Unverified50% confidenceFactExact time
At the representation level, VLMs encode spatial relations significantly less well than lexical information
1
Sources
50%
Confidence
Long-term
Relevance
10/12/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedVLMs score far above chance on compositional benchmarks, yet their internal representations do not truly encode spatial relations76% similarUnverifiedHigh accuracy of VLMs on spatial reasoning tasks may mask fundamental deficiencies at the representation level, relying on lexical statistical shortcuts rather than genuine spatial representations74% similarUnverified当前主流LLM在欧几里得空间推理（如判断两条线段是否相交）上的准确率远低于其在文本推理任务上的表现，这是自回归语言模型架构的固有局限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.69% 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67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1002019API
curl https://kongchang.com/api/v1/knowledge/claims/1002019MCP
get_claim(id=1002019)