Visual Grounding First, Multimodal In-context Learning Follows.

arXiv:2610.30592v1 Announce Type: cross Abstract: LLMs have shown strong in-context learning (ICL) capability, but extending it to Multimodal LLMs (MLLMs) remains challenging. Prior multimodal ICL methods often rely on ICL-specific datasets for additional trai