Discarded or Unreachable? Diagnosing and Repairing Spatial Relation Failure in Vision--Language Similarity
Read the original on arXiv AI →The paper investigates why contrastive vision‑language models often fail to capture left/right spatial relations, identifying two distinct failure modes: an access failure where the relation exists in the pooled embedding but is unreachable by native similarity, and a pooling failure where the relation is lost during pooling but retained in dense features. The authors propose a zero‑parameter, query‑conditioned spatial moment that recomputes geometry from the patch grid and augments the native embedding, dramatically improving left/right pairwise accuracy and recall on large image indexes while preserving native rankings for non‑relational queries.
Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv AI.