acceptodds
Under review as a conference paper at ICLR 2027

SCOPE: Learning Spherical Coordinate Grounding for Spatial Understanding in MLLMs

Abstract

Spatial understanding is pivotal for machines to advance toward physical world intelligence, yet current MLLMs still struggle to estimate object distances, judge relative directions, and reason about 3D layout. One line of work adds 3D encoders, which changes the architecture and weakens general multimodal ability. The other fine-tunes on spatial reasoning data, where supervision targets final answers rather than the 3D locations of relevant objects. Models may then rely on language priors without learning where objects are in 3D space. We therefore seek to train MLLMs to localize objects in 3D, which raises a question: which location representation best supports MLLMs' spatial reasoning? Under a protocol that varies only this choice across three MLLMs, we compare four representations common in 3D vision and find spherical coordinates the most effective across all three models. In spherical coordinates, distance and direction occupy separate axes: an observer's yaw rotation changes only azimuth, while moving directly toward an object changes only distance. To train MLLMs with spherical coordinate grounding, we develop an automatic pipeline that generates grounding samples from annotated 3D scans, yielding SphereQA, a large dataset of 300K samples in egocentric and allocentric reference frames. We then introduce SCOPE, an MLLM trained on SphereQA to predict spherical coordinates through per-axis classification with a spherical coordinate tokenizer instead of raw-text generation. Its two variants improve over their respective base MLLMs by 10.2/11.1 points on average across four spatial benchmarks, achieve the best average accuracy among the evaluated open-source and specialist MLLMs, and preserve general multimodal ability.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multimodal Large Language Model, Visual Reasoning, Spatial Intelligence

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_KeAwOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper focuses on enhancing the spatial understanding ability of MLLM. First, the paper explores the effects of four location representations on enhancing the spatial understanding ability of MLLM, and finds that using spherical coordinates is the most effective. In order to train MLLMs using spherical coordinate annotation, an automated process was developed to generate grounding samples from annotated 3D scans and construct 300K training data. In addition, by using a spherical coordinate tokenizer for per-axis classification instead of the original text generation method, spherical coordinates can be predicted. Based on constructed training data and vocabulary expansion, SCOPE achieves effective improvement in spatial position understanding dataset.

Critical strengths
  1. The paper explores an important issue in spatial understanding, which location representation is most suitable for enhancing the spatial understanding ability of MLLM. The effects of four location representations were compared through controlled experiments to determine the gain brought by spherical coordinates.
  2. The paper developed an automated pipeline for constructing spherical coordinate grounding data and generated 300K validated training samples. The experiments showed that these data can enhance the spatial understanding ability of MLLM.
  3. The paper proposes a spherical coordinate tokenizer that effectively predicts spherical coordinates using MLLM by dividing each axis into half-open bins and assigning one token per bin.
  4. SCOPE trained on the constructed training data and with vocabulary expansion achieves better performance on multiple spatial understanding datasets.
Critical weaknesses
  1. The spherical tokenizer’s quantization error and bin design are underexplored. Azimuth and elevation are binned every 5◦, and distance uses growing bin widths. Direct localization yields a mean error of 0.51 m and a median error of 0.36 m for SCOPE-8B, which may be sufficient for these benchmarks but limiting for high-precision spatial applications.
  2. How SCOPE uses spherical coordinate tokens at inference time is not clearly explained. During training, SCOPE performs per-axis classification over spherical tokens. However, most benchmark questions in Table 1 are multiple-choice or numerical questions that do not require explicit output.
  3. SCOPE is trained on the constructed dataset SPHEREQA and an auxiliary mixture of public spatial QA and general VQA. Figure 6a also analyses the impact of different parts of the data on model performance, but the paper lacks an explanation of how these different data are scheduled during training.
Questions for authors
  1. Can spherical coordinate grounding training bring performance improvement on other 3D scene understanding and spatial question answering benchmarks (such as VSI Bench)?
  2. How does the bin granularity of each spherical-coordinate axis, namely azimuth, elevation, and radial distance (for example, 5◦ angular bins and growing-width distance bins), affect 3D localization error and final performance on the four spatial benchmarks, and can the authors provide an ablation over different bin granularities and justify the chosen granularity?
3 comments
AuthorsOpenReview ↗

Responses to Reviewer KeAw

Thank you for your helpful review and for recognizing the importance of studying spatial representations and the value of our automated grounding data pipeline. We address each of your concerns below, and changes in the revised paper are marked in orange.

  1. [W1, Q2] Angular bin granularity and localization precision

    We agree that bin granularity matters for localization precision. We halved the angular bin width from 5° to 2.5°, keeping all other settings unchanged. As shown below, finer bins slightly improve localization but yield a lower average spatial QA score (45.8 vs. 46.5). Halving the bin width doubles the angular vocabulary, reducing the average number of training examples per token. Learning these finer distinctions may therefore require more training data. Since each configuration requires retraining on the full 600K-sample corpus, time and compute constraints during the rebuttal period limited us to this additional setting. We will explore more angular and radial settings in future work.

    Angular binsMindCubeMMSIViewSpatialReVSIAvg.Loc. mean (m) ↓Loc. median (m) ↓<0.5 m (%) ↑<1 m (%) ↑
    5°49.336.549.850.446.50.510.3665.8689.58
    2.5°49.536.148.449.245.80.490.3468.4490.42
  2. [W2] Coordinate tokens at inference

    SCOPE does not generate coordinate tokens on the Table 1 benchmarks. We use each benchmark’s default prompts in lmms-eval without the <SPHERE_MODE> prefix, so SCOPE directly outputs an option letter or a number. We have clarified this in the Inference Setup paragraph (Sec. 5.1). Coordinate tokens benefit spatial reasoning through training, even though they are not generated at inference. Since grounding and spatial QA share model parameters, learning to locate objects in 3D helps the model reason about their spatial relationships. Two ablations support this:

    • Grounding-only training. Training on SphereQA alone, without spatial QA data, improves the average score by 4.3 points (Fig. 6(a)), demonstrating that object-localization supervision transfers to spatial QA.

    • Tokenized supervision. With the same training data and spherical coordinate targets, tokenized supervision outperforms raw-text supervision by 2.0 points (Table 3). This shows that coordinate tokens make grounding supervision more effective for downstream spatial reasoning.

  3. [W3] Joint training schedule

    Sorry for the confusion. We mix SphereQA with the auxiliary data and train them jointly for one epoch in a single stage. We have revised the Model Training paragraph in Sec. 5.1 to clarify this schedule.

  4. [Q1] Evaluation on additional spatial benchmarks

    Thank you for the suggestion. Our main evaluation includes ReVSI [1], a rebuilt version of VSI-Bench with corrected 3D annotations and human-verified questions. Following your suggestion, we have additionally evaluated SCOPE and its base models on VSI-Bench, SPAR-Bench-Tiny, and CV-Bench using the same evaluation settings as Table 1. SCOPE improves both backbones on all three benchmarks, further supporting the effectiveness of our method across spatial reasoning benchmarks.

    ModelVSI-BenchSPAR-Bench-TinyCV-Bench
    Qwen2.5-VL-7B35.136.178.0
    SCOPE-7B50.345.780.6
    Qwen3-VL-8B55.839.886.3
    SCOPE-8B58.253.788.0

[1] ReVSI: Rebuilding Visual Spatial Intelligence Evaluation for Accurate Assessment of VLM 3D Reasoning. ICML 2026.

Reviewer_KeAwOpenReview ↗

Responses to rebuttal

Rebuttal has solved my problem, and I will maintain my positive score.

AuthorsOpenReview ↗

Follow-up Response to Reviewer KeAw

Thank you for following up. We're glad that our response has addressed your concerns, and we appreciate your positive assessment of our work.

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

This paper proposes SCOPE that enhances spatial reasoning in MLLMs through spherical coordinate grounding and a dedicated coordinate tokenizer. The paper is well-motivated, and the experiments provide evidence for the effectiveness of spherical coordinates over other geometric representations.
However, there still remain several concerns, such as the fairness of comparisons with existing methods, the significance of performance improvements, the justification of key design choices, and the comprehensiveness of the evaluation. These limitations weaken the paper's empirical claims.

Critical strengths
  1. Geometric representation. The paper investigates spatial reasoning from the perspective of geometric representations. The separability of spherical coordinates provides an intuitive motivation, and the controlled experiments demonstrate their advantages over alternative coordinate representations.
  2. Dataset construction and ablation studies. The authors construct SphereQA, a dataset containing 300K object-level grounding samples. The ablation studies provide useful insights into the contributions of coordinate representations, tokenization, and semantic embedding initialization.
  3. Practical architecture design. SCOPE improves spatial reasoning without requiring additional 3D geometric inputs at inference time or substantial modifications to the MLLM backbone.

Overall, I've learned some useful points from the paper. It solves some problems in my own work done last year.

Critical weaknesses

W1. Unfair comparisons due to differences in training data and supervision. SphereQA leverages annotated 3D scans, including object geometry, camera poses, and object locations, to provide explicit object-level 3D grounding supervision. In contrast, the compared methods differ substantially in their training datasets, supervision granularity, and task distributions. The internal ablations demonstrate the benefits of spherical coordinates over Cartesian coordinates, this is good. However, the cross-method comparisons do not sufficiently isolate the contribution of the proposed approach. More controlled comparisons under consistent training data, backbones are needed.

W2. Limited performance improvements over existing methods. Compare with SenseNova-SI-1.1-7B and VST-SFT-7B, the improvement of SCOPE-7B is marginal. The gap in the capabilities of the foundation model has a significant impact on the final outcome. Considering that all these methods are based on Qwen2.5-VL-7B, the current results are not sufficient to prove that SCOPE has a significant overall performance advantage. Furthermore, it is suggested to add a comparison between SCOPE-8B and SenseNova-SI-1.1-Qwen3-VL-8B, which is also based on Qwen3-VL-8B.

W3. Incomplete benchmark and baseline coverage. The evaluation does not include relevant baselines such as Spatial-MLLM [1] and does not include VSI-Bench, a widely adopted benchmark for spatial reasoning. Including VSI-Bench would improve comparability with prior work.

W4. Insufficient justification of data efficiency claims. SphereQA relies on high-quality 3D scene annotations, including meshes, camera poses, and object-level geometry. Comparing data efficiency only by the number of QA pairs overlooks differences in annotation costs and supervision richness. The claimed data efficiency advantage should therefore be interpreted more cautiously.

[1] Wu, Diankun, et al. "Spatial-mllm: Boosting mllm capabilities in visual-based spatial intelligence." Advances in neural information processing systems 38 (2025): 13569-13597.
[2] Yang, Jihan, et al. "Thinking in space: How multimodal large language models see, remember, and recall spaces." 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025.

Questions for authors

Besides weaknesses, here are some other questions I want to raise:

  1. Can your proposed model perform well on 3DVG tasks? Seeing that object-level spherical grounding is a central contribution, the direct grounding evaluation in Appendix A.4 only measures 3D object-center localization. It does not evaluate complete 3D bounding box prediction, which is a common task. This evaluation is considerably narrower than standard 3D visual grounding and does not fully establish object-level grounding capability. Here you can refer to these works [1][2].

  2. There is also a small suggestion. We recommend adding controlled experiments with different bin granularities, Token-only localization accuracy, and a continuous Regression Head, so as to further verify the effectiveness and necessity of the proposed Tokenizer.

[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.
[2] Xu, Beining, et al. "S 2^ 2-MLLM: Boosting Spatial Reasoning Capability of MLLMs for 3D Visual Grounding with Structural Guidance." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026.

I may reconsider my score based on the rebuttal. The authors may add extra experimental experiments. If not enough time, you can just add important ones.

4 comments
AuthorsOpenReview ↗

Responses to Reviewer ibNM (1/2)

Thank you for your constructive feedback and recognition of our work. We address each of your concerns below, and changes in the revised paper are marked in orange.

  1. [W1] Controlled comparisons of training data and supervision

    Thank you for acknowledging our representation ablations. We clarify that spherical coordinate supervision is our core contribution, not an unfair comparison. It offers an underexplored supervision perspective: localizing objects in spherical coordinates to support spatial reasoning (Sec. 3). To further address your concern, we clarify our training data and present controlled comparisons below.

    • Training data across methods. SphereQA is generated automatically from public 3D scans without additional manual labeling. SCOPE uses 0.6M training samples, compared with the reported spatial training corpora of 4.1M for VST-SFT-7B and 8M for SenseNova-SI-1.1-7B (Fig. 6(c)).

    • Controlled comparison. We use the same backbone under the same sample budget. Replacing half of the existing public training data with SphereQA improves the average score at both 300K and 600K, supporting the benefit of incorporating spherical grounding supervision. We additionally compare models sharing the Qwen3-VL-8B backbone in [W2].

      Training data (Qwen3-VL-8B)300K600K
      Existing public data only42.644.3
      SphereQA + existing public data (SCOPE)43.946.5
  2. [W2] Comparisons with Qwen3-VL-8B-based spatial MLLMs

    On Qwen2.5-VL-7B, SCOPE achieves an average score of 41.5, compared with 40.9 for SenseNova-SI-1.1-7B and 39.6 for VST-SFT-7B, using only 0.6M training samples, 13.3× and 6.8× fewer than their 8M and 4.1M. Notably, SenseNova-SI-1.1-7B is trained on MindCube-Train, whereas our MindCube evaluation is zero-shot.

    Following your suggestion, we have added SenseNova-SI-1.1-8B and MultiHopSpatial-8B to Table 1. Both share the Qwen3-VL-8B backbone with SCOPE-8B. SCOPE-8B achieves the best zero-shot results across all four benchmarks among these models. * marks in-domain MindCube results.

    ModelMindCubeMMSIViewSpatialReVSI
    SenseNova-SI-1.1-8B71.8*35.149.544.1
    MultiHopSpatial-8B33.229.240.044.6
    SCOPE-8B49.336.549.850.4
  3. [W3] Spatial-MLLM and VSI-Bench evaluation

    We address the baseline and the benchmark in turn.

    • Spatial-MLLM. We have added Spatial-MLLM to Table 1, and SCOPE-8B outperforms it on all four benchmarks:

      ModelMindCubeMMSIViewSpatialReVSI
      Spatial-MLLM-4B39.024.335.541.9
      SCOPE-8B49.336.549.850.4
    • VSI-Bench. Our main evaluation already covers VSI-Bench through ReVSI [1], a rebuilt version of VSI-Bench that addresses annotation errors and question answerability from sampled frames. Following your suggestion, we additionally evaluate both backbones on the original VSI-Bench using the same evaluation settings as Table 1. SCOPE improves both backbones, with gains of 15.2 and 2.4 points for the 7B and 8B models, respectively:

      ModelVSI-Bench
      Qwen2.5-VL-7B35.1
      SCOPE-7B50.3 (+15.2)
      Qwen3-VL-8B55.8
      SCOPE-8B58.2 (+2.4)

[1] ReVSI: Rebuilding Visual Spatial Intelligence Evaluation for Accurate Assessment of VLM 3D Reasoning. ICML 2026.

AuthorsOpenReview ↗

Responses to Reviewer ibNM (2/2)

  1. [W4] Scope of the data-efficiency claim

    Thank you for pointing this out. We have revised the Data Efficiency paragraph in Sec. 5.3 to clarify that our claim concerns the number of SFT samples, rather than annotation cost or supervision richness.

  2. [Q1] Direct 3D bounding box prediction

    We evaluate direct 3D bounding box prediction on the same SUN RGB-D queries as App. A.4, reporting accuracy at 3D IoU thresholds of 0.25 and 0.5. SCOPE improves [email protected] by 11.9 points over its backbone. [email protected] remains low for both models, indicating limited precision in full-box prediction. SphereQA currently supervises object centers rather than box sizes, and adding size supervision is a direction for future work. Additionally, we have also added Qwen-3D and S²-MLLM to the discussion of 3D visual grounding methods in Sec. 2.

    Model[email protected][email protected]
    Qwen3-VL-8B27.380.86
    SCOPE-8B39.313.13
  3. [Q2] Bin granularity, token-only localization, and regression

    Thank you for these helpful suggestions. We address each point below.

    • Bin granularity. We halved the angular bin width from 5° to 2.5°, keeping all other settings unchanged. As shown below, finer bins slightly improve localization but yield a lower average spatial QA score (45.8 vs. 46.5). Halving the bin width doubles the angular vocabulary, reducing the average number of training examples per token. Learning these finer distinctions may therefore require more training data. Each configuration requires retraining on the full training set, so rebuttal time and compute constraints limited us to this additional setting. We will explore more angular and radial settings in future work.

      Angular binsMindCubeMMSIViewSpatialReVSIAvg.Loc. mean (m) ↓Loc. median (m) ↓<0.5 m (%) ↑<1 m (%) ↑
      5°49.336.549.850.446.50.510.3665.8689.58
      2.5°49.536.148.449.245.80.490.3468.4490.42
    • Token-only localization. We prompt SCOPE to output only spherical coordinate tokens. We then map each token to its bin center and compute the 3D object-center error on the same SUN RGB-D queries as App. A.4. Token-only outputs achieve 0.63 m mean error, compared with 1.32 m for the base model and 0.51 m for SCOPE’s numerical outputs.

      ModelOutputMean (m) ↓Median (m) ↓<0.5 m (%) ↑<1 m (%) ↑
      Qwen3-VL-8BNumerical1.321.193.3432.33
      SCOPE-8BNumerical0.510.3665.8689.58
      SCOPE-8BToken-only0.630.4758.4885.63
    • Continuous regression head. We agree this is an interesting alternative. However, it requires modifying the architecture and retraining the model, which is difficult to complete within the limited rebuttal period. We will explore it in future work.

Reviewer_ibNMOpenReview ↗

response to rebuttal

Thank you for your rebuttal. I will keep my positive score!

AuthorsOpenReview ↗

Follow-up Response to Reviewer ibNM

Thank you for your thoughtful review and for taking the time to consider our response. We appreciate your positive assessment.

Reviewer_8NaLOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes SCOPE, a method for improving MLLM spatial reasoning through spherical coordinate grounding. The work consists of three main parts: A controlled prompting study compares four 3D grounding representations and finds that spherical coordinates achieve the best results. The authors construct SphereQA, a dataset of 300K object-level grounding questions generated from annotated 3D scans, covering single-image, multi-image, and video inputs. SCOPE introduces a spherical coordinate tokenizer that discretizes distance and angles into special tokens for coordinate prediction. Experiments on four spatial benchmarks show improvements of 10.2 and 11.1 points over two base MLLMs, while largely preserving general multimodal capabilities.

Critical strengths
  1. Simple and well-motivated idea. Spherical coordinates naturally separate distance and direction. The proposed approach provides explicit 3D grounding supervision without requiring major architectural changes.
  2. Useful investigation of grounding representations. The controlled prompting study highlights how different coordinate representations affect MLLM spatial reasoning, providing an interesting motivation for the proposed method.
  3. Scalable dataset construction. SphereQA provides large-scale object-level 3D grounding supervision from existing annotated scenes. It covers both egocentric and allocentric settings with different visual input formats.
  4. Strong empirical results. SCOPE achieves consistent overall improvements across four spatial benchmarks. The direct object localization results in Appendix A.4 also provide evidence that the model learns more accurate 3D grounding.
  5. Comprehensive experiments. The paper includes ablations on coordinate representations, tokenization, training data composition, and scaling. It also evaluates the preservation of general multimodal capabilities.
Critical weaknesses
  1. Limited evidence for the general advantage of spherical coordinates. The controlled representation study in Section 3 is conducted only on MindCube-Tiny, where many questions involve distance, direction, and viewpoint transformations. Spherical coordinates directly expose the quantities needed for some of these tasks. For example, left/right or near/far comparisons can often be answered by directly comparing azimuth or distance, while Cartesian coordinates require additional computation. Therefore, the advantage of spherical coordinates on such questions is somewhat expected. The paper does not provide sufficient evidence that this advantage extends to a broader range of spatial reasoning tasks. Evaluating more diverse tasks and reporting per-task results would make the findings more convincing.
  2. The contributions of the proposed supervision and training data are not sufficiently disentangled. SCOPE primarily improves spatial reasoning through an auxiliary object-level coordinate grounding task, rather than introducing an explicit geometric reasoning mechanism. Although the ablations show some benefits from SphereQA, it remains unclear how much of the improvement comes from the grounding supervision itself, increased data diversity, or the choice of coordinate representation. In particular, the final model is trained on a mixture of SphereQA and existing spatial/general QA datasets. In Fig. 6(a), the 600K Auxiliary-only setting already achieves 44.3%, compared with 46.5% for the full method. Moreover, the paper does not clarify how the 600K Auxiliary-only dataset is constructed. If it repeats the original 300K auxiliary samples, the comparison may also reflect differences in the number of unique training examples rather than the benefit of spherical grounding supervision alone.
  3. Missing comparisons with strong recent spatial MLLMs. Table 1 omits several competitive spatial MLLMs with published results on the same benchmarks. For example, 3DThinker (CVPR 2026) reports 76.0 on MindCube-Tiny, 43.3 on MMSI-Bench, and 68.6 on ViewSpatial with a Qwen2.5-VL-7B backbone, substantially outperforming SCOPE on these benchmarks. Stronger SenseNova-SI variants (e.g., SenseNova-SI-1.3) also report significantly higher scores. Although evaluation settings may differ, the current baseline selection does not adequately establish SCOPE's competitiveness against recent spatial specialists.
  4. The transfer from coordinate grounding to general spatial reasoning is not sufficiently explained. SphereQA mainly supervises static object locations, while the evaluated benchmarks cover broader abilities such as object counting, motion understanding, and route planning. Moreover, the newly introduced spherical tokens are not explicitly generated during ordinary spatial QA inference. While such auxiliary supervision may improve the model's shared representations, the paper provides limited analysis of how this learned grounding ability transfers to general spatial reasoning. The subtask results in Appendix A.3 also show that improvements are particularly strong in camera-centered positioning and distance-related tasks, while gains are smaller or less consistent in other abilities, such as object orientation and motion reasoning. A more detailed investigation of how spherical grounding supervision benefits these broader tasks would substantially strengthen the paper's contribution.
Questions for authors
  1. Could the authors provide per-task results for the controlled representation study in Section 3? Does the advantage of spherical coordinates extend beyond tasks that directly involve distance or direction comparisons?
  2. Could the authors include a tokenized Cartesian coordinate baseline in Table 3 to better isolate the benefit of spherical coordinates from that of coordinate tokenization?
  3. How is the 600K Auxiliary-only dataset in Fig. 6(a) constructed? Does it contain 600K unique samples, or does it repeat the original 300K auxiliary samples?
  4. Could the authors include comparisons with stronger recent spatial MLLMs in the main table?
  5. Could the authors further explain why and how spherical coordinate grounding supervision improves general spatial reasoning tasks where the special coordinate tokens are not explicitly used during inference? In particular, why are the improvements less consistent across different spatial abilities?
  6. Could the authors provide stronger evidence that the gains come from learning better spatial representations, rather than simply from additional or higher-quality training data?
6 comments
AuthorsOpenReview ↗

Responses to Reviewer 8NaL (1/3)

Thank you for your review and for recognizing our well-motivated idea and comprehensive experiments. We address each of your concerns below, and changes in the revised paper are marked in orange.

  1. [Q1, W1] Per-task evidence for spherical representations

    Thank you for recognizing our motivation that spherical coordinates, which naturally separate distance and direction, facilitate spatial reasoning. We clarify that MindCube-Tiny evaluates layout construction and cross-view reasoning from partial observations, rather than only direct comparisons of distances or directions. Following your suggestion, we report per-task results on MindCube-Tiny and further evaluate two subtasks on MMSI-Bench, as follows.

    • Per-task results on MindCube-Tiny. As shown below, spherical coordinates perform best on Among and Rotation for all three models. On Around, geometry prompts generally do not help over image only. We have added this table to the appendix (Table 6).

      ModelInputAroundAmongRotationOverall
      Qwen3-VL-32B-InstructImage Only36.436.726.034.6
      Depth36.037.027.034.9
      3D Box35.637.326.034.8
      Cartesian34.438.526.535.2
      Spherical36.439.831.037.3
      InternVL3.5-14BImage Only46.841.838.042.3
      Depth44.447.041.045.2
      3D Box43.647.742.045.6
      Cartesian39.647.841.044.6
      Spherical42.848.743.046.2
      GPT-5.4-miniImage Only49.238.329.039.1
      Depth48.441.230.540.9
      3D Box46.842.727.540.8
      Cartesian40.442.730.539.8
      Spherical47.643.231.542.0
    • Additional per-task results on MMSI-Bench. To further assess the effectiveness of spherical representations across spatial reasoning tasks, we evaluate two additional subtasks from MMSI-Bench: object–object relations (Obj–Obj) and multi-step reasoning (MSR). As shown below, spherical coordinates outperform Cartesian coordinates on both subtasks for all three models and achieve the best or tied-best result in five of six cases. These results indicate the effectiveness of spherical representations across the evaluated subtasks.

      ModelTaskImage OnlyDepth3D BoxCartesianSpherical
      Qwen3-VL-32B-InstructObj–Obj30.933.031.927.733.0
      MSR28.330.330.328.831.3
      InternVL3.5-14BObj–Obj22.324.523.421.326.6
      MSR27.833.830.329.833.3
      GPT-5.4-miniObj–Obj24.526.623.428.729.8
      MSR24.725.325.827.328.3
  2. [Q2] Tokenized Cartesian baseline

    We train a tokenized Cartesian baseline with the same training setup and semantic initialization as our spherical tokens, binning each Cartesian axis like the radial axis and mirroring it around zero. Spherical tokens still outperform Cartesian tokens on all four benchmarks, so the advantage of spherical coordinates holds with tokenization fixed. We have added this baseline to Table 3.

    Tokenized representationMindCubeMMSIViewSpatialReVSIAvg.
    Cartesian47.834.948.548.344.9
    Spherical49.336.549.850.446.5
  3. [Q3] Unique samples in the auxiliary-only comparison

    Sorry for the confusion. The 600K auxiliary-only set contains 600K unique samples. We add 300K non-overlapping samples from the same four sources in the original proportions. This is now clarified in Sec. 5.3.

AuthorsOpenReview ↗

Responses to Reviewer 8NaL (2/3)

  1. [Q4, W3] Comparisons with recent spatial MLLMs

    We clarify that directly comparing the reported results of 3DThinker and SenseNova-SI-1.3 with SCOPE constitutes an unfair comparison, as their settings differ from ours in three key ways.

    • 3DThinker and SenseNova-SI are trained on MindCube data. Our training data contain neither MindCube-Train nor scenes used in evaluation (App. A.1). Our MindCube results are therefore zero-shot, making a direct comparison with their in-domain results an unfair comparison.

    • 3DThinker relies on an extra 3D model. It adds an extra 3D foundation model (VGGT) and a projector during training to align its latent reasoning tokens with 3D features. SCOPE keeps the MLLM architecture unchanged.

    • SenseNova-SI-1.3 uses 14M training samples. This is about 23× the 0.6M samples used by SCOPE, placing it in a different data regime.

    Following your suggestion, we have added four recent spatial MLLMs to Table 1: 3DThinker-S1-3B, Spatial-MLLM-4B, SenseNova-SI-1.1-8B, and MultiHopSpatial-8B. For 3DThinker, we use the released 3B checkpoint because its 7B weights are unavailable. * marks in-domain MindCube scores and the corresponding averages. SCOPE-8B achieves the best zero-shot results on every benchmark among these models.

    ModelMindCubeMMSIViewSpatialReVSIAvg.
    3DThinker-S1-3B65.9*22.530.012.432.7*
    SenseNova-SI-1.1-8B71.8*35.149.544.150.1*
    Spatial-MLLM-4B39.024.335.541.935.2
    MultiHopSpatial-8B33.229.240.044.636.8
    SCOPE-8B49.336.549.850.446.5
  2. [Q5, W4] Transfer from grounding to spatial reasoning

    Thank you for these insightful questions. We address the two parts in turn.

    • Why grounding helps without coordinate tokens at inference. Coordinate tokens provide training supervision but are not generated during spatial QA. Since grounding and QA share model parameters, learning to locate objects in 3D helps the model reason about their spatial relationships. Two ablation results support this:

      • Grounding-only training: Training on SphereQA alone, without spatial QA data, improves the average score by 4.3 points (35.4 → 39.7, Fig. 6(a)).
      • Tokenized supervision: Spherical tokens outperform raw-text coordinates by 2.0 points (Table 3), although neither setting generates coordinate tokens during spatial QA.
    • Why the gains vary across abilities. This variation is expected, as different spatial abilities depend to varying degrees on the object locations supervised by SphereQA.

      • Position and distance. We observe substantial gains in camera-centered positioning on MMSI (up to +30.3) and distance estimation on ReVSI, both closely related to the supervised object locations.
      • Relations and counting. Gains also extend to Obj–Obj relations on MMSI (+11.7/+8.5) and object counting on ReVSI (+26.7/+5.4), indicating benefits beyond predicting individual object locations.
      • Orientation and motion. Gains are smaller or less consistent on these tasks, which require information about object orientation or temporal changes that SphereQA does not explicitly supervise.
AuthorsOpenReview ↗

Responses to Reviewer 8NaL (3/3)

  1. [Q6, W2] Separating representation and training-data contributions

    Our paper already examines both factors through the representation ablations in Table 3 and the data-composition study in Fig. 6(a). We summarize these results below, together with the newly added tokenized Cartesian baseline (see [Q2]).

    • Spatial representation. Table 3 uses the same training data with different coordinate representations. Spherical tokens outperform Cartesian and spherical raw text by 4.1 and 2.0 points on average, respectively. They also outperform tokenized Cartesian coordinates on all four benchmarks (see [Q2]). These comparisons isolate the benefit of the coordinate representation from changes in training data.

    • Training data. Fig. 6(a) compares data compositions under the same sample budget. Training on SphereQA alone improves the average score by 4.3 points without spatial QA data. Replacing half of the auxiliary QA with SphereQA improves the average score by 1.3 points at 300K and 2.2 points at 600K (44.3 to 46.5), supporting the benefit of spherical grounding supervision at a fixed sample budget.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.