Research map / Vision-language and multimodal / Multimodal LLMs

GIQ: Benchmarking 3D Geometric Reasoning of Vision Foundation Models with Simulated and Real Polyhedra

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Spatial reasoning in VLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces GIQ, a benchmark of 224 synthetic (Mitsuba-rendered) and real paper-model polyhedra images with meshes, testing monocular 3D reconstruction, symmetry probing, mental rotation, and zero-shot classification in vision and VLM models.

Problem

Vision foundation models and VLMs look strong on standard benchmarks, but their understanding of precise 3D geometric properties (symmetry, convexity, face geometry, complexity) is unclear; existing 3D analysis datasets lack fine-grained geometric ground truth.

Approach

Use polyhedra, which have well-defined categories and symmetry groups, as an unambiguous testbed; pair 224 shapes with synthetic renderings and photographed paper models in indoor/outdoor settings, and evaluate reconstruction, probing, mental rotation, and zero-shot naming.

Key results

• Platonic-solid single-image reconstruction by Shap-E, Stable Fast 3D, and OpenLRM is unreliable; even a cube is often reconstructed incorrectly on wild images.
• DINOv2 reaches up to 93% balanced accuracy on 4-fold symmetry for wild images with a linear probe trained on synthetic data.
• On the hard mental rotation split, SigLIP with a non-linear probe reaches 69% and DINOv2 67%, vs. 68.05% average human accuracy from a 42-participant study.

Surprising findings

• Non-linear probes gave comparable average performance to linear probes for symmetry detection.
• Chain-of-thought prompting yielded minimal gains and multi-view inputs only marginal improvements for frontier VLM zero-shot classification.

Limitations

• Wild images are limited to paper models of polyhedra; 48 of 92 Johnson solids are included in the wild set.
• Excluded Johnson solids from symmetry detection due to viewing-angle ambiguities.
• Mental rotation mostly evaluated with a single hard split and a small human study, limiting generality.

Experimental setup

Datasets: GIQ synthetic renderings GIQ wild images

Benchmarks: GIQ

Metrics: balanced accuracy accuracy

Baselines: Shap-E Stable Fast 3D OpenLRM CLIP DINO DINOv2 SigLIP MAE VGGT DUSt3R MASt3R Claude 3.7 Sonnet

Builds on: Mitsuba DINOv2 Objaverse TripoSR LRM

Closest papers by meaning

• G$^2$VLM: Geometry Grounded Vision Language Model with Unified 3D Reconstruction and Spatial Reasoning (CVPR 2026 · similarity 0.87)
• PolyhedronNet: Representation Learning for Polyhedra with Surface-attributed Graph (ICLR 2025 · similarity 0.87)
• Beyond 3D VQAs: Injecting 3D Spatial Priors into Vision-Language Models for Enhanced Geometric Reasoning (CVPR 2026 · similarity 0.87)
• Thinking in Structures: Evaluating Spatial Intelligence in Constraint-Governed Spaces (ICML 2026 · similarity 0.87)
• Think with 3D: Geometric Imagination Grounded Spatial Reasoning from Limited Views (CVPR 2026 · similarity 0.86)
• Imagine3D-LLM: Teaching MLLMs to Imagine 3D Scenes Before Answering (NeurIPS 2026 · similarity 0.86)