Consistent Yet Wrong: Evidence Insensitivity in Spatial Vision-Language Models
For researchers and practitioners in robotics, autonomy, and embodied AI, this work exposes a critical flaw in using consistency as a measure of spatial reasoning in VLMs, providing a diagnostic framework to evaluate beyond accuracy.
The paper reveals that modern vision-language models (VLMs) often produce view-invariant, consistent answers for spatial queries even when those answers are incorrect, indicating weak coupling between predictions and visual evidence. The authors introduce ViewDiag, a multi-view evaluation protocol, and show that models exhibit high prediction stability paired with substantial error, challenging the use of cross-view consistency as a proxy for geometric understanding.
Spatial reasoning is fundamental to robotics, autonomy, and embodied AI, yet modern vision-language models (VLMs) remain unreliable on metric distance queries. A common assumption is that consistent predictions across viewpoints reflect geometric grounding. We test this assumption and find the opposite: leading VLMs often produce view-invariant and consistent answers even when those answers are incorrect, indicating weak coupling between predictions and viewpoint-specific visual evidence. We introduce \textbf{ViewDiag}, a controlled multi-view evaluation protocol built from Hypersim, ScanNet, and KITTI360, comprising 176 object-pair tracks across 80 scenes with 2--10 views per track. The protocol evaluates models along three axes: metric accuracy, distributional concentration, and a latent feature probe for internal collapse that distinguishes decision collapse from representation collapse. Across diverse models, we observe a consistent pattern of high prediction stability paired with substantial error, clustering in a regime characterized by strong consistency but low accuracy. \noindent These results challenge the common use of cross-view consistency as a proxy for geometric understanding. Instead, we show that stable predictions may reflect prior-driven collapse rather than evidence-sensitive reasoning. ViewDiag provides a controlled benchmark and diagnostic framework for evaluating spatial VLMs beyond accuracy alone. The code and data can be found \href{https://github.com/SDivakarBhat/Consistent_Yet_Wrong.git}{here}