Research map / Vision-language and multimodal / Multimodal LLMs

Seeing from Another Perspective: Evaluating Multi-View Understanding in MLLMs

AAAI 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 1 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 All-Angles Bench, a human-annotated multi-view QA benchmark (~2,100 questions, 90 scenes, six tasks) showing that 27 MLLMs fall well short of humans, especially on cross-view correspondence and camera pose estimation.

Problem

MLLMs used as embodied agents often fail at multi-view geometric consistency and cross-view correspondence, and no benchmark systematically evaluated multi-view understanding.

Approach

A multi-view benchmark built from Ego-Exo4D and EgoHumans scenes with six task categories, human-verified multiple-choice questions, and paired rephrased/reordered questions that measure answer consistency, plus analysis of CoT prompting and camera pose visualization.

Key results

• Human annotators reach 88.9% accuracy on camera pose estimation ordering, while Gemini-2.0-Flash, Qwen2.5-VL-72B and InternVL2.5-38B trail by over 50 points.
• Ovis2-34B and Qwen2.5-VL-72B outperform Gemini-2.0 and Claude-3.7-Sonnet on object manipulation and relative direction.
• GPT-4o shows around 70% inconsistency (IC) on relative distance, while other models sit near 40%; all models exceed 40% IC on relative direction.
• Identification CoT gives GPT-4o substantial gains on partial-visibility counting, but gains diminish or reverse for InternVL2.5-38B.

Surprising findings

• Many open-source MLLMs perform worse than random guessing on camera pose estimation.
• CoT prompting gives limited or negative benefit for models already strong at multi-view counting, such as InternVL2.5-38B.

Limitations

• Benchmark is small (about 2,100 questions over 90 scenes) and multiple-choice with three options.
• Human evaluation uses only a 250-question subset with two evaluators per question split.
• Closed-source models' training strategies are unknown, limiting causal analysis of failures.

Experimental setup

Datasets: Ego-Exo4D EgoHumans

Benchmarks: All-Angles Bench

Metrics: accuracy consistency (CC/WW/IC on paired questions)

Baselines: Gemini-2.0-Flash Claude-3.7-Sonnet GPT-4o Qwen2.5-VL-72B InternVL2.5-38B Ovis2-34B Deepseek-VL2 Cambrian LLaVA-OneVision LLaVA-NeXT-Video

Builds on: Chain-of-thought prompting Self-consistency VLMEvalKit

Closest papers by meaning

• Seeing Across Views: Benchmarking Spatial Reasoning of Vision-Language Models in Robotic Scenes (ICLR 2026 · similarity 0.93)
• From Correspondence to Actions: Human-Like Multi-Image Spatial Reasoning in Multi-modal Large Language Models (ICML 2026 · similarity 0.92)
• MMPerspective: Do MLLMs Understand Perspective? A Comprehensive Benchmark for Perspective Perception, Reasoning, and Robustness (NeurIPS 2025 · similarity 0.92)
• Multi-SpatialMLLM: Multi-Frame Spatial Understanding with Multi-Modal Large Language Models (CVPR 2026 · similarity 0.92)
• Actial: Activate Spatial Reasoning Ability of Multimodal Large Language Models (NeurIPS 2025 · similarity 0.92)
• Imagine3D-LLM: Teaching MLLMs to Imagine 3D Scenes Before Answering (NeurIPS 2026 · similarity 0.91)