GamiBench: Evaluating Spatial Reasoning and 2D-to-3D Planning Capabilities of MLLMs with Origami Folding Tasks

TL;DR

GamiBench addresses the gap in evaluating spatial reasoning in multimodal LLMs by tying 2D crease patterns to 3D folds across multiple viewpoints and time. It introduces origami-inspired tasks, two reasoning regimes (SSSU and Multi-Step), and new metrics for viewpoint consistency and physical feasibility, enabling a holistic assessment of geometric understanding. The benchmark reveals that even leading models struggle with multi-view coherence and feasibility under sequential folding, highlighting a need for geometry-aware training and architectures. By providing a standardized, cross-view planning framework, GamiBench aims to drive advances in spatial reasoning core to real-world manipulation and embodied AI.

Abstract

Multimodal large language models (MLLMs) are proficient in perception and instruction-following, but they still struggle with spatial reasoning: the ability to mentally track and manipulate objects across multiple views and over time. Spatial reasoning is a key component of human intelligence, but most existing benchmarks focus on static images or final outputs, failing to account for the sequential and viewpoint-dependent nature of this skill. To close this gap, we introduce GamiBench, a benchmark designed to evaluate spatial reasoning and 2D-to-3D planning in MLLMs through origami-inspired folding tasks. GamiBench includes 186 regular and 186 impossible 2D crease patterns paired with their corresponding 3D folded shapes, produced from six distinct viewpoints across three visual question-answering (VQA) tasks: predicting 3D fold configurations, distinguishing valid viewpoints, and detecting impossible patterns. Unlike previous benchmarks that assess only final predictions, GamiBench holistically evaluates the entire reasoning process--measuring cross-view consistency, physical feasibility through impossible-fold detection, and interpretation of intermediate folding steps. It further introduces new diagnostic metrics--viewpoint consistency (VC) and impossible fold selection rate (IFSR)--to measure how well models handle folds of varying complexity. Our experiments show that even leading models such as GPT-5 and Gemini-2.5-Pro struggle on single-step spatial understanding. These contributions establish a standardized framework for evaluating geometric understanding and spatial reasoning in MLLMs. Dataset and code: https://github.com/stvngo/GamiBench.

Figures (5)

• Figure 1: 2D normal crease pattern (top-left), right-view 3D final fold state (top-right), top-view alternative 3D final fold state (bottom-left), and 2D impossible crease pattern (bottom-right) of an example butterfly variation (komatsu_butterfly). Impossible crease pattern has inverted creases and missing valley/mountain folds.
• Figure 2: Visual mapping of the task flow in GamiBench. The model receives a 2D crease pattern (komatsu_dolphin) and text prompt, encodes candidate 3D folds, and outputs the most plausible match among multiple choices.
• Figure 3: Task-specific template. Our QA template includes instructions, question, and answer for the SSSU task.
• Figure 4: GamiBench Closed-Source Evaluations. Regular accuracy (top) and Impossible accuracy (bottom) of models, sorted in descending order from left to right, complex and simple.
• Figure 5: GamiBench Open-Source Evaluations. Regular accuracy (top) and Impossible accuracy (bottom) of models, sorted in descending order from left to right, complex and simple.