Viewpoint-aware spatial reasoning is a persistent bottleneck with poor calibration of spatial uncertainty¶

Claim · All claims
Evidence · 1 study · design r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q2 i? On a Relative Position failure case, 90.0% of models predict a definite lateral relation and 73.3% predict a definite vertical relation despite ground truth specifying none; only 43.3% correctly identify depth, and no model resolves all three dimensions correctly. → Luyao Zhu 2026

Evidence¶

Luyao Zhu 2026¶

Luyao Zhu, Xun Wei Yee, Wei Li, Mun Thye Mak, Wee Siong Ng. (2026). MUSE: Benchmarking Large Vision-Language Models on Multi-modal Understanding in Situated Education. arXiv preprint. https://arxiv.org/abs/2609.19088 doi:10.48550/arxiv.2609.19088

q2 · i? · design · r2

Case-based error analysis of Relative Position responses across the evaluated models (Figure 10b). The article reports a strong forced-relation bias and states that models oscillate between asserting definite relations under ambiguous evidence and predicting None while missing valid depth cues, revealing weak viewpoint-aware spatial reasoning.

"Although the ground truth specifies no definite lateral or vertical relation, 90.0% and 73.3% of models, respectively, predict one; depth reasoning is also unreliable, with only 43.3% correctly identifying the girl as in front of the boy."

Discussion¶

Related Claims¶

• Ground-truth validation found the detector identified 55% of human-coded interactions, with misses driven by visual constraints such as teachers outside boundary boxes or obscured — related
• Generative AI generates responses based on probability and pattern prediction, not reasoning or understanding — a broader claim this one bears on
• LLaMA's LoRA adaptation compresses its prediction range to roughly 2.0–4.0, while Gemma maintains a broader distribution closer to the ground truth — related
• Reasoning with ground-truth auxiliary diagrams consistently outperforms reasoning from the original diagram alone on GeoVAD-Bench — related
• All evaluated models struggle with spatial reasoning and precise visual replication on Turtle Graphics tasks — a narrower finding that bears on this claim
• Trust and reliance can diverge, and self-reported reliance measures cannot classify calibration without output-quality ground truth — related