From Hallucination to Grounding: Diagnosing Visual Spatial Intelligence via CRISP

2026-06-24

AI Summary

Current evaluations of vision-language models struggle to disentangle linguistic priors from genuine spatial reasoning capabilities, leading to ambiguous assessments of visual-spatial intelligence. This work proposes CRISP, a novel evaluation paradigm that leverages consistency metrics to explicitly measure the alignment between implicit perception and explicit reasoning. By integrating 3D scene graphs with an Oracle intervention protocol, CRISP decouples perceptual bottlenecks from reasoning capacity, enabling fine-grained diagnostic insights. The framework reveals for the first time that closed-source models possess strong reasoning abilities yet suffer from metric estimation biases and underutilization of implicit structural cues, whereas open-source models are primarily limited by multi-hop compositional reasoning. The study systematically uncovers a pervasive perception-reasoning disconnect in current models, offering a new pathway toward multimodal alignment beyond end-to-end training.

Abstract

Current VLM evaluations often conflate language priors with genuine spatial reasoning. To address this, we introduce CRISP, a novel structural-diagnostic evaluation paradigm that assesses visual spatial intelligence through consistency, the alignment between implicit perception and explicit reasoning. Unlike traditional black-box QA, CRISP utilizes metric 3D Scene Graphs and an oracle intervention protocol to decouple latent reasoning capabilities from perceptual bottlenecks. This granular diagnosis uncovers a systematic perception-reasoning disconnect. Crucially, we reveal that while proprietary models possess robust latent reasoning engines, they suffer from inaccurate metric estimation and a critical failure to leverage their implicit structural representations. Conversely, open-source models remain fundamentally bottlenecked by their lack of multi-hop compositional reasoning. By shifting the focus from merely ``guessing correctly'' via language priors to genuinely ``perceiving, verifying, and reasoning,'' CRISP offers a rigorous roadmap for multimodal alignment beyond end-to-end post-training. The code and dataset are available at https://github.com/iiyamayuki/CRISP-Bench.

Problem

• visual spatial intelligence
• language priors
• spatial reasoning
• multimodal alignment
• perception-reasoning disconnect

Innovation

• CRISP
• visual spatial intelligence
• 3D Scene Graphs
• oracle intervention
• multimodal alignment