TY  - RPRT
TI  - Cosine Misleads: Auxiliary Losses Reshape Vision Language Models, Not Their Latents
AU  - XiuYu Zhang
AU  - Junfeng Fang
AU  - Zhenkai Liang
PY  - 2026
UR  - https://arxiv.org/abs/2606.05753
ID  - 2606.05753
ER  -