Research map / Vision-language and multimodal / Multimodal LLMs
Paper Folding Puzzles: Can Multimodal Large Language Models Perform Spatial Reasoning?
Dibin Zhou, Yantao Xu, Zongming Huang, Zengwei Yan, Wenhao Liu, Yongwei Miao, Jianfeng Ren, Fuchang Liu
Read the paperCheck your idea against 137,217 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 a paper-folding puzzle benchmark to test whether multimodal LLMs can perform spatial and abstract visual reasoning.
Problem
MLLMs lag human performance on abstract visual reasoning (AVR), and existing AVR benchmarks are limited to narrow, repetitive 2D patterns with simple spatial relations and few reasoning dimensions.
Approach
Uses paper-folding puzzles, inspired by human spatial reasoning, as a benchmark that requires inferring latent folding rules from visual question sets and generalizing them to novel scenarios, with richer spatial relationships than prior 2D AVR tasks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| compare humans with best MLLM on spatial reasoning puzzles | MLLMs on PFP · PFP, five tasks | accuracy (%): human 72.64 vs best model 41.67 on single-step; 86.50 vs 41.00 on multi-step | Hurt |
| increase folding complexity from 1-Fold to 3-Fold and corner-fold Others (ablation) | MLLMs · PFP Single-Step Reasoning | accuracy (%): o4-mini 34.43 (1-Fold) -> 26.92 (Others) | Hurt |
| add viewpoint variation to the 3D folding task (easy vs hard) (ablation) | humans · PFP 3D Folding | accuracy (%): 85.00 -> 80.33 | Hurt |
| add viewpoint variation to the 3D folding task (ablation) | most MLLMs · PFP 3D Folding | accuracy (%): o4-mini 30.67 -> 30.67 | No clear effect |
| change candidate net shapes in 2D unfolding (easy vs hard) (ablation) | humans · PFP 2D Unfolding | accuracy (%): 84.00 -> 83.33 | No clear effect |
| use closed-source reasoning model o4-mini instead of open-source models | MLLMs · PFP overall | accuracy (%): Multi-step 41.00 vs open-source best 27.00 | Improved |
| scale up Qwen2.5-VL from 7B to 72B (ablation) | Qwen2.5-VL · PFP Single-Step Reasoning | accuracy (%): 22.83 -> 31.67 | Improved |
| scale up Qwen2.5-VL from 32B to 72B (ablation) | Qwen2.5-VL · PFP Single-Step Reasoning | accuracy (%): 29.17 -> 31.67 | No clear effect |
| use thinking or reasoning models versus non-reasoning GPT-4o on 3D tasks | MLLMs · PFP 3D Folding and 2D Unfolding | accuracy (%): 3D Folding o4-mini 30.33 vs GPT-4o 26.50 | No clear effect |
Closest papers by meaning
- Unfolding Spatial Cognition: Evaluating Multimodal Models on Visual Simulations (ICLR 2026 · similarity 0.92)
- ORIGAMISPACE: Benchmarking Multimodal LLMs in Multi-Step Spatial Reasoning with Mathematical Constraints (NeurIPS 2025 · similarity 0.92)
- Is A Picture Worth A Thousand Words? Delving Into Spatial Reasoning for Vision Language Models (NeurIPS 2024 · similarity 0.92)
- SpatialScore: Towards Comprehensive Evaluation for Spatial Intelligence (CVPR 2026 · similarity 0.92)
- Spatial-MLLM: Boosting MLLM Capabilities in Visual-based Spatial Intelligence (NeurIPS 2025 · similarity 0.92)
- Actial: Activate Spatial Reasoning Ability of Multimodal Large Language Models (NeurIPS 2025 · similarity 0.92)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
