\mathrm{res}}=[q_{r+1},\ldots,q_{r+D_{\mathrm{res}}}] denote the focused residual-subspace basis. Since both are extracted from the same eigendecomposition, they satisfy

QP⊤Vres=0.Q_P^\top V_{\mathrm{res}}=0.

Thus, the initially sampled direction vres=Vreszv_{\mathrm{res}}=V_{\mathrm{res}}z is strictly orthogonal to the principal subspace.

However, the subsequent projection onto the local tangent space of the known-class center ckc_k gives

\operatorname{Normalize}\left( v_{\mathrm{res}}-\langle v_{\mathrm{res}},c_k\rangle c_k \right).$$ As correctly noted by the reviewer, this operation may reintroduce principal-subspace components. Specifically, before normalization, $$Q_P^\top\left(v_{\mathrm{res}}-\langle v_{\mathrm{res}},c_k\rangle c_k\right) = -\langle v_{\mathrm{res}},c_k\rangle Q_P^\top c_k.$$ Therefore, the resulting direction satisfies the local tangent-space constraint but does not necessarily retain exact orthogonality to the global principal subspace. **A2. Additional geometric analysis.** To quantify the practical magnitude of this deviation, we conducted an additional geometric analysis across all five RWD datasets and both ViT-B/16 and ViT-L/14 backbones. We measured the principal-subspace energy ratio of the resulting unit tangent directions: $$E_{\mathcal P}=\\|Q_P^\top v^\perp\\|_2^2.$$ The mean energy ratios are summarized below: | Dataset | ViT-B/16 | ViT-L/14 | |---|---:|---:| | Aerial | $5.496\times10^{-4}$ | $6.167\times10^{-4}$ | | Aquatic | $4.162\times10^{-4}$ | $5.699\times10^{-4}$ | | Game | $6.563\times10^{-4}$ | $6.417\times10^{-4}$ | | Medical | $6.683\times10^{-4}$ | $7.345\times10^{-4}$ | | Surgery | $6.453\times10^{-4}$ | $4.963\times10^{-4}$ | Across all ten configurations, the mean principal-subspace energy ratios range from $4.162\times10^{-4}$ to $7.345\times10^{-4}$, indicating that tangent-space projection introduces only a small principal-space component on average. These results support our intended **residual-initialized tangent-space exploration**: exploration directions are initially sampled from the focused residual subspace and subsequently projected onto the local tangent space to enable spherical traversal. **Manuscript revisions.** We have corrected the strict subspace inclusion in **Sec. 3.1 (Eq. 1)**, clarifying that the final direction belongs to $T_{c_k}\mathcal M$ without requiring it to remain in the original residual subspace. In **Sec. 3.2 (following Eq. 2)**, we have explicitly explained how tangent-space projection may reintroduce principal-subspace components. We have also added the quantitative results to the revised Appendix, under **"Principal-Subspace Energy After Tangent Projection"** in the **Extended Analysis** section. These revisions distinguish exact initial residual-space orthogonality from the final local tangent-space constraint, while the additional measurements demonstrate that the reintroduced principal-space component is small on average. We sincerely appreciate the reviewer's insightful comments. We hope that the mathematical clarifications, controlled numerical verification, additional prototype-construction comparisons, and geometric measurements provided across these three responses address your concerns and support a reconsideration of the overall assessment.

The authors' response has partially addressed my concerns. I raise my score to weak accept.

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

MGME builds unknown-class prototypes for open-world detection on a frozen OWL-ViT pipeline. From the covariance of normalized known features it takes a band of residual eigenvectors, samples directions in that band, projects them onto the tangent space of each known center, and steps along the sphere with the map cos(θ)c + sin(θ)v. Candidates whose cosine with any other known center exceeds a threshold are dropped, and the survivor most aligned with the mean embedding of LLM-proposed domain concepts is kept. Training adds an orthogonality penalty on assembled text prototypes. Inference still multiplies the unknown score by the baseline attribute-relevance and known-class outlier gates. The method is trained and evaluated on the five RWD datasets with ViT-B/16 and ViT-L/14.

Critical strengths

S1. On a shared OWL-ViT implementation, MGME raises overall unknown mAP from 6.8 to 11.8 on ViT-B/16 and from 15.9 to 19.2 on ViT-L/14, while overall known mAP also rises (Table 1: 18.3 to 19.2, and 36.9 to 38.1). Appendix C.3 states that FOMO, UMB, and PASS were retrained in this pipeline with the same splits, support samples, backbones, attribute pools, and evaluation code. The largest cells, such as Medical B/16 unknown mAP 9.1 against FOMO 2.1, are large enough to matter if they replicate.

S2. The component ablations separate the spherical step from the selection heuristics. On ViT-L/14, removing the geodesic step drops Aerial unknown mAP from 12.6 to 5.6 and Medical unknown mAP from 15.1 to 1.8 (Table 2). Removing semantic steering leaves Aerial at 10.2 and Medical at 14.8, still above the corresponding PASS and UMB entries in Table 1. This is the evidence that most directly supports a geometric change inside the pipeline.

Critical weaknesses

W1. The paper’s explanation does not match the mechanism it measures. Section 4.2 says FOMO, UMB, and PASS anchor the unknown space with the arithmetic mean of known embeddings. FOMO instead scores an unknown by the product of attribute in-distribution and known-class out-of-distribution (Zohar et al., arXiv:2312.05745, Section 4.4). PASS selects attributes by partial optimal transport rather than by a class-mean prototype (Yang et al., CVPR 2025). Appendix A.3 says MGME inherits both gates unchanged and changes the unknown logit and the orthogonality loss. Table 4 further shows that one geodesic prototype already reaches 11.3 unknown mAP on Aerial, against 12.6 for three prototypes, while Algorithm 1 selects every slot by the same maximum alignment with t_domain and does not repel previously chosen prototypes. The mean-collapse claim is therefore not isolated. The missing control is the same detection head with the unknown prototype replaced by the normalized mean of the known centers, together with the exact formula of the Euclidean-add ablation.

W2. The unknown-discovery claim is not yet auditable, and the stability claim is stronger than the table. Semantic steering depends on a fixed list of LLM concepts, but that list, its prompt, and its overlap with the Task-1 unknown names are not in the paper. Appendix E.2 says the exact model and prompts are reported; the only named generator in Appendix C.3 is GPT-3.5 for the baseline attribute pools. Qualitative examples include unknowns such as Radius, Ulna, Scalpel, and Tennis court, which a domain prompt can easily name. There is also no run-to-run interval. Table 1’s macro gain coexists with Medical B/16 current-known mAP of 3.1 against FOMO’s 11.5, and Surgery L/14 unknown mAP of 13.5 against PASS’s 16.6. Section 4.2 describes these degradations as minor and discusses the Surgery case. A reader cannot yet tell how much of the unknown gain is geometric, how much is the concept list, or how stable either part is.

Questions for authors

Q1. In the same OWL-ViT head, what are unknown and known mAP if the generated set is replaced by the normalized mean of the known centers, and if the prototype logit is removed so that only the inherited p_ID and p_OOD remain? Please also give the Euclidean-add formula and its step size. (W1.) A gap similar to Table 1 under that mean control would support the geometric change without relying on the mean-collapse account of FOMO and PASS. A gap that disappears would make the current explanation the wrong basis for acceptance.

Q2. What LLM, prompt, and fixed concept list were used for t_domain on each dataset, and how many of those strings match Task-1 unknown class names? (W2.) Low overlap, with most of the unknown mAP retained after the overlapping strings are removed, would reduce the leakage concern. High overlap, or a large drop after removal, would mean semantic steering is supplying the unknown vocabulary.

Q3. Were U, θ, τ_safe, and D_res selected on the same Task-1 splits reported in Table 1, and are those entries single runs? Intervals over seeds for the overall unknown mAP and for the Medical B/16 and Surgery L/14 cells would show whether the macro gain survives the negative cells.

4 comments
AuthorsOpenReview ↗

Response to Reviewer ndCN (Part 1/4)

We sincerely thank the reviewer for the detailed and constructive assessment of our work and for recognizing MGME's improvements in open-world detection. We appreciate the opportunity to clarify the relationship between existing attribute-based methods and our unknown-prototype construction.

Q1 (W1): Clarification of mean-based unknown modeling and the contribution of MGME.

A1. Attribute modeling, unknown-prototype construction, and unknown inference. These are three distinct but mutually compatible components of attribute-based OWOD frameworks. FOMO employs attribute selection and refinement, UMB additionally models attribute-dependent empirical probabilities, and PASS introduces partial optimal transport (POT) for attribute selection and alignment. These mechanisms primarily determine how attribute embeddings EattE_{\mathrm{att}} and their assignment weights WW represent known classes.

As described in Sec. 3.5, the assembled known-class representations are

Pcls=W⊤Eatt∈RK×d.P_{\mathrm{cls}}=W^\top E_{\mathrm{att}}\in\mathbb R^{K\times d}.

Unknown-prototype construction is a separate operation. FOMO, UMB, and PASS construct a single unknown proxy by averaging the assembled known-class representations:

pu=1K∑k=1KPcls[k,:].p_u=\frac1K\sum_{k=1}^{K}P_{\mathrm{cls}}[k,:].

Thus, their different attribute modeling strategies do not contradict their shared mean-based unknown-prototype construction. This is the intended meaning of our statement in Sec. 4.2 that these methods rely on "the arithmetic mean of known embeddings to anchor the unknown space." We apologize that the original wording may have conflated unknown-prototype construction with the complete unknown-scoring mechanisms. We have clarified this distinction in the revised manuscript.

A2. What MGME changes. Building upon the existing attribute-based detection framework, MGME introduces the orthogonalization loss Lortho\mathcal L_{\mathrm{ortho}} to reduce entanglement among assembled known-class text prototypes and improve their discriminability. More importantly, we replace the single unknown proxy pup_u with a set of geometrically constructed prototypes Pu=pii=1U\mathcal P_u=\\{p_i\\}_{i=1}^{U}.

At the raw-logit level of Eq. (7), the corresponding unknown responses are

L_U^*(v)=\max_i\cos(v,p_i).$$ As in previous methods, when only one unknown proxy is used, the maximum-response aggregation reduces to its sole response. MGME extends this formulation to multiple unknown prototypes. **The unknown-inference formulation remains consistent with previous methods**, as expressed in Eq. (8): $$P_U(v)=\sigma(L_U^*(v))\\,p_{\mathrm{ID}}(v)\\,p_{\mathrm{OOD}}(v).$$ As also clarified in Appendix A.3, the inherited formulations of $p_{\mathrm{ID}}$ and $p_{\mathrm{OOD}}$ remain unchanged. MGME modifies the unknown representations underlying $L_U^*$ and introduces $\mathcal L_{\mathrm{ortho}}$, rather than redesigning the inference gates. Our contribution therefore lies in **constructing more expressive unknown representations within the existing detection framework**, using residual-guided geometric exploration, boundary pruning, and semantic selection. **Manuscript revisions.** We have clarified the mean-based unknown-proxy construction and its multi-prototype extension in Sec. 3.5, including their integration into Eq. (7). We have also revised Sec. 4.2 to distinguish attribute modeling, unknown-prototype construction, and unknown inference.
AuthorsOpenReview ↗

Response to Reviewer ndCN (Part 2/4)

Q1 (W1, continued): Direct mean-prototype and inherited-gate controls.

A3. Additional controls. Following the reviewer's suggestion, we compare two additional variants in the same OWL-ViT detection head:

BASE-MEAN: Replace the generated unknown-prototype set with the normalized mean of known class centers, pmean=Normalize⁡(1K∑kck)p_{\mathrm{mean}}=\operatorname{Normalize}(\frac1K\sum_k c_k), while retaining the unknown-logit term and inherited gates.

BASE-GATES: Remove the prototype-logit factor σ(LU∗)\sigma(L_U^*), retaining only the inherited pIDp_{\mathrm{ID}} and pOODp_{\mathrm{OOD}} gates.

The Task-1 U/K-mAP results are summarized below.

MethodAQ UAQ KAE UAE KGame UGame KMedical UMedical KSurgery USurgery KOverall UOverall K
ViT-B/16
BASE-MEAN7.541.70.017.217.36.14.68.810.315.57.917.9
BASE-GATES5.48.42.811.426.70.00.21.54.73.47.94.9
MGME9.444.92.118.922.59.49.18.115.814.611.819.2
ViT-L/14
BASE-MEAN18.453.83.442.932.121.34.723.316.539.615.036.2
BASE-GATES5.10.015.412.94.20.01.60.08.01.56.82.9
MGME20.652.812.642.034.422.015.128.313.545.519.238.1

Compared with BASE-MEAN, MGME improves overall U/K-mAP by +3.9/+1.3 on B/16 and +4.2/+1.9 on L/14. These results provide more direct evidence for the effectiveness of MGME's unknown-prototype construction relative to a single normalized mean proxy, addressing the reviewer's concern without relying solely on the mean-collapse interpretation of FOMO and PASS.

BASE-GATES also shows substantially lower overall performance, particularly in known detection. Removing the prototype-logit factor LU∗L_U^* increases the relative ranking of unknown predictions, reducing the number of retained known detections. This indicates that prototype responses and the inherited gates play complementary roles in the detection framework, further supporting the importance of explicit unknown-prototype representations.

A4. Prototype quantity and diversity. We clarify that UcU_c in Appendix Table 4 denotes the number of prototypes generated per known-class center, rather than the total number of unknown prototypes. The total number satisfies Utotal≤KUcU_{\mathrm{total}}\leq K U_c after candidate selection. Thus, Uc=1U_c=1 already represents a multi-prototype configuration, unlike a single global mean proxy.

Increasing UcU_c from 1 to 3 improves Aerial U/K-mAP from 11.3/39.6 to 12.6/42.0, while Medical changes from 15.4/25.3 to 15.1/28.3. This indicates that one prototype per known-class center already provides substantial detection capability, whereas increasing the number offers dataset-dependent benefits and may introduce redundancy.

Regarding Algorithm 1, we agree that the same domain-level semantic attractor tdomaint_{\mathrm{domain}} is used across prototype slots and that no explicit pairwise repulsion is imposed. We therefore do not claim that the current selection rule guarantees pairwise prototype diversity.

A5. Euclidean-addition ablation. We have also clarified the implementation requested by the reviewer. For each known-class center ckc_k, its nearest other known-class center cj∗c_{j^*} is selected, and the unknown prototype is generated as

\operatorname{Normalize}\left( c_k + 0.5[(c_{j^{\ast}} - c_k) + 0.1\lVert c_{j^{\ast}} - c_k\rVert_2\xi] \right),$$ where $\xi$ is a random unit vector. This variant evaluates an alternative inter-class Euclidean-addition strategy against MGME's residual-guided, angularly controlled prototype construction. The interpolation coefficient is 0.5, followed by L2 normalization. Its lower U-mAP in Table 2 supports the effectiveness of MGME's overall prototype-construction strategy relative to this particular alternative. **Manuscript revisions.** We have added the normalized-mean and gates-only controls to the Appendix D.5 and referenced them in Sec. 4.2; clarified prototype counts and the selection rule in the Appendix; and specified the Euclidean-addition formula and purpose in Sec. 4.3. We appreciate the reviewer's suggestions, which have helped us clarify the contribution of unknown-prototype construction and provide more targeted empirical evidence for MGME.
AuthorsOpenReview ↗

Response to Reviewer ndCN (Part 3/4)

Q2 (W2): LLM concept vocabulary, overlap with unknown categories, and its influence on unknown discovery.

A1. Concept vocabulary and provenance. We thank the reviewer for pointing out the insufficient documentation of the semantic concept vocabulary. We apologize for this omission.

To maintain consistency with the FOMO-based evaluation protocol, MGME reuses fixed, pre-generated domain-relevant concept lists rather than generating a new vocabulary specifically for our method. The same lists are used by BASE-ZS+LLM, ensuring consistent semantic inputs.

We clarify that attribute generation and unknown-class proposals serve different purposes. The GPT-3.5 configuration in Appendix C.3 concerns attribute-pool construction, whereas MGME's semantic steering directly uses the fixed concept lists without additional LLM queries. Unlike BASE-ZS+LLM, which employs proposed class names as detection queries, MGME aggregates their text embeddings into a domain-level semantic attractor tdomaint_{\mathrm{domain}} to guide the selection of geometrically generated unknown prototypes.

Nevertheless, we acknowledge that overlap with actual unknown-class names may influence semantic selection.

A2. Vocabulary overlap and exact-name removal. We have provided the complete concept lists in the revised Appendix, including their exact-name overlaps with Task-1 unknown classes:

DatasetMatches / ConceptsOverlapping names
Aquatic3 / 12Puffin, Stingray, Starfish
Aerial2 / 21Harbor, Bridge
Game3 / 20Ekko, Fiora, Tahm Kench
Medical2 / 53Radius, Ulna
Surgery2 / 22Scissors, Hook

Vocabulary sizes include the generic object or objects entry. In total, 12 exact-name overlaps are identified across 128 concept entries.

Following the reviewer's suggestion, we removed all identified exact-name overlaps and re-evaluated Task-1 unknown detection:

BackboneVariantAQAEGameMedicalSurgeryOverall
B/16MGME9.42.122.59.115.811.8
B/16Exact-name removal8.62.420.79.115.411.2
L/14MGME20.612.634.415.113.519.2
L/14Exact-name removal18.411.032.415.213.818.1

After removing these overlapping names, MGME retains approximately 95.4% and 94.3% of its originally reported overall U-mAP on ViT-B/16 and ViT-L/14, respectively.

Moreover, the resulting overall U-mAP remains above the normalized-mean controls reported in our Q1 response: 11.2 vs. 7.9 on B/16 and 18.1 vs. 15.0 on L/14.

These results demonstrate that MGME's unknown-detection improvements are not solely attributable to the presence of exact unknown-class names in the semantic vocabulary. The remaining domain-relevant concepts can still guide the selection of geometrically generated unknown prototypes, while the residual-guided construction provides the candidate representations.

Manuscript revisions. We have clarified the use of fixed concept lists and their distinction from attribute-generation prompts in Sec. 3.4 and Appendix C.3. We have also included the complete concept vocabularies, exact-name overlap statistics, and corresponding removal experiments in the Appendix.

We appreciate the reviewer's suggestion, which has helped us clarify the semantic guidance mechanism and provide direct evidence of its robustness to exact unknown-label overlap.

AuthorsOpenReview ↗

Response to Reviewer ndCN (Part 4/4)

Q3 (W2): Hyperparameter selection and run-to-run stability.

We thank the reviewer for raising these important questions regarding hyperparameter selection and the reliability of the reported improvements.

A1. Hyperparameter selection. We clarify that the reported hyperparameters were not selected using the Task-1 test set. Following the FOMO/RWD evaluation protocol, we preserve the same Task-1 class partitions and use a separate validation subset drawn from the training data for model selection.

Specifically, the number of prototypes per known class UcU_c and the initial geodesic angle θstart\theta_{\mathrm{start}} were selected based on known-class AP (K-AP) on the held-out validation subset, rather than unknown AP on the test set.

The residual-subspace dimension Dres=150D_{\mathrm{res}}=150 was instead fixed heuristically as a shared residual-band width for both embedding dimensions (d=512d=512 and d=768d=768), without an additional parameter sweep. Likewise, the safety threshold τsafe=0.92\tau_{\mathrm{safe}}=0.92 was retained as a fixed design setting. Neither parameter was optimized using the reported test results.

The Task-1 sensitivity analyses in the Appendix characterize the effects of different configurations, rather than serving as the basis for test-set hyperparameter selection. No hyperparameter was selected based on the Task-1 test results reported in Table 1.

A2. Run-to-run stability and negative cases. We agree that aggregate gains should be supported by run-to-run variability analysis. We therefore conducted additional repeated-run evaluations and computed the standard deviations of Task-1 U/K-mAP across all five datasets and both backbones.

BackboneMetric SDAquaticAerialGameMedicalSurgery
B/16U-mAP0.04270.04330.22250.00000.3789
B/16K-mAP0.05420.02010.05970.02090.0309
L/14U-mAP0.07850.07570.51070.38650.1737
L/14K-mAP0.01390.01470.04890.01060.1405

The U-mAP standard deviations are consistently small across the evaluated datasets and backbones, with a maximum of 0.5107 AP points. These limited fluctuations across repeated runs provide further evidence for the stability and reproducibility of MGME's unknown-detection performance, supporting the overall U-mAP improvements reported in Table 1.

We also examined the two negative cases highlighted by the reviewer:

  • Medical B/16 (Task-2 CK-mAP): MGME achieves 3.1 versus FOMO's 11.5, with an additional CK-mAP standard deviation of 0.1680 (PK-mAP SD: 0.1200).
  • Surgery L/14 (Task-1 U-mAP): MGME achieves 13.5 versus PASS's 16.6, with a U-mAP standard deviation of 0.1737.

These cases also exhibit low run-to-run variability. Thus, we acknowledge the observed degradations as domain-specific performance limitations, rather than dismissing them as unstable measurements.

Fine-grained Medical and Surgery categories present challenging visual and semantic similarities, which may contribute to difficult discrimination trade-offs. However, the available results do not isolate a single cause for these degradations.

Importantly, stable overall improvements do not imply uniform superiority in every individual metric. The additional repeated-run results support the reproducibility of MGME's overall performance while confirming that localized limitations remain in challenging domains.

Manuscript revisions. We have clarified the validation-based and fixed-design hyperparameter settings in the experimental protocol and Appendix C.4. We have also added the run-to-run standard deviations to the Appendix and revised Sec. 4.2 to acknowledge the Medical and Surgery performance degradations.

We appreciate the reviewer's suggestions, which have helped us improve the transparency of our evaluation protocol and provide a more precise assessment of MGME's robustness.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers and the Area Chair

We sincerely thank all reviewers for their thoughtful and constructive feedback, as well as the Area Chair for coordinating the discussion. We greatly appreciate the reviewers' careful consideration of our rebuttal and their encouraging follow-up assessments.

Throughout the rebuttal process, we have carefully addressed the concerns raised by the reviewers and further refined our manuscript through additional experiments, theoretical clarifications, and implementation details. The major updates are summarized below.

1. Clarification of Contributions and Distinctions from Prior Methods.

We clarified the distinction between attribute modeling, unknown-prototype construction, and unknown inference in existing attribute-based OWOD approaches. Although prior methods employ different attribute modeling strategies, they construct a single unknown proxy by averaging assembled known-class representations. In contrast, MGME replaces this mean-based proxy with multiple geometrically constructed unknown prototypes while retaining the inherited inference formulation. Together with discriminative prototype orthogonalization, our main contribution lies in more expressive unknown representations rather than a redesigned unknown-scoring mechanism. We also provided direct mean-prototype and inherited-gate controls to further clarify these contributions.

2. Geometric Clarifications and Additional Controlled Experiments.

We clarified the mathematical equivalence between geodesic traversal and normalized Euclidean addition under matched angular displacements, and distinguished this equivalence from the alternative inter-class Euclidean-addition construction evaluated in our original ablation. We further conducted controlled comparisons with Random Tangent, Direct LLM, and Known Centers under a fixed adapted checkpoint, providing additional evidence for the effectiveness of residual-guided prototype construction. We also clarified the tangent-space orthogonality constraint and quantified principal-subspace alignment, boundary-pruning behavior, and prototype similarities to known instances.

3. Experimental Transparency and Robustness Analysis.

We disclosed the LLM-generated concept lists and their overlap with unknown class names, evaluated performance after removing overlapping concepts, and provided additional multi-run variability statistics. We also clarified the hyperparameter-selection protocol and distinguished validation-based choices from fixed design settings.

Together, these revisions provide additional empirical evidence for MGME and a more precise characterization of its contributions and geometric properties. The revised manuscript incorporates the corresponding experiments and discussions in the main text and Appendix, with modifications highlighted in blue for ease of reference.

We sincerely appreciate the reviewers' valuable suggestions and the Area Chair's efforts throughout the discussion. Their feedback has helped us improve the rigor, clarity, and reproducibility of our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.