- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper studies how geometry foundation models should be connected to VLMs for spatial reasoning. The authors argue that existing routes—QA-only fusion of geometry features and alignment of VLM latents to geometry features—can capture geometry without ensuring that the answering VLM actually uses it. They provide a diagnostic showing that zeroing 3DThinker latents leaves MindCube accuracy essentially unchanged and that answer tokens barely attend to them. At inference, the VLM answers directly from continuous geometric tokens without generating geometry CoT. Experiments across seven spatial benchmarks and three VLM families, plus ablations, geometry interventions, attention analysis, and failure cases, support the design.
Critical strengths
-
The paper moves the objective from “reconstruct/align geometry accurately” to “learn a geometric interface the receiving VLM can actually read and use.” The diagnostic that aligned latents can be zeroed without accuracy loss is a strong motivation, and the two-stage separation—adapt geometry to the VLM, then freeze the interface and learn to answer—is conceptually clean and practically relevant for embodied/autonomous systems.
-
GeoReader reaches 94.3 on MindCube-Tiny and 70.6 average across six benchmarks, with gains across Qwen2.5-VL, Qwen3-VL, and Qwen3.5 families. The ablations isolate readability pretraining, pair supervision, Reader freezing, and question-conditioned injection, while geometry interventions and attention analysis provide evidence that the model actually uses the supplied geometry rather than ignoring it.
Critical weaknesses
-
Stage 2 uses a multi-source QA collection derived from VSI-style, SpatialLadder, SPAR, ViewSpatial, MMSI, and CV-Bench-style data, while evaluation includes these benchmarks. The authors state that official held-out test questions are excluded and training scenes do not overlap with evaluation scenes, which is reassuring, but the paper should more explicitly document question-generation/split boundaries and whether all compared baselines received comparable QA data and tuning budgets. This matters for interpreting the six-benchmark averages, though the controlled MindCube ablations and Table 3 mitigate the concern.
-
The 3DThinker comparison partly relies on a released checkpoint and cited numbers. Table 4 Panel B includes adapted VLM-3R/SpaceMind-style controls, which is good, but a same-backbone, same-QA-data, same-tuning-budget feature-alignment control would make the central “readability vs. alignment” claim more directly supported.
Questions for authors
-
Please provide the exact split protocol between the multi-source Stage 2 QA collection and the six evaluation benchmarks, including whether any training questions are generated from the same benchmark scenes/templates as the test questions. Also clarify whether all baselines were trained/tuned with comparable QA data and budgets.
-
Stage 1 trains Readers with annotation-derived geometry, while main results use DA3-predicted geometry. How large is the domain gap, and can you quantify Reader sensitivity to pose/depth estimation errors beyond the aggregate results in Table 15?
-
Each view contributes one geometric token. Is there an ablation on the number of geometric tokens per view or a discussion of the information bottleneck? Also, after Stage 2 fine-tunes the VLM, how stable is the frozen Reader interface?
4 comments
Follow-Up to Your Review (Second Reminder)
Dear Reviewer UXxg,
Thank you again for the time and care you put into reviewing our work!
We know this is a busy period, so we wanted to follow up briefly on our point-by-point responses and the revised manuscript, where we highlighted the changes for easy reference.
If you have a chance to look over our responses, we would appreciate hearing whether they address your main concerns. If anything remains unclear, we would be happy to clarify it here.
Thanks again for your support!
Sincerely yours,
The Authors of Submission 4793
partially addressed my concerns
Thanks for the authors' responses and additional experiments. They have partially addressed my concerns and improve the experimental completeness of the paper.
Thank You for Your Response
Dear Reviewer UXxg,
Thank you for taking the time to read our rebuttal and assess the additional experiments. We appreciate your positive assessment of GeoReader and your rating in favor of acceptance.
We are glad that the new experiments addressed some of your concerns and made the evaluation more complete. If any concerns remain, we would be happy to discuss them here.
Thank you again for your time and feedback!
Warm regards,
The Authors of Submission 4793
Follow-Up to Your Review
Dear Reviewer
UXxg,Thank you again for the time and effort you put into reviewing our work!
We understand that this might be a super busy period for you. We wanted to briefly let you know that we have posted point-by-point responses to your questions and concerns and uploaded a revised manuscript with the changes highlighted.
Whenever you have a chance, we would be grateful for your feedback on whether the responses address your concerns. We would be happy to clarify anything that remains open.
Thank you!
Best,
The Authors of Submission
4793