Unverified50% confidenceFactExact time
VLMs score far above chance on compositional benchmarks, yet their internal representations do not truly encode spatial relations
1
Sources
50%
Confidence
Long-term
Relevance
10/12/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedHigh accuracy of VLMs on spatial reasoning tasks may mask fundamental deficiencies at the representation level, relying on lexical statistical shortcuts rather than genuine spatial representations78% similarUnverifiedAt the representation level, VLMs encode spatial relations significantly less well than lexical information76% similarUnverifiedIn hard sample detection, VLMs have stronger semantic capabilities but at a high computational cost, while lightweight detection models are fast but struggle with hard samples, forming a trade-off between accuracy and efficiency.75% similarUnverifiedVLM在描述整体场景时表现出色，却在精确定位微小目标时频频出错，这是因为当画面元素密集时单个patch可能包含多个人物导致细节信息在下采样中被稀释或丢失72% similarUnverifiedVLMs have limitations when processing high-precision engineering drawings: insufficient accuracy in recognizing small text, a tendency to hallucinate and fabricate dimensions, and limited understanding of GD&T professional symbols.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1002020API
curl https://kongchang.com/api/v1/knowledge/claims/1002020MCP
get_claim(id=1002020)