a research thread on visual information in multimodal LLMs

My recent work follows one question: how can multimodal language models make better use of the information produced by a visual encoder?

The first study examines visual layer selection and asks where useful visual signals appear inside the encoder. The second studies multi-layer feature fusion, comparing ways to combine complementary information across layers. The third examines connector selection, focusing on what happens when visual information is preserved or compressed before it reaches the language model.

Together, these projects form a connected research thread across visual representation, feature integration, and vision-language alignment. The corresponding papers and resources are available on the publications page.