acceptodds
Under review as a conference paper at ICLR 2027

Geometry-Guided Cross-View Reasoning in Multimodal Large Language Models

15 views

Abstract

Multi-view spatial reasoning requires MLLMs to associate observations of the same physical scene region across viewpoints. Yet existing geometry-aware MLLMs mainly enrich individual visual tokens with 3D or camera information, leaving such cross-view correspondences to emerge implicitly from appearance-driven attention. We introduce GeoMLLM, a geometry-guided MLLM that grounds cross-view information exchange in camera geometry. GeoMLLM equips each visual patch with a camera-ray representation and uses pairwise ray and epipolar cues to constrain which tokens can plausibly correspond across views. During training, depth-derived soft correspondences further supervise cross-view attention, teaching the model to route information between observations of the same physical scene region. Built on Qwen3-VL-4B, GeoMLLM achieves state-of-the-art performance on SPAR-Bench, reaching overall, with on multi-view tasks ( points) and on Cross-view Perception ( points). GeoMLLM also consistently improves over its matched backbone across additional multi-image and 3D spatial reasoning benchmarks. These results demonstrate that explicitly modeling camera geometry in cross-view token interactions provides a strong inductive bias for multi-view spatial reasoning.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multimodal Large Language Model, Spatial Intelligence, 3D Computer Vision

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_nWiWOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

To address the multi-view spatial reasoning in MLLMs, the paper proposes GeoMLLM, a novel framework that grounds cross-view information exchange in camera geometry. By focusing on correlations of tokens from different views, the paper tackles the ambiguity of cross view correspondence. On the method level, they employ a camera-ray embedding and geometry-aware attention, with a geometry-aware correspondence supervision. The model achieves exceptionally strong results for SPAR-Bench etc, beating flagship VLMs and human performance.

Critical strengths
  1. The research question is important. People in 3D VLMs and video understanding have been talking about or troubled by this problem. A solution for this will be impactful for the research community.
  2. The design is generally elegant. The Geometry bias, leveraging epipolar calculations, is a simple and effective way to inject geometric constraints into the model, which is worth recognition in academia.
  3. Exceptionally strong performance across benchmarks, which demonstrates the value of the method. Impressively, the performance was obtained from 3B/4B small models.
  4. Analysis of Figure 3 is thoughtful. It demonstrates that the model is still robust under camera perturbations, so predicted cameras are not expected to significantly decrease performance.
Critical weaknesses
  1. The first point of contribution is overclaiming. The “cross-view” issue has been a long-standing research problem. One example is Qwen3d [1].
  2. (minor) It is confusing to talk about Qwen2.5 results in intro while in abstract the results are reported with Qwen3. Consistency is needed when authors renew a paper with new experimental results.

[1] Lin, Lucy, et al. "Qwen-3d: A generalist 3d vision-language model for spatial understanding." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2026.

Questions for authors
  1. In Line 148, what is the reasoning behind suppressing the attention between two tokens from different views that does not correspond? In my opinion, one token in image0 may find extra information in image1 helpful.
  2. Could you please clarify Line 210? Why the second term is complementary to the first one? What is the underlying logic of the design of Eq.(2)?
3 comments
AuthorsOpenReview ↗

Response to Reviewer nWiW

Thank you for recognizing the importance of our research question, the geometric design, and the strong empirical results. We appreciate your constructive comments and suggestions.

Weakness 1 — Scope of the first contribution

The first point of contribution is overclaiming. The “cross-view” issue has been a long-standing research problem. One example is Qwen3d.

Authors’ Response

Thank you for highlighting this point.

  1. Contribution scope

    We agree that cross-view reasoning is a long-standing research problem and that our first contribution could be stated more precisely.

  2. Planned revisions

    • We will revise the first contribution and the corresponding statements in the introduction to more precisely describe our specific contribution within this established research area.
    • We will also explicitly acknowledge Qwen-3D in the related-work discussion.
  3. Comparison with GTA and PRoPE

    • To place our contribution within the established literature on cross-view reasoning, we additionally compare the full GeoMLLM model with GTA and PRoPE, two existing geometry-aware attention methods.

    • All methods use Qwen2.5-VL-3B with the same training datasets, evaluation datasets, and experimental environment. The table reports overall benchmark scores, with the full GeoMLLM results taken from manuscript Tables 1–2.

    MethodSPAR-Bench fullMMSI-BenchViewSpatial-BenchCV-Bench
    GTA70.0128.4741.8176.17
    PRoPE68.7028.0139.1281.68
    GeoMLLM (full)82.1832.1643.8182.57

Weakness 2 — Consistency of reported results

(minor) It is confusing to talk about Qwen2.5 results in intro while in abstract the results are reported with Qwen3. Consistency is needed when authors renew a paper with new experimental results.

Authors’ Response

Thank you for noting this inconsistency.

  1. Consistent primary results

    We will consistently foreground the Qwen3-VL-4B results on SPAR-Bench full in the abstract and introduction.

    Evaluation categoryResult
    Overall83.79
    Multi-view tasks83.49
    Cross-view Perception88.13
  2. Explicit result attribution

    We will explicitly identify the backbone and evaluation split associated with each result throughout the paper.

Question 1 — Motivation for the geometric attention bias

In Line 148, what is the reasoning behind suppressing the attention between two tokens from different views that does not correspond? In my opinion, one token in image0 may find extra information in image1 helpful.

Authors’ Response

Thank you for raising this point.

  • Motivation: The geometric bias reduces erroneous aggregation between visually similar but non-corresponding patches, especially under repeated structures or large viewpoint changes.
  • Complementary information: It is a soft prior rather than a hard mask, so the content-based score, residual connections, and downstream LLM can still use useful cross-view information.

We will clarify this point around Eqs. (2)–(3).

Question 2 — Complementary terms in the geometric attention bias

Could you please clarify Line 210? Why the second term is complementary to the first one? What is the underlying logic of the design of Eq.(2)?

Authors’ Response

Thank you for asking us to clarify the design of Eq. (2).

The two terms in Eq. (2) provide complementary geometric signals:

Bijh=η[−λψ(rij)vij+gh(ϕij)Δh(ϕij)],B_{ij}^{h}=\eta\left[-\lambda\psi(r_{ij})v_{ij}+g^{h}(\phi_{ij})\Delta^{h}(\phi_{ij})\right],

  • First term: The analytic epipolar penalty suppresses geometrically inconsistent pairs, but cannot distinguish all candidates near the same epipolar line.
  • Second term: The gated, head-specific residual uses richer geometry, such as ray angles, camera baseline, visibility, and degeneracy, to refine the bias. It also applies to same-view pairs, where the epipolar penalty is inactive.
  • Overall logic: The first term provides a stable geometric prior, while the second adapts it using learned pairwise information.

We will clarify this logic around Eq. (2).

Dear Reviewer nWiW,

Thank you again for your thoughtful review and constructive suggestions, which have been very helpful in improving our work.

We are writing to gently follow up on our rebuttal and would be grateful to hear whether our responses have addressed your main concerns.

We would be happy to provide any additional explanations or clarifications if needed.

Should you find our responses satisfactory, we would sincerely appreciate your consideration of updating your overall evaluation.

Thank you very much for your time and continued consideration.

Best regards,

Authors

Dear nWiW,

With just over 24 hours remaining before the discussion deadline, we would greatly appreciate your feedback on our responses. Thank you for your time!

Reviewer_6pUWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper presents GeoMLLM, a geometry-guided MLLM that grounds cross-view information exchange in camera geometry. The method represents each visual patch with a gated camera-ray representation (Equation 1) and uses pairwise ray and epipolar cues to constrain which tokens can plausibly correspond across views. The loss function is designed to enhance cross-view attention, enabling the model to route information between observations of the same physical scene region. The authors fine-tune Qwen3-VL-4B on the training sets of SPAR-7M-RGBD and LLaVA-Hound and report that it achieves state-of-the-art performance on SPAR-Bench. This paper reports cross-benchmark transfer on MMSI-Bench (general multi-image spatial reasoning), ViewSpatial-Bench (multi-view localization and reasoning), and CVBench (single-image 2D and 3D spatial understanding).

Critical strengths
  • The high-level idea of this proposed method (encoding camera-ray representation into each visual patch and learning their correspondence through attention scores) sounds reasonable.
  • The experiments show strong empirical results on the in-domain and out-of-domain settings.
Critical weaknesses
  • This paper lacks the detail in the proposed method, which makes me hesitate to recommend it (see Questions For Authors)
  • It is unclear which component (e.g., camera-ray encoding and loss over attention scores) contributes to the performance improvement.
Questions for authors

It was difficult for me to understand the details of the method because many variables are undefined. For example,

  • L141: What is CC? (dimensionality of embeddings?)
  • L157: What are upu_p and vpv_p?
  • L157: What are RkR_k and tkt_k?
  • L189: What is σ\sigma? (sigmoid function?)
  • L189: How is MLPgate{\rm MLP}_{\rm gate} constructed?
  • L206: What is hh? (index for attention head?)
  • L206: What are η\eta and λ\lambda?
  • L208: What is eˉij\bar{e}_{ij}?
  • L208: What are hhh^h and Δh\Delta^h?
  • L230: Are the parameters wa,cw_{a,c} trained by the loss function of Equation 5?
  • L233: Does the variable Ah,i,cA_{h,i,c} correspond to AijhA_{ij}^h in Equation 3?
  • L233: What is HH (the number of attention heads?)
  • L237: I don't see the underlying idea of why the loss function is defined by the head-averaged distribution even after reading L239-241.
  • L237: ϵ\epsilon is undefined.

How is the sensitivity of the hyperparameter βcorr\beta_{\rm corr} for the performance?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 6pUW

Thank you for recognizing the geometric motivation and empirical results. We address your comments below.

Weakness 1 and Question 1:

This paper lacks the detail in the proposed method, which makes me hesitate to recommend it (see Questions For Authors)

Authors’ Response

We thank you for highlighting the insufficient level of detail in the description of the proposed method. We will revise the method section as a whole to clarify the motivation, design, and connections among the three components. We will also define all symbols at first use and use consistent notation throughout. We address your specific questions below.

LineSymbol(s)Definition
L141CCChannel dimension of each visual patch token xk,px_{k,p}.
L157u~p\tilde{u}_pHomogeneous patch-center pixel coordinate (up,vp,1)⊤(u_p,v_p,1)^\top.
L157KkK_k, Tkc2wT_k^{c2w}Intrinsic matrix and camera-to-world transformation of view kk.
L157RkR_k, tkt_kRotation and translation, with Xw=RkXcam+tkX^w=R_kX^{cam}+t_k.
L189σ(⋅)\sigma(\cdot), [⋅,⋅][\cdot,\cdot], ⊙\odotElementwise sigmoid, concatenation, and elementwise multiplication.
L189MLP⁡gate\operatorname{MLP}_{gate}A two-layer fully connected network defined by MLP⁡gate(z)=W2GELU⁡(W1z+b1)+b2\operatorname{MLP}_{gate}(z)=W_2\operatorname{GELU}(W_1z+b_1)+b_2. The input z∈R2Cz\in\mathbb{R}^{2C} concatenates the visual token and ray embedding, each of dimension CC. The layer parameters are W1∈RC×2CW_1\in\mathbb{R}^{C\times 2C}, W2∈RC×CW_2\in\mathbb{R}^{C\times C}, and b1,b2∈RCb_1,b_2\in\mathbb{R}^{C}. Eq. (1) applies an elementwise sigmoid to the output of the second linear layer.
L189α\alphaLearned scalar controlling the ray-embedding residual magnitude.
L206hh, η\eta, λ\lambdaHead index, learned overall bias scale, and learned positive epipolar-penalty scale.
L208eˉij\bar e_{ij}Epipolar point-to-line distance normalized by patch scale (Appendix C.2).
L208hhh^h, Δh\Delta^hWe could not locate these two symbols at L208. If you are referring to gh(ϕij)g^h(\phi_{ij}) and Δh(ϕij)\Delta^h(\phi_{ij}) in Eq. (2) at L206, they denote a sigmoid gate and a head-specific scalar residual, respectively, computed by MLPs from ϕij\phi_{ij} (Appendix C.2).
L230wa,cw_{a,c}Fixed bilinear targets constructed from depth reprojection and normalized over valid neighbors. The loss optimizes model parameters producing attention probabilities (Appendix C.3).
L233Ah,i,cA_{h,i,c}Same as AichA^h_{ic} in Eq. (3), evaluated before dropout in the final connector layer, with source token ii and target token cc.
L233HHH=8H=8 attention heads.
L237ϵ\epsilonSmall positive constant for numerical stability in the logarithm.

L237: I don't see the underlying idea of why the loss function is defined by the head-averaged distribution even after reading L239-241.

Authors’ Response

Thank you for the comment. Averaging probabilities before the logarithm rewards the probability mass that the heads collectively assign to the geometric targets. This lets the heads share responsibility for matching the target distribution while retaining flexibility in their individual attention patterns. We will clarify this motivation in Section 3.3.

AuthorsOpenReview ↗

Response to Reviewer 6pUW

Weakness 2:

It is unclear which component (e.g., camera-ray encoding and loss over attention scores) contributes to the performance improvement.

Authors’ Response

Thank you for the constructive comment. We provide additional SFT-only results for Qwen2.5-VL-3B and Qwen3-VL-4B, alongside the MHSA connector variant and the component ablations in Table 3.

1. Experimental setup

The SFT-only baselines retain the native Qwen architectures and use the shared training mixture of LLaVA-Hound and SPAR examples. Both are trained for one epoch with a global batch size of 64 and a learning rate of 10−510^{-5}, with the vision encoder frozen.

2. Comparisons with SFT-only and MHSA baselines

BackboneConfigurationOverallMulti-viewSingle-viewCross-view
Qwen2.5-VL-3BPure Qwen + SFT65.0767.6766.7852.70
Qwen2.5-VL-3BGeoMLLM (MHSA)74.8875.8074.5373.07
Qwen2.5-VL-3BGeoMLLM (full)82.1882.3880.3986.37
Qwen3-VL-4BPure Qwen + SFT69.4171.5771.2857.93
Qwen3-VL-4BGeoMLLM (full)83.7983.4982.5188.13
  • Qwen2.5-VL-3B. Full GeoMLLM improves over the SFT-only baseline by 17.11 points overall and 33.67 points on cross-view tasks.
  • Qwen3-VL-4B. Full GeoMLLM improves over the SFT-only baseline by 14.38 points overall and 30.20 points on cross-view tasks.
  • Geometry-aware attention versus MHSA. With Qwen2.5-VL-3B, replacing standard multi-head self-attention in the connector with our geometry-aware attention improves the overall score from 74.88 to 82.18, a gain of 7.30 points.

3. Individual component contributions

Table 3 further measures the individual component contributions with Qwen2.5-VL-3B, sharing the connector depth, initialization, training data, and optimization schedule.

Component addedStarting configurationSPAR-Bench overall gain
Geometry-aware attentionRay embedding+10.60
Ray embeddingGeometry-aware attention+4.15
Correspondence supervisionBoth components+2.85

Correspondence supervision additionally improves the multi-view average by 2.13 points.

4. Manuscript revision

We will add the SFT-only results and expand the discussion of Table 3 to make the contributions of camera-ray embedding, geometry-aware attention, and correspondence supervision explicit.

Q2. Hyperparameter sensitivity

How is the sensitivity of the hyperparameter βcorr\beta_{\mathrm{corr}} for the performance?

Authors’ Response

Thank you for the question. We have completed a sensitivity sweep with βcorr∈{0.1,0.3,0.5,0.7,1.0}\beta_{\mathrm{corr}}\in\{0.1,0.3,0.5,0.7,1.0\}. All scores below use the same SPAR-Bench full evaluation protocol.

  • Sensitivity sweep results

The βcorr=0.1\beta_{\mathrm{corr}}=0.1 row uses the Qwen2.5-VL-3B results in manuscript Table 1. Δ\Delta denotes the difference from the βcorr=0.1\beta_{\mathrm{corr}}=0.1 reference, in percentage points.

(a) Overall and view-group scores

βcorr\beta_{\mathrm{corr}}OverallMulti-viewSingle-viewCross-viewΔ\Delta Overall
0.182.1882.3880.3986.370.00
0.382.1682.4480.2486.47-0.02
0.582.1082.6380.3685.17-0.08
0.782.0482.7180.1685.04-0.14
1.081.7782.0880.3184.75-0.41

Overall averages all 20 tasks. Multi-view averages the nine tasks marked MV, and single-view averages the eight single-view depth, distance, object-relation, and spatial-imagination tasks. Cross-view averages position matching, camera-motion inference, and viewpoint-change inference.

(b) Sensitivity pattern

  • Overall performance. The overall score remains close to 82.1 for βcorr\beta_{\mathrm{corr}} from 0.1 to 0.7. The largest score in this sweep is 82.18 at 0.1, followed by 82.16 at 0.3, 82.10 at 0.5, and 82.04 at 0.7. Increasing the weight to 1.0 reduces the score to 81.77.
  • Multi-view performance. The multi-view average increases from 82.38 at 0.1 to 82.71 at 0.7, then decreases to 82.08 at 1.0.
  • Cross-view performance. The cross-view average is highest at 0.3 with 86.47. It is 86.37 at 0.1, 85.17 at 0.5, 85.04 at 0.7, and 84.75 at 1.0.
  • Single-view performance. The single-view average changes from 80.39 at 0.1 to 80.24, 80.36, 80.16, and 80.31 as the weight increases to 0.3, 0.5, 0.7, and 1.0.

Dear Reviewer 6pUW,

Thank you again for your thoughtful review and constructive suggestions, which have been very helpful in improving our work.

We are writing to gently follow up on our rebuttal and would be grateful to hear whether our responses have addressed your main concerns.

We would be happy to provide any additional explanations or clarifications if needed.

Should you find our responses satisfactory, we would sincerely appreciate your consideration of updating your overall evaluation.

Thank you very much for your time and continued consideration.

Best regards,

Authors

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

This paper proposes GeoMLLM to improve cross-view spatial reasoning in MLLMs by explicitly modeling geometric correspondences between visual tokens across different viewpoints. The method incorporates camera-ray embeddings into visual tokens and introduces geometry-aware attention that uses pairwise ray and epipolar constraints to guide cross-view interactions. It also uses depth-derived correspondences to supervise attention during training. Experiments on SPAR-Bench and several other spatial reasoning benchmarks show improvements on multi-view tasks. The authors further conduct component ablations, camera perturbation experiments, and qualitative analyses to evaluate the effectiveness of the proposed method.

Critical strengths
  1. The paper is clearly written and the overall pipeline is straightforward. The motivation and methodology are clearly presented.
  2. The paper provides visualizations of cross-view attention and several case studies, which help illustrate how the proposed method works and complement the quantitative results.
Critical weaknesses
  1. Incorporating geometric constraints into attention and integrating camera information into MLLMs have been explored in prior work. For example, GTA (Miyato et al., 2024) and PRoPE (Li et al., 2025) introduce geometry-aware attention mechanisms, while CA-MLLM (Zhang et al., 2026), VG-LLM (Zheng et al., 2025), SpaceMind (Zhao et al., 2025) and CAMCUE (Zhang et al., 2026) incorporate camera or 3D geometric information into MLLMs. Although the paper compares against CA-MLLM and VG-LLM, these comparisons do not use fully matched training settings. There are no controlled comparisons with alternative geometry-aware attention mechanisms such as GTA or PRoPE. This makes it difficult to assess the additional contribution of the proposed design over existing approaches.
  2. Although the paper provides component-level ablations, the zero-component configuration in Table 3 uses an external VG-LLM baseline rather than a fully matched SFT-only model. Including an SFT-only baseline with the same backbone, connector architecture, training data, and optimization settings would help clarify the overall contribution of the proposed method.
Questions for authors
  1. Could the authors provide controlled comparisons with existing geometry-aware attention methods, such as GTA and PRoPE, under the same backbone and training settings?
  2. Could the authors provide a fully matched SFT-only baseline without the proposed geometry-aware components, while keeping the backbone, connector architecture, and training settings unchanged? This would help isolate the overall contribution of the three proposed modules from the effects of spatial fine-tuning and architectural differences.
3 comments
AuthorsOpenReview ↗

Response to Reviewer ttTm

Thank you for recognizing the clear presentation and qualitative analyses, and for your constructive suggestions.

Weaknesses 1 and Question 1:

Incorporating geometric constraints into attention and integrating camera information into MLLMs have been explored in prior work. For example, GTA (Miyato et al., 2024) and PRoPE (Li et al., 2025) introduce geometry-aware attention mechanisms, while CA-MLLM (Zhang et al., 2026), VG-LLM (Zheng et al., 2025), SpaceMind (Zhao et al., 2025) and CAMCUE (Zhang et al., 2026) incorporate camera or 3D geometric information into MLLMs. Although the paper compares against CA-MLLM and VG-LLM, these comparisons do not use fully matched training settings. There are no controlled comparisons with alternative geometry-aware attention mechanisms such as GTA or PRoPE. This makes it difficult to assess the additional contribution of the proposed design over existing approaches.

Could the authors provide controlled comparisons with existing geometry-aware attention methods, such as GTA and PRoPE, under the same backbone and training settings?

Authors’ Response:

Thank you for your suggestion. To address your concern about the additional contribution of GeoMLLM over existing geometry-aware attention methods, we provide direct comparisons with GTA and PRoPE.

1. Experimental setup

All methods use Qwen2.5-VL-3B and share the same training datasets, evaluation datasets, and experimental environment. We use the full GeoMLLM scores from manuscript Tables 1–2.

2. Quantitative results

(a) MMSI-Bench

MethodAttr.MSRMotionRelationOverall
GTA26.1624.2422.6431.8928.47
PRoPE24.6726.7716.6333.1228.01
GeoMLLM (full)36.9827.2729.4132.2932.16

(b) ViewSpatial-Bench

MethodCameraPersonOverall
GTA34.0446.9941.81
PRoPE36.0841.1439.12
GeoMLLM (full)33.4450.7243.81

(c) CV-Bench

Method2D3DOverall
GTA70.5181.8376.17
PRoPE74.6988.6781.68
GeoMLLM (full)73.6491.5082.57

(d) SPAR-Bench full

MethodOverallMulti-viewSingle-viewCross-view
GTA70.0172.1270.7661.70
PRoPE68.7070.9169.2960.48
GeoMLLM (full)82.1882.3880.3986.37

Cross-view averages position matching, camera-motion inference, and viewpoint-change inference.

Weaknesses 2 and Question 2:

Although the paper provides component-level ablations, the zero-component configuration in Table 3 uses an external VG-LLM baseline rather than a fully matched SFT-only model. Including an SFT-only baseline with the same backbone, connector architecture, training data, and optimization settings would help clarify the overall contribution of the proposed method.

Could the authors provide a fully matched SFT-only baseline without the proposed geometry-aware components, while keeping the backbone, connector architecture, and training settings unchanged? This would help isolate the overall contribution of the three proposed modules from the effects of spatial fine-tuning and architectural differences.

Authors’ Response:

Thank you for your comment. To assess GeoMLLM's contribution beyond spatial fine-tuning, we add SFT-only baselines for Qwen2.5-VL-3B and Qwen3-VL-4B. We also report the MHSA (multi-head self-attention) connector variant from manuscript Table 1.

1. Experimental setup

The SFT-only baselines use native Qwen architectures and the shared training mixture of LLaVA-Hound and SPAR examples. Both are trained for one epoch with a global batch size of 64 and a learning rate of 10−510^{-5}, with the vision encoder frozen. The full GeoMLLM and MHSA reference scores are taken from manuscript Tables 1–2.

2. Quantitative results

BackboneConfigurationOverallMulti-viewSingle-viewCross-view
Qwen2.5-VL-3BPure Qwen + SFT65.0767.6766.7852.70
Qwen2.5-VL-3BGeoMLLM (MHSA)74.8875.8074.5373.07
Qwen2.5-VL-3BGeoMLLM (full)82.1882.3880.3986.37
Qwen3-VL-4BPure Qwen + SFT69.4171.5771.2857.93
Qwen3-VL-4BGeoMLLM (full)83.7983.4982.5188.13

Dear Reviewer ttTm,

Thank you again for your thoughtful review and constructive suggestions, which have been very helpful in improving our work.

We are writing to gently follow up on our rebuttal and would be grateful to hear whether our responses have addressed your main concerns.

We would be happy to provide any additional explanations or clarifications if needed.

Should you find our responses satisfactory, we would sincerely appreciate your consideration of updating your overall evaluation.

Thank you very much for your time and continued consideration.

Best regards,

Authors

Dear ttTm,

With just over 24 hours remaining before the discussion deadline, we would greatly appreciate your feedback on our responses. Thank you for your time!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.