The authors propose GeoPID, which operates on the hidden representations of VLMs without additional training.
Geometric decomposition of VLM representations improves visual grounding without retraining
By analyzing redundant, unique, and synergistic information in vision and language subspaces, GeoPID identifies and amplifies vision-unique features at inference.
Big Tech
Seulgi Kim · Zhixiong Zhang · Xinwei Zhang · Jie Ling · Ronn Shaw
Georgia Institute of Technology · Amazon
Research Digest··3 min read
The authors introduce GeoPID, a training-free framework that decomposes multimodal representations in vision-language models (VLMs) into redundant, modality-unique, and synergistic components using geometric relationships between visual and textual subspaces.
Why this paper
From Amazon and Georgia Institute of Technology
In one line
GeoPID decomposes VLM representations into vision-unique components and amplifies them, achieving 7.63% accuracy gain without retraining.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§