acceptodds
Under review as a conference paper at ICLR 2027

Thinking with Latent Geometry Tokens for Multi-Image Spatial Reasoning

16 views

Abstract

Multi-image spatial reasoning remains challenging for multimodal large language models (MLLMs). A key bottleneck lies in integrating partial observations into coherent internal spatial representations, a capability known as spatial mental modeling. Existing approaches often rely on dense spatial annotations or external tools, limiting scalability or increasing inference overhead. To reduce these dependencies, recent work explores modeling spatial representations in latent space. However, effectively encoding geometric information in latent space and incorporating it during reasoning is underexplored. We propose Thinking with Latent Geometry Tokens, a two-stage framework that grounds latent states in compact geometric representations and encourages flexible latent-mode entry. Specifically, we use camera and scene tokens from a pretrained 3D foundation model to supervise latent states, providing viewpoint-conditioned and cross-view geometry information while jointly training on text generation. We then augment Group Relative Policy Optimization (GRPO) with latent-mode entry point exploration, using successful exploratory trajectories to guide policy updates. Experiments on three multi-image benchmarks demonstrate our consistent gains over spatial reasoning baselines, with further improvements in single-image settings.

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

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_zmSDOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies spatial mental modeling for multi-image spatial reasoning. Instead of invoking external 3D tools at inference or relying on dense spatial annotations, it aligns latent states with compact camera and scene tokens extracted from a pretrained VGGT-Ω model. Stage 1 combines MSE-based geometric alignment with next-token prediction. Stage 2 augments GRPO with latent-mode entry exploration. The default backbone is Qwen2.5-VL-3B. The training set contains 15K examples sampled from MindCube-Train and SPAR-7M, including 10K for SFT and 5K for RL, with reasoning traces generated by Qwen3.5-122B. Across three multi-image benchmarks, the method improves the overall score from 29.4 to 55.6 over Qwen2.5-VL-3B and exceeds HATCH-3B by 12.0 points. It also improves performance on the single-image SPAR-Bench-SI and CV-Bench.

Critical strengths
  1. The controlled comparison with different methods provides the clearest evidence for the method.
  2. The ablation of latent-entry exploration controls for the additional rollout budget. When failed prompts receive the same number of additional standard rollouts without entry-position bias, performance on MindCube-Tiny is 3.8 points lower. Removing either camera or scene tokens reduces performance by 2.8 and 4.4 points, respectively, while removing the entire latent segment causes a 14.9-point drop. These ablations are well aligned with the proposed components.
Critical weaknesses
  1. The overall framework is highly similar to 3DThinker. Both methods align latent states with geometry extracted by a pretrained 3D foundation model, combine geometry alignment with next-token prediction during SFT, and subsequently apply GRPO. The novelty and conceptual distinction from 3DThinker should therefore be articulated more carefully.
  2. The primary Qwen backbone is relatively dated. Although the paper evaluates different model families and scales, the main experiments are centered on Qwen2.5-VL. Validation on newer Qwen3.6/3.8, would better establish whether the method remains beneficial with stronger spatial representations and reasoning capabilities. This is consequential because improvements obtained on an older backbone may shrink as the underlying model becomes more capable.

I would like to raise my score if the author can solve my concerns.

Questions for authors
  1. Please report the overlap between the 15K training examples and MindCube-Tiny, MMSI-Bench, and SPAR-Bench at the image, scene, and question-template levels. If SPAR-Bench overlaps with SPAR-7M, or if MindCube-Tiny shares scenes with the MindCube training data, the main table cannot be interpreted as evidence of out-of-domain generalization.
  2. If HATCH and 3DThinker are trained using the same Qwen3.5-122B-generated reasoning traces as your method, while removing only the proposed geometric latent components, how much of the gain remains on MindCube-Tiny and MMSI-Bench? This comparison would help separate the contribution of the method from that of the teacher-generated data.
8 comments
AuthorsOpenReview ↗

Rebuttal by Authors [1/4]

Dear Reviewer zmSD,

Thank you for your constructive suggestions. We are glad that you found the controlled comparisons and ablations to provide clear evidence for our method. We sincerely hope our responses below could address your concerns.

W1. Novelty and conceptual distinction from 3DThinker.

Thank you for your valuable suggestion. We agree that articulating the distinction from 3DThinker more carefully would strengthen the paper. We have revised the manuscript in Sections 2.3 and 4.2 as you suggested. With the shared training framework (SFT+RL), our method differs from 3DThinker in the following respects:

  • Alignment mechanism: 3DThinker maps latent states and image embeddings jointly to dense VGGT features, whereas our method maps the latent states alone to compact geometry tokens (Eq. 3, L206-213). We adopt this design to encourage the geometric target to be recovered from the latent state itself, rather than also relying on the image embeddings.
  • Latent segment: 3DThinker aligns a fixed number of latent states as a whole, whereas we assign one camera latent and one scene latent to each input view (L193-202). Each latent state thus receives its own supervision signal and has a defined role, with both roles shown to contribute in our ablation (Figure 6).
  • Latent-mode entry: 3DThinker enters latent mode at a fixed point before text generation and uses standard rollouts for GRPO. In contrast, we augment GRPO with latent-mode entry exploration (L261-269) to encourage flexible latent-mode entry.

We compare 3DThinker and our method at the SFT stage directly in Appendix C.1. With the same geometric teacher (VGGT-Ω), backbone, and training data, ours outperforms 3DThinker by 7.4 points on MindCube-Tiny (Figure 7). In addition, when 3DThinker is retrained on our 15K samples, our method outperforms it on all benchmarks at both stages (please see our response to Q2).

AuthorsOpenReview ↗

Rebuttal by Authors [2/4]

W2. Validation on newer Qwen backbone.

Thank you for pointing this out. We agree that demonstrating applicability on newer backbones is important. Following your suggestion, we have conducted new experiments on Qwen3.8-27B across both multi-image and single-image spatial reasoning benchmarks. As shown in Table E, our method consistently improves over the base model on these benchmarks, indicating that it remains beneficial with a newer backbone. We have added these results to Appendix C.2 of the revised manuscript.

Table E: Results with the Qwen3.8-27B backbone on MindCube-Tiny, MMSI-Bench, SPAR-Bench-MV, SPAR-Bench-SI, and CV-Bench.

ModelsMindCube-TinyMMSI-BenchSPAR-Bench-MVSPAR-Bench-SICV-Bench
Qwen3.8-27B (Base)46.031.845.752.887.5
Ours (SFT)84.536.655.857.589.0
Ours (SFT+RL)88.440.563.261.090.3
AuthorsOpenReview ↗

Rebuttal by Authors [3/4]

Q1. Clarification on overlap between training samples and benchmarks.

Thank you for raising this question. Our 15K training samples consist of 5K from the SPAR-7M training split and the full MindCube-Train set (10K). When sampling from SPAR-7M, we selected only multi-image samples and excluded all scenes that appear in SPAR-Bench or in the ScanNet and Matterport3D portions of MMSI-Bench. The overlap at the image, scene, and question-template levels is summarized in Table F, and we have added the details to Appendix A.3 of the revised manuscript.

Table F: Overlap between our training samples and each benchmark.

BenchmarkImageSceneQuestion templateSetting
MindCube-TinyNoYesYesIn-domain
SPAR-BenchNoNoYesIn-domain
MMSI-BenchNoNoNoOut-of-domain
CV-BenchNoNoNoOut-of-domain
  • In-domain benchmarks. SPAR-Bench shares no images or scenes with our SPAR-7M samples, but shares question templates. MindCube-Tiny shares no images with MindCube-Train, but shares scenes and question templates. We regard both as in-domain benchmarks.
  • Out-of-domain benchmarks. MMSI-Bench and CV-Bench share no images, scenes, or question templates with our training samples. MMSI-Bench aggregates images from eight sources (ScanNet, nuScenes, Matterport3D, Ego4D, AgiBot-World, DTU, DAVIS 2017, and Waymo). It requires the model to reason about domains absent from our training data, such as autonomous driving and robotic manipulation. CV-Bench is built from ADE20K, COCO, and three Omni3D subsets (Hypersim, SUN RGB-D, and nuScenes), covering diverse indoor/outdoor environments from datasets entirely distinct from our training sources.
Reviewer_ydEJOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a two-stage training framework for multi-image and single-image spatial reasoning. It uses camera and scene geometry priors from VGGT-Ω to improve spatial reasoning performance. The method achieves clear improvements across several benchmarks and demonstrates some generalization ability.

Critical strengths
  • The two-stage training framework is intuitive, and the design of the second stage offers useful insights.
  • The method achieves substantial improvements on several benchmarks and shows some generalization ability. The ablation studies provide evidence supporting the effectiveness of several components of the training recipe.
Critical weaknesses
  • Most experiments use Qwen2.5-VL-3B, while the evaluation across different backbones is restricted to MindCube-Tiny. The limited evaluation across model families, scales, and benchmarks makes it difficult to assess the method’s general applicability.
  • The improvement over HATCH on SPAR-Bench-MV is marginal and does not sufficiently support the strength of the paper’s claims.
  • The ablation studies are conducted only on MindCube-Tiny and do not report how individual components affect performance on out-of-distribution benchmarks. Their contributions to generalization therefore remain unclear.
  • The discussion of related work is incomplete. e.g.
    • Latent-GRPO: Group Relative Policy Optimization for Latent Reasoning (Deng et al., 2026).
    • Adaptive Latent Agentic Reasoning (Jung et al., 2026).
    • Actial: Activate Spatial Reasoning Ability of Multimodal Large Language Models (Zhan et al., 2025).
    • GeoAnchor: Collaborative Reasoning via Latent Decomposition for 3D Spatial Understanding (Li et al., 2026).
    • Beyond Visual Memory: Mechanistic Diagnostics of Latent Visual Reasoning (Guo et al., 2026).
Questions for authors

See the weakness.

6 comments
AuthorsOpenReview ↗

Rebuttal by Authors [1/3]

Dear Reviewer ydEJ,

We thank you for your thoughtful comments and recognition of our intuitive two-stage training framework, the insightful design of the second stage, and the substantial empirical improvements. We sincerely hope that the following responses address your concerns.

W1. Evaluation across backbones and benchmarks.

Thank you for your valuable feedback. To further assess the general applicability of our method, we have conducted new experiments on all benchmarks from Tables 1 and 2, with backbones covering different model families and scales: Qwen3-VL-8B [1], InternVL3.5-8B [2], and Qwen3.8-27B [3]. All models are trained on our curated 15K dataset. Results in Table C show that our method consistently improves over the corresponding base models at both the SFT and SFT+RL stages. These results complement the cross-backbone comparison with 3DThinker in Figure 5 and support the general applicability of our method across model families and scales. We have added these results to Appendix C.2 of the updated version.

Table C: Results across model families and scales on MindCube-Tiny, MMSI-Bench, SPAR-Bench-MV, SPAR-Bench-SI, and CV-Bench.

BackboneMethodMindCube-TinyMMSI-BenchSPAR-Bench-MVSPAR-Bench-SICV-Bench
Qwen3-VL-8BBase33.430.837.940.885.1
Ours (SFT)74.534.351.448.786.8
Ours (SFT+RL)80.837.259.054.589.1
InternVL3.5-8BBase36.728.239.637.581.7
Ours (SFT)78.432.953.944.982.6
Ours (SFT+RL)82.735.058.347.484.0
Qwen3.8-27BBase46.031.845.752.887.5
Ours (SFT)84.536.655.857.589.0
Ours (SFT+RL)88.440.563.261.090.3
AuthorsOpenReview ↗

Rebuttal by Authors [2/3]

W2. Improvement over HATCH on SPAR-Bench-MV.

Thank you for bringing up this point. We agree that our improvement over HATCH on SPAR-Bench-MV is modest (54.1 vs. 53.7). We would like to respectfully clarify that we do not intend to claim a large margin on SPAR-Bench-MV. As stated in Section 3.2 (L398-400), our gains over HATCH are most pronounced on MindCube-Tiny (+27.6) and MMSI-Bench (+7.8). To state our results more precisely, we have revised Section 3.2 in the updated version to describe the gains on each benchmark more accurately.

W3. Ablation studies beyond MindCube-Tiny.

Thank you for your thoughtful suggestion. We have extended the component ablations to all remaining benchmarks (MMSI-Bench, SPAR-Bench-MV, SPAR-Bench-SI, and CV-Bench) with the same Qwen2.5-VL-3B-based variants as in Figure 6. Table D shows that each component contributes on every benchmark, including the out-of-domain MMSI-Bench and CV-Bench. We have added these results to Appendix C.3 of the revised manuscript.

Table D: Ablation study on the MMSI-Bench, SPAR-Bench-MV, SPAR-Bench-SI, and CV-Bench.

ComponentsMMSI-BenchSPAR-Bench-MVSPAR-Bench-SICV-Bench
Qwen2.5-VL-3B (base model)25.624.923.970.7
SFT w/o camera & scene latent26.037.333.468.0
SFT w/o camera latent28.341.840.271.2
SFT w/o scene latent29.542.436.271.9
Full method SFT w/o RL32.146.042.673.0
GRPO w/o latent-mode entry exploration32.949.444.373.5
Full method34.854.147.775.8
AuthorsOpenReview ↗

Rebuttal by Authors [3/3]

W4. Discussion of related work.

Thank you for highlighting these works. We have included all these relevant works in Section 4 of the revised manuscript. We summarize the distinctions between our method and the mentioned related works below.

  • RL for latent reasoning (Latent-GRPO [4], ALAR [5]): These two works focus on text-only LLMs, whereas our method targets multimodal LLMs for multi-image spatial reasoning.
  • Spatial reasoning (Actial [6], GeoAnchor [7]): Actial is a data and training recipe to strengthen viewpoint understanding in MLLMs, with the model architecture unchanged and reasoning performed purely in text. For GeoAnchor, we would like to respectfully clarify that it was discussed in Section 4.2 of the original submission. GeoAnchor is a latent visual reasoning approach that targets single-image spatial reasoning and requires task-specific annotations. In contrast, our work addresses multi-image inputs and requires no such annotations for latent supervision.
  • Analysis of latent visual reasoning (Beyond Visual Memory [8]): This work is primarily a diagnostic study of latent visual reasoning, in which a probe method is developed to test whether latent contents are used. Our work instead focuses on grounding latent states in geometry tokens for multi-image spatial reasoning.

References

[1] Qwen3-VL Technical Report

[2] InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency

[3] Qwen3.8-Max: A New Bar for Coding and Cowork

[4] Latent-GRPO: Group Relative Policy Optimization for Latent Reasoning

[5] Adaptive Latent Agentic Reasoning

[6] Actial: Activate Spatial Reasoning Ability of Multimodal Large Language Models

[7] GeoAnchor: Collaborative Reasoning via Latent Decomposition for 3D Spatial Understanding

[8] Beyond Visual Memory: Mechanistic Diagnostics of Latent Visual Reasoning


We sincerely thank you again for the thoughtful comments. They have motivated additional experiments, ablations, and clarifications, which we believe have significantly strengthened our manuscript. Please do not hesitate to let us know if you have any further comments. We will be glad to discuss them.

Reviewer_tzLzOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents "Thinking with Latent Geometry Tokens", a two-stage framework that grounds latent states in compact geometric representations and encourages flexible latent-mode entry. The method uses camera and scene tokens from a pretrained 3D foundation model and supervises latent states with viewpoint-conditioned and cross-view geometry information, while jointly training for text generation. Then it applies GRPO with latent-mode entry-point exploration, so that successful exploratory trajectories can guide policy updates. This study uses MindCube-Tiny (spatial mental modeling from limited views), MMSI-Bench (multi-image spatial reasoning), SPAR-Bench-MV (spatial perception and reasoning across multiple views), and Qwen2.5-VL-3B as the default backbone for the experiments. This paper reports that the proposed method consistently outperforms spatial reasoning baselines.

Critical strengths
  • The motivation for latent-mode entry exploration is interesting. The observation that different entry points recover different failed instances provides a compelling motivation for exploring when to invoke latent reasoning and constitutes a clear contribution to the incorporation of representations from 3D foundation models into VLM reasoning.
  • The component ablations clearly demonstrate the effectiveness of the individual components. In particular, the comparison with additional unbiased rollouts controls for both the additional sampling budget and the retention of successful trajectories, providing useful evidence for the effectiveness of entry exploration beyond simply increasing the number of samples.
Critical weaknesses
  • Limited methodological novelty beyond latent-mode entry exploration. The overall framework of incorporating geometric representations from a 3D foundation model into VLM reasoning has already been explored by 3DThinker (Chen et al., 2026). The geometry-grounded latent learning component appears to be an incremental modification of this framework, replacing dense geometric features with camera and scene tokens available in VGGT-Ω. While this modification is empirically useful, its conceptual novelty appears limited. Consequently, latent-mode entry exploration seems to be the primary methodological contribution.
  • Insufficient analysis of the primary contribution. Given Weakness 1, the paper would benefit from placing greater emphasis on latent-mode entry exploration and providing a deeper analysis of its behavior. For example, it remains unclear whether the learned policy selects different entry points for different question categories or largely converges to a generally favorable position. Further analysis (e.g., entry-point distributions before and after RL, or their relationship to question categories and correctness) would help establish what the exploration procedure actually learns.
  • The baseline comparisons do not adequately establish improvements over the full 3DThinker method. The performance of 3DThinker-3B in Tables 1 and 2 is substantially lower than that reported in the original paper. For example, the original paper reports 75.2 on MindCube-Tiny after SFT+RL, whereas Table 1 reports 62.7, corresponding to the SFT-only model. It also reports 41.9 on MMSI-Bench and 78.4 on CV-Bench for its SFT+RL model trained on the larger curated dataset, compared with 19.0 and 47.7 in this paper. Appendix B explains that the main comparisons use the publicly available SFT checkpoint. However, comparing the proposed SFT+RL method against an SFT-only version of the closest baseline does not establish superiority over the full method. Moreover, even the original paper's SFT-only model trained on the larger dataset achieves 37.7 on MMSI-Bench and 74.5 on CV-Bench, suggesting that the discrepancies extend beyond the absence of RL. Although these models differ in their training data and potentially their evaluation settings, the magnitude of the discrepancies warrants further explanation. The main tables should clearly identify the baseline checkpoint and training stage, and ideally include the original reported results as separate entries with the relevant differences explicitly documented. Figure 5 provides a useful comparison of MindCube-Tiny, but it does not resolve these concerns for the other benchmarks.

Chen et al. 2026. Think with 3D: Geometric Imagination Grounded Spatial Reasoning from Limited Views. CVPR.

Questions for authors

Nothing in particular (see Critical Weaknesses).

4 comments
AuthorsOpenReview ↗

Rebuttal by Authors [1/3]

Dear Reviewer tzLz,

We appreciate your constructive feedback. We are glad that you found the motivation for latent-mode entry exploration compelling and the ablation studies supportive of our design choices. We sincerely hope our responses below could address your concerns.


W1. Methodological novelty regarding geometry-grounded latent learning.

Thank you for raising this point, and for recognizing latent-mode entry exploration as a clear contribution of our work. We would like to respectfully clarify that geometry-grounded latent learning differs from 3DThinker beyond the choice of features, in two respects.

  • Alignment mechanism: 3DThinker jointly aligns latent states and image embeddings to dense geometry features, whereas our method maps the latent states alone to compact geometry tokens (Eq. 3, L206-213). We adopt this design to encourage the geometric target to be recovered from the latent state itself, rather than also relying on the image embeddings.
  • Latent segment: 3DThinker aligns a fixed number of latent states as a whole, whereas we assign one camera latent and one scene latent to each input view (L193-202). Each latent state thus receives its own supervision signal and has a defined role, with both roles shown to contribute in our ablation (Figure 6).

We compare 3DThinker and our method at the SFT stage directly in Appendix C.1. With the same geometric teacher (VGGT-Ω), backbone, and training data, ours outperforms 3DThinker by 7.4 points on MindCube-Tiny (Figure 7). We have further clarified these distinctions in Sections 2.3 and 4.2 of the revised manuscript.

AuthorsOpenReview ↗

Rebuttal by Authors [2/3]

W2. Analysis of latent-mode entry distribution.

Thank you for your thoughtful comments. We appreciate your suggestions to further analyze the behaviour of latent-mode entry exploration to strengthen our work. As suggested, we have added a new analysis in Appendix C.4 (Figure 8). For each question category, we report the distribution of latent-mode entry positions after RL and the accuracy change on changed-entry questions in Table A. Entry positions, e.g., (0, 25%), are relative to the length of output tokens, and “Start (unchanged)” denotes entry at the beginning of the response, the same as before RL.

Based on the results in Table A, we find that 74.3% of the questions across the two benchmarks change their entry position after RL. In addition, the entry distribution varies across question categories on both the in-domain and out-of-domain benchmarks, suggesting that the learned policy does not converge to a single favorable position. Regarding correctness, accuracy improves after RL on the changed-entry questions in every category, further supporting the effectiveness of latent-mode entry exploration.

Table A: Distribution of latent-mode entry positions after RL and net accuracy gain on changed-entry questions.

BenchmarkCategoryStart (unchanged)(0, 25%)[25%, 50%)[50%, 75%)[75%, 100%]Net gain on changed-entry questions (pp)
MindCube-Tiny (In-domain)Rotation14.0%22.5%45.0%17.0%1.5%+6.4
Among25.2%57.5%13.0%2.7%1.7%+6.9
Around39.6%47.6%9.6%1.6%1.6%+10.6
MMSI-Bench (Out-of-domain)Position26.6%37.5%21.8%11.7%2.3%+2.3
Attribute42.3%39.2%11.5%5.4%1.5%+2.7
Motion22.7%35.3%26.7%14.0%1.3%+3.4
MSR10.6%19.2%38.9%25.8%5.6%+4.0
AuthorsOpenReview ↗

Rebuttal by Authors [3/3]

W3. Baseline comparisons with 3DThinker.

Thank you for your constructive suggestion.

  • Baseline settings. To support reproducibility, we evaluate open-weight baselines using their authors’ publicly released checkpoints in main tables, as described in Appendix B. To our knowledge, the only public 3DThinker checkpoint is the Qwen2.5-VL-3B based model trained with SFT on the 3DThinker-10K dataset. We use this checkpoint as our baseline. As you noted, the original paper reports MMSI-Bench and CV-Bench results for models trained on a much larger dataset, using 237K samples for SFT and a 57K-sample subset for RL. These differences in training data likely contribute to the observed performance gap. Following your suggestion, to make baseline settings clearer in main tables, we have now explicitly labeled 3DThinker’s training stage and dataset in the revised Tables 1 and 2.
  • Additional comparisons with 3DThinker. Thank you for acknowledging the value of the comparison with full 3DThinker (SFT+RL) in Figure 5. We agree that extending this comparison to additional benchmarks would better demonstrate the improvements of our method. To this end, we have retrained 3DThinker on our 15K curated dataset using its official implementation and default settings. The results across all benchmarks are presented below and have been added to Appendix C.1.

Table B: Comparison with 3DThinker retrained on our 15K training samples.

ModelMindCube-TinyMMSI-BenchSPAR-Bench-MVSPAR-Bench-SICV-Bench
Qwen2.5-VL-3B (Base)37.825.624.923.970.7
3DThinker (SFT)62.126.240.834.271.5
Ours (SFT)71.032.146.042.673.0
3DThinker (SFT+RL)72.928.046.136.072.2
Ours (SFT+RL)77.834.854.147.775.8

Thank you again for the constructive feedback. These insights have motivated us to conduct additional analyses and experiments that we believe have significantly strengthened our manuscript. Please do not hesitate to let us know if you have any further comments. We would be glad to discuss any further questions you may have.

AuthorsOpenReview ↗

Sincere Thanks and Follow-up to Reviewer tzLz

Dear Reviewer tzLz,

Thank you again for your time and effort in reviewing our paper. We would be grateful to hear whether our clarifications and the additional analyses in the revised manuscript have resolved your concerns. We remain committed to addressing any further points you may have during the discussion phase.

Best regards,

The Authors of Submission 17778

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Remarks by Authors

Dear Reviewers, ACs, SACs, and PCs,

Thank you for your time and effort in ensuring a high-quality review process for ICLR. We respectfully provide a summary of the discussion for the ACs' reference. For clarity and simplicity, we refer to Reviewers tzLz, ydEJ, and zmSD as R1, R2, and R3, respectively.

We are grateful that the reviewers recognized the key strengths of our work:

  1. An intuitive two-stage training framework (R2) for multi-image spatial reasoning.
  2. Compelling motivation (R1) and insightful design (R2) of latent-mode entry exploration.
  3. Substantial improvements on several benchmarks (R2), with controlled comparisons that provide clear evidence for the method (R3).
  4. Ablations that support the proposed components (R1, R2, R3), with the rollout budget controlled for entry exploration (R1, R3).

The main concerns were addressed as follows:

  1. R1 and R3 asked about the conceptual distinction from 3DThinker and further comparisons with it. We clarified how our alignment mechanism, latent segment design, and latent-mode entry differ, and retrained 3DThinker with its official implementation on the same 15K training samples as ours. Our method outperforms it on all five benchmarks at both the SFT and SFT+RL stages.
  2. R2 and R3 asked about the generalizability of our method. Additional experiments show that it consistently improves Qwen3-VL-8B, InternVL3.5-8B, and Qwen3.8-27B on all five benchmarks (R2, R3). The extended ablations show that each component contributes on every benchmark (R2), including the out-of-domain MMSI-Bench and CV-Bench, which share no images, scenes, or question templates with our training data (R3).
  3. R1 asked for a deeper analysis of latent-mode entry exploration. Our new analysis shows that the learned entry positions vary with the question category rather than converging to a single favorable position, and that accuracy improves on changed-entry questions in every category.

Following our rebuttal, R1 and R3 raised their scores from 2 to 3, and R2 maintained the positive rating of 3. All additional clarifications and experiments have been incorporated into the revised manuscript (highlighted in blue) and will remain in the final version.

We sincerely thank the reviewers for the thoughtful feedback and for carefully engaging with our rebuttal. We hope this summary is helpful to the ACs in making the final decision.

Best regards,

The Authors of Submission 17778

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.