Back to AIBriefs
AnalysisAI Models

Hyper-ICL: Attention Calibration for Multimodal In-Context Learning

Arxiv CS.CV (computer vision)

Proposes Hyper-ICL, using hyperbolic anchor distillation to calibrate attention in multimodal in-context learning. Improves MLLM performance on few-shot tasks without fine-tuning.

·
7 days ago
Hyper-ICL: Attention Calibration for Multimodal In-Context Learning — AIBriefs