acceptodds
Withdrawn from ICLR 2027

Seeing in 2D, Reasoning in 3D: What Explicit Geometry Reveals about Vision-Language Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Does making geometry explicit remove the need for vision and is it enough for 3D reasoning? End-to-end accuracy alone cannot distinguish the benefit of supplying geometric information from the ability to solve problems once that information is available. We introduce GeoText, a structured representation for diagnostic input interventions over four geometric components: view elements, 3D shape structure, given measurements and constraints, and spatial relations. Using OrthoMind-Bench, a collection of 5,357 Chinese K–12 solid-geometry problems, we evaluate six open-weight vision-language models under single-component and cumulative interventions, each with and without images. Providing 3D shape descriptions yields the largest standalone gains across all six models, improving accuracy by 13.2–28.7 percentage points; substantial gains persist even after view descriptions are supplied. Additional givens provide much smaller gains once views and shape are available. Explicit geometry also changes the marginal value of vision: images improve accuracy by 10.5–27.9 points without GeoText, but their contribution becomes near-zero or negative in four models when all GeoText components are supplied. Yet, with both images and all GeoText components, model error rates remain 32.6–47.8%. These results reveal a separation between reduced visual dependence and successful geometric problem solving: external geometric descriptions can substantially reduce the benefit of images while leaving considerable failures unresolved. GeoText provides a diagnostic framework for measuring which geometric information improves performance, how it changes the value of vision, and what failures persist under explicit geometric support.

Primary area:
datasets and benchmarks
Keywords:
Vision-Language Models, 3D Geometric Reasoning, Structured Geometric Representations, Diagnostic Interventions, Visual Dependence

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_wW7DOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies which explicit geometry representation is the most useful for solid-geometry problem solving with vision-language models (VLMs). To figure out this question, the paper introduces GeoText, which is a structured representation for diagnostic analysis over four geometric input information: view elements, 3D shape structure, given measurements and constraints, and spatial relations. To evaluate on this question, a new benchmark called OrthoMindBench containing mostly three-view drawing questions is constructed. The paper evaluates on the constructed benchmarks across different VLM models, and comes to the conclusion that explicit geometry information given to the model can substantially reduce the contribution of images for solid-geometry problem solving.

Critical strengths
  • The paper presents a valuable study of analyzing how much each geometric component contribute to the solid-geometry problem solving question with VLMs. The conclusion of "explicit geometry information dominates the benefits from images" seems to be natural, but can also be informative to provide guidance for future works on solving mathematical geometry problems.

  • The paper constructs OrthoMindBench, which can be useful for future research on solid-geometry problems, especially for the controlled experiments for structured representations to study the individual component contribution to the performance of solving mathematical geometry problems.

Critical weaknesses
  • Compared to existing benchmarks on solid-geometry tasks like SolidGeo [1], the questions in the constructed OrthoMind-Bench seems to have limited diversity. In Lines 147-148, the paper states that over 80% of the questions are three-view drawings, and the rest of the questions being in unspecified categories. In comparison, SolidGeo states in its Section 3.1 that the questions have 8 fine-grained domains, and are categorized into three levels of difficulty. Therefore, the constructed benchmark seems to be less comprehensive and carefully curated than the existing benchmark.

  • The details of the proposed benchmark is also not clear presented. For example, in Lines 149-150, the paper states that "two batches of 1,420 and 3,937 problems have disjoint identifiers and are not difficulty strata". However, I feel confused about how are these two batches of benchmarking data separated, and why we even need two batches of data. The illustration of the dataset breakdown in Appendix C demonstrates that these two batches have very different coverage of the properties like "Relations", as one batch is all covered while the other only has less than 10% covered, which makes people feel even more confused on the purpose of having two batches. Also, there is seldom an example of how the question looks like. The only visual available in the paper is the three-view diagram in Figure 1. Other than that, for example, for the example question presented in Section A.2, the problem only has text description without any illustration, which makes audience very hard to get a taste of what the example questions really look like.

The above two points are the major weaknesses of the paper. Other weaknesses include that the title of the paper is confusing and kind of misleading, as the title kind of suggest 3D-aware spatial reasoning within VLMs, which is nowhere close to the actual topic studied in the paper, which is basically reasoning on geometry for AI-for-Math problems. This can attract many wrong audience to this paper, and I think the keywords like "math" needs to be added in the title.

[1] Wang et al. SolidGeo: Measuring Multimodal Spatial Math Reasoning in Solid Geometry. NeurIPS 2025.

Questions for authors
  • What is the specific question breakdown for the proposed OrthoMind-Bench? For example, for the rest of 20% questions beyond three-view drawings, what question types do they belong to? Also, how is the question diversity compared to existing benchmarks on K-12 solid-geometry questions including SolidGeo?

  • What is the purpose of having two batches for the proposed benchmark? What is each split of the benchmark tailored for?

Reviewer_bNojOpenReview ↗
Rating
1
Confidence
2(−1)
Informativeness
1
Presentation
2(+1)
Soundness
2(+1)
Read review
Paper summary

This paper addresses the difficulty of distinguishing visual information acquisition failures from subsequent reasoning failures in vision-language models (VLMs) on 3D geometry problems. This paper proposes GeoText, a structured geometric representation framework that categorizes geometric information into four components: views, 3D structure, given conditions, and spatial relations. Controlled experiments on 5,357 solid geometry problems using six models investigate the effects of different information combinations, with and without images. Results show that 3D structural descriptions yield the greatest accuracy improvement. Explicit geometric information substantially reduces image dependence, yet 32.6%–47.8% of problems remain unsolved, highlighting the gap between reduced visual dependence and improved geometric reasoning.

Critical strengths

The research idea is relatively novel. GeoText provides an interesting intervention-based framework for evaluating VLMs' 3D geometric understanding.

Critical weaknesses
  1. The paper lacks concrete examples showing model outputs and results under different information intervention settings. This makes it difficult to intuitively understand how individual geometric components affect reasoning and when GeoText improves or fails to improve problem-solving performance.
  2. Several figures and captions appear clearly inconsistent, undermining the clarity and rigor of the presentation. For example, the caption of Figure 3 claims to illustrate the marginal contribution of images when explicit geometric information is provided. However, its four panels actually show the distributions of information dimensions, difficulty levels, geometry types, and task categories, which do not match the caption.
Questions for authors
  1. Could you provide concrete benchmark examples to illustrate the practical effectiveness of the GeoText framework?
  2. Could you review and correct the inconsistencies between figures and captions throughout the paper, ensuring that figure labels align with the explanations of experimental results in the main text?
1 comment

Thank you for the clarification and revisions. Your response has largely addressed my concerns. However, I still find the empirical validation insufficient and do not consider the work ready for acceptance. I will therefore maintain my original rating.

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

This paper investigates how explicit geometric information affects vision-language models’ ability to solve 3D geometry problems from 2D diagrams. It introduces GeoText, a diagnostic framework that organizes model-generated descriptions into four components: view elements, 3D shape structure, given measurements and constraints, and spatial relations. Using OrthoMind-Bench, a collection of Chinese K–12 solid-geometry problems, the authors evaluate six open-weight VLMs by supplying these components individually and cumulatively, with and without images, while retaining the original questions. This design measures the benefit of each description, its incremental value alongside other information, and the remaining contribution of visual input.

Critical strengths
  1. The consistent benefit of Shape descriptions is a clear empirical strength. They improve performance across all evaluated models, both independently and after view descriptions are supplied. The latter comparison shows that describing the projections does not exhaust the value of describing the solid. This supports the usefulness of the complete Shape representation without attributing the gains to an isolated reconstruction capability.
  2. The comparisons with and without images reveal how geometric support changes the value of visual input. Images help every model at baseline, but their contribution becomes negligible or negative for some models under full GeoText support, while remaining positive for others. This model-dependent pattern is more informative than a general claim that textual geometry makes vision unnecessary.
  3. The intervention design usefully separates standalone benefits from gains conditional on previously supplied information. The contrasting results for Givens illustrate why this distinction matters. The coverage analysis also helps interpret the limited gains from Relations, considering both a fully covered subset and the possibility that earlier components already express relational information.
  4. The manuscript generally interprets its measurements carefully. It distinguishes net accuracy gains from repaired errors and acknowledges the algebraic relationship between reduced image benefit and larger text-only gains. It also avoids treating full GeoText as verified complete geometry or residual errors as purely reasoning failures. These qualifications strengthen the conclusions, although some figure labels should be brought into alignment with them.
Critical weaknesses
  1. The diagnostic conclusions require stronger validation of the supplied descriptions. The paper does not quantitatively assess their correctness, completeness, or potential to reveal answers. Excluding explicitly computed results does not guarantee that the remaining descriptions contain no information that directly resolves the question. In particular, Shape includes dimensions and relations alongside structural information, so its gains may reflect both useful geometric support and partial answer disclosure.
  2. The distinction from closely related diagnostic work needs sharper articulation. MathVerse receives limited comparison at the level of its information categories and input variants, while MathLens and Your Reasoning Benchmark May Not Test Reasoning are absent from the references despite their relevant representation and modality comparisons. GeoText’s solid-geometry components and conditional interventions may offer a useful distinction, but this should be established through a focused methodological comparison rather than an indiscriminate expansion of benchmarks.
  3. Several figures conflict with their captions and the paper’s stated interpretation. Figure 3 shows dataset composition rather than image utility, while Figure 4 is a stacked bar chart described as a scatter plot. The label “Fixed by Oracle structure” also incorrectly presents net improvement as repaired errors and GeoText as oracle information. Although the tables support reconstruction of the main numerical findings, the visual presentation needs correction.
  4. The final-answer scoring procedure remains insufficiently specified. Limited string normalization cannot establish general mathematical equivalence, particularly when most answers are open-form. The rechecking model’s prompt, triggering conditions, adjudication rules, and handling of failed checks are not documented, and the resulting correctness labels are not validated. These omissions complicate interpretation of small accuracy differences and residual error rates.
  5. The benchmark’s construction is less clearly documented. The paper reports counts and labeling procedures but leaves data sources, selection criteria, semantic and image deduplication, and reference-answer verification unclear. Disjoint identifiers alone do not establish distinct problems. These details are needed to assess the benchmark’s reusability and whether recurring templates disproportionately influence the aggregate findings.
Questions for authors
  1. Could you provide the GeoText generation prompt and specify all inputs available to Gemini-3.1-Pro-Preview, including any reference answers or worked solutions? A modest stratified audit distinguishing inaccurate descriptions, missing necessary information, and fields that directly answer the question would be helpful. Reporting the main Shape comparisons on a verified subset would clarify how much of the improvement comes from supporting subsequent reasoning.

  2. What additional diagnostic insight does the four-component design provide beyond MathVerse’s information variants, MathLens’s aligned evaluations, and Wang et al.’s description-based comparisons? In particular, which finding requires separating Views from Shape and could not be obtained by comparing complete descriptions with images? Linking this distinction to the existing experiments would be more informative than substantially expanding the benchmark suite.

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

Related papers

Discussion (0)

Sign in to comment.