Research map / Vision-language and multimodal / Multimodal LLMs

SpinBench: Perspective and Rotation as a Lens on Spatial Reasoning in VLMs

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 SpinBench, a cognitively grounded diagnostic benchmark of 2.7k VQA items across seven spatial task groups (perspective taking, rotation, translation, relative pose) and evaluates 43 VLMs, finding egocentric bias, poor rotation understanding, and inconsistency under reformulation.

Problem

Vision-language models' spatial reasoning, particularly perspective taking and rotational/viewpoint understanding, is poorly diagnosed; existing benchmarks entangle spatial reasoning with other abilities or lack controlled variation.

Approach

A progressively structured benchmark decomposing perspective taking into single-object diagnostic subtasks (identity matching, rotation, translation, canonical view) and multi-object scenes, with controlled variations in frame of reference, premise presence, symmetric and syntactic augmentation, and multi-image inputs, scored with chance-corrected kappa and pairwise consistency.

Key results

• GPT-5 ranks first in overall accuracy and consistency (97.1% consistency), with InternVL3-38B the top open-source model.
• Humans achieve 91.2% accuracy, and VLM accuracy correlates negatively with human response time (r=-0.54, p<0.05).
• On face rotation, top egocentric models (Gemini 2.5 Pro, kappa 0.94) are the worst on the allocentric variant (kappa -0.66), showing a strong egocentric bias.

Surprising findings

• Adding a depth map or using higher-resolution inputs consistently decreased or failed to improve accuracy for most models.
• 16 of 39 models perform below chance on premise-based perspective-taking tasks where relations are given in the text.

Limitations

• Does not cover containment, support, or vertical relations such as in, on, above.
• Tasks restricted to horizontal 2D plane and horizontal orbits.

Experimental setup

Datasets: Infinigen scenes ABO Multi-View Car Dataset Stereo Face Database YCB

Benchmarks: SpinBench

Metrics: accuracy Cohen's kappa pairwise consistency

Baselines: GPT-5 GPT-4o Gemini 2.5 Pro InternVL3-38B Qwen2.5-VL-32B SpaceOm SpaceThinker Cosmos-Reason1

Builds on: CLEVR Infinigen

Closest papers by meaning

• SpatialRGPT: Grounded Spatial Reasoning in Vision-Language Models (NeurIPS 2024 · similarity 0.92)
• InternSpatial: A Comprehensive Dataset for Spatial Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.92)
• SpatiaLab: Can Vision–Language Models Perform Spatial Reasoning in the Wild? (ICLR 2026 · similarity 0.91)
• 3ViewSense: Spatial and Mental Perspective Reasoning from Orthographic Views in Vision-Language Models (ICML 2026 · similarity 0.91)
• Seeing from Another Perspective: Evaluating Multi-View Understanding in MLLMs (AAAI 2026 · similarity 0.91)
• Thinking in Structures: Evaluating Spatial Intelligence in Constraint-Governed Spaces (ICML 2026 · similarity 0.91)