TriViewBench: Controlled Complexity Scaling for Multi-View Structural Reasoning in MLLMs

TriViewBench is a new diagnostic benchmark designed to evaluate how Multimodal Large Language Models (MLLMs) handle structural reasoning across multiple viewpoints. Unlike existing benchmarks that use unconstrained real-world images, TriViewBench uses synthetic 3D scenes where object count and occlusion are explicitly parameterized. This allows researchers to isolate specific factors that cause MLLMs to fail, such as the difficulty of reconciling objects across different perspectives.

The researchers generated 1,923 synthetic scenes rendered from front, side, and top-down views. These scenes were organized into four complexity levels based on object density and occlusion severity. The benchmark includes over 14,000 question-answer pairs categorized into three types: Local Decision (simple spatial relations), Object Counting (enumerating objects), and Global Recovery (reconstructing full scene structure). The team evaluated 18 leading open- and closed-source MLLMs using both direct and Chain-of-Thought (CoT) prompting.

The study reveals a consistent 'capability hierarchy' across all 18 models: they are best at local decisions, struggle significantly with object counting, and collapse when tasked with global structural recovery. As complexity increases, performance drops sharply—particularly for Global Recovery, which saw an 80% relative decline.

Error analysis identified two distinct failure modes in counting: 'occlusion blindness' (undercounting objects hidden in a single view) and 'identity confusion' (overcounting the same object when viewed from multiple angles). Notably, Chain-of-Thought prompting failed to improve overall performance, suggesting that the models' inability to maintain a coherent 3D spatial representation is a more fundamental limitation than their reasoning process.

This study demonstrates that current MLLMs lack the structural grounding necessary for reliable multi-view reasoning. By providing a controlled framework, TriViewBench allows developers to pinpoint exactly where models fail, moving beyond generic accuracy scores to understand the underlying perceptual and spatial bottlenecks in multimodal AI.