Seonglae ChoIntroduces a variable-width architecture for decoder-only transformers, keeping the early and late layers wide while making the middle layers narrower.
Variable-Width Transformers do not change width during inference; the width simply differs by layer. Residual connections use parameter-free carry-forward residual resizing, which copies inactive coordinates so dimensional changes work smoothly without learnable projection layers.
Referring Expression Instance Retrieval and A Strong End-to-End Baseline
Using natural language to query visual information is a fundamental need in real-world applications. Text-Image Retrieval (TIR) retrieves a target image from a gallery based on an image-level...
https://arxiv.org/abs/2506.18246
