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

2026-06-24

AI Summary

This work addresses the lack of systematic benchmarks for evaluating multimodal large language models (MLLMs) on multi-view spatial reasoning under controllable complexity. To this end, we introduce TriViewBench, the first benchmark based on synthetic 3D scenes that enables parametric control over object count and occlusion levels. Using a unified prompting protocol, chain-of-thought analysis, and cross-view consistency evaluation, we systematically assess 18 prominent MLLMs. Our experiments reveal a monotonic decline in model performance with increasing structural complexity, with pronounced degradation in object counting and global scene reconstruction tasks. Furthermore, we identify two distinct failure modes—occlusion-induced blind spots and cross-view identity confusion—offering fine-grained diagnostic insights into MLLMs’ spatial reasoning capabilities.

Abstract

Multimodal Large Language Models (MLLMs) demonstrate strong performance on standard visual question answering benchmarks, yet their scalability under controlled structural complexity remains poorly understood. We introduce TriViewBench, a controlled three-view visual reasoning benchmark constructed from synthetic 3D scenes with explicitly parameterized object count and occlusion. The benchmark contains 1,923 scenes and over 14K Question-Answer (QA) pairs organized into four complexity levels and three reasoning categories: Local Decision, Object Counting, and Global Recovery. We evaluate 18 open- and closed-source MLLMs under a unified prompting protocol. All 18 models exhibit an identical capability hierarchy without exception (Local Decision > Object Counting > Global Recovery), and performance degrades monotonically with complexity: Local Decision tasks decline modestly (12.11% relative drop), while Object Counting degrades substantially (59.14%) and Global Recovery collapses severely (80.02%). Error analysis on Object Counting reveals two mechanistically independent failure modes: single-view tasks are dominated by undercounting due to occlusion blindness, whereas the multi-view task reverses to overcounting due to cross-view identity confusion. Chain-of-Thought (CoT) prompting yields near-zero overall benefit ($Δ= -0.16\%$) and its effect on Global Recovery is strongly capability-gated, suggesting that the bottleneck lies in cross-view spatial representation rather than reasoning strategy. These findings reveal fundamental scalability limitations in current MLLMs and position TriViewBench as a controlled diagnostic framework for analyzing structural reasoning failures.

Problem

• Multimodal Large Language Models
• Structural Reasoning
• Controlled Complexity
• Multi-View Reasoning
• Scalability

Innovation

• controlled complexity
• multi-view reasoning
• structural reasoning
• occlusion handling
• spatial representation