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

Researchers introduce TriViewBench, a controlled benchmark for evaluating multimodal AI models' ability to reason across multiple 3D views with varying complexity. Testing 18 MLLMs reveals a universal capability hierarchy and severe performance degradation on complex tasks, particularly in cross-view spatial reasoning, suggesting fundamental limitations in current AI architecture.

TriViewBench represents a critical diagnostic tool for identifying weaknesses in multimodal large language models that standard benchmarks fail to capture. While MLLMs perform well on existing visual question-answering datasets, this research demonstrates they collapse under controlled structural complexity, with performance declining up to 80% on global recovery tasks. The uniform hierarchy across all 18 tested models—regardless of architecture or training approach—indicates these limitations reflect fundamental design constraints rather than model-specific issues.

The benchmark's value lies in its systematic parameterization of complexity through occlusion and object count, enabling researchers to pinpoint where reasoning fails mechanistically. The study reveals two distinct failure modes in object counting: single-view undercounting due to occlusion blindness versus multi-view overcounting from identity confusion across perspectives. These findings suggest MLLMs struggle with spatial representation and cross-view consistency rather than purely logical reasoning.

For the AI development community, this work underscores that scaling model size or improving instruction-following techniques addresses only part of the problem. The near-zero benefit from Chain-of-Thought prompting indicates the bottleneck resides in perceptual-spatial understanding, not reasoning strategy. This has implications for real-world applications requiring spatial reasoning—autonomous systems, 3D scene understanding, and multi-camera inference.

Moving forward, addressing these structural reasoning limitations likely requires architectural innovations beyond current transformer-based approaches. TriViewBench provides a standardized framework for measuring progress on these fundamental challenges, making it essential for researchers developing next-generation multimodal models.

• →All 18 MLLMs tested show identical capability hierarchy: Local Decision > Object Counting > Global Recovery, indicating universal architectural limitations.
• →Performance degrades severely with complexity: 80% drop on global recovery tasks reveals fundamental scalability constraints.
• →Cross-view identity confusion and occlusion blindness represent mechanistically independent failure modes requiring distinct architectural solutions.
• →Chain-of-Thought prompting provides negligible benefit (−0.16%), suggesting reasoning bottleneck lies in spatial representation, not logic.
• →TriViewBench establishes a controlled diagnostic framework for systematically evaluating multimodal model weaknesses.