Skip to content
conftrace_
2026 EMNLP EMNLP 2026

GraSP-VL: Length as a Semantic Granularity Interface for Vision-Language Representations

AbstractFrozen vision-language models already encode object, attribute, relation, and caption information, but a single vector exposes all of these signals through the same access pattern. We ask whether vector length can become a semantic access control: short prefixes answer coarse queries, while longer prefixes add finer distinctions. GraSP-VL learns one shared near-orthogonal transform for image and text embeddings and assigns a fixed, configurable prefix contract. At inference, the transform is applied once and ordinary cosine retrieval uses the prefix required by the task. On a 20,147-example COCO/Flickr30K pool, GraSP-VL reaches a staircase score of 53.01 and hard-negative selectivity of 89.76 with full-space drift below 10-6. Full-dimensional ImageNet-V2 classification and Flickr30K/COCO Karpathy retrieval are unchanged, while held-out human-caption retrieval improves at shorter prefixes. These results show that length can reorganize access to frozen VLM semantics rather than merely compressing them. Code and project page are available at https://lizesheng13.github.io/Grasp-VL/.

More from EMNLP 2026

All EMNLP 2026 papers →