What drives attention sinks? A study of massive activations and rotational positional encoding in large vision-language models.

@article{zhang-2026-drives, author = {Xiaofeng Zhang and Yuanchao Zhu and Chaochen Gu and Jiawei Cao and Hao Cheng and Kaijie Wu}, title = {{What drives attention sinks? A study of massive activations and rotational positional encoding in large vision-language models}}, journal = {Information Processing and Management}, volume = {63}, number = {2}, pages = {104431}, year = {2026}, url = {https://doi.org/10.1016/J.IPM.2025.104431}, doi = {https://doi.org/10.1016/J.IPM.2025.104431} }