Research map / Vision-language and multimodal / Multimodal LLMs

MMPerspective: Do MLLMs Understand Perspective? A Comprehensive Benchmark for Perspective Perception, Reasoning, and Robustness

NeurIPS 2025 · 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 MMPerspective, a 10-task multiple-choice benchmark (2,711 images, 5,083 QA pairs) evaluating MLLMs' perspective perception, reasoning, and robustness under perspective-preserving transformations, and evaluates 43 MLLMs.

Problem

Whether multimodal large language models internalize perspective geometry (vanishing points, horizon lines, line relations, perspective type) and keep consistent geometric interpretations under appearance-preserving image transformations remains untested.

Approach

A structured benchmark spanning perception, reasoning, and robustness dimensions for perspective, with a consistency-based robustness metric (binary and graded) computed over perspective-invariant perturbations (cropping, flipping, masking) of each seed image, plus a Blender-based pipeline producing images with ground-truth vanishing point coordinates.

Key results

• GPT-4o and Gemini-2-flash achieve the highest overall accuracy (57.7% and 57.6% as stated in text), while Gemini-2-flash (CoT) reaches 57.2% overall and 69.8% on VPP.
• Overall accuracy correlates with model size at r=0.81 across families, while robustness correlates at only r=0.34; vision encoder size correlates with robustness at r=0.15.
• Chain-of-thought prompting yields average gains of +5.59 points overall accuracy and +6.63 points robustness across GPT-4o, Gemini-1.5-flash, and Gemini-2-flash.

Measured effects reported in the paper

Surprising findings

• Many large models with high perception accuracy score very low on robustness, e.g. InternVL3-38B has 9.1% binary robustness despite strong perception.
• One-shot in-context learning helps GPT-4o (50.0% to 58.3% overall) but hurts GPT-4o-mini (37.6% to 34.1%).

Limitations

• Multiple-choice format only, without open-ended or dynamic spatial tasks.
• Dataset skews toward architectural and indoor scenes.
• Some tasks contain ambiguous visual cues, so errors may reflect subjective interpretation.

Experimental setup

Datasets: RPVP dataset web-collected architectural and indoor images self-shot real-world fish-eye and linear perspective photos Blender-rendered scenes with ground-truth VP coordinates

Benchmarks: MMPerspective

Metrics: accuracy binary robustness score graded robustness score

Baselines: GPT-4o Gemini-2-flash Gemini-1.5-flash GPT-4o-mini InternVL3-78B Qwen2.5-VL-72B

Builds on: RPVP dataset Blender-MCP

Closest papers by meaning

• Seeing from Another Perspective: Evaluating Multi-View Understanding in MLLMs (AAAI 2026 · similarity 0.92)
• CoSpace: Benchmarking Continuous Space Perception Ability for Vision-Language Models (CVPR 2025 · similarity 0.90)
• Why Is Spatial Reasoning Hard for VLMs? An Attention Mechanism Perspective on Focus Areas (ICML 2025 · similarity 0.90)
• SpinBench: Perspective and Rotation as a Lens on Spatial Reasoning in VLMs (ICLR 2026 · similarity 0.90)
• Spatial-MLLM: Boosting MLLM Capabilities in Visual-based Spatial Intelligence (NeurIPS 2025 · similarity 0.89)
• Actial: Activate Spatial Reasoning Ability of Multimodal Large Language Models (NeurIPS 2025 · similarity 0.89)