Last released Oct 2, 2026
SIMIT-ICL: vision-language models that imagine their own in-context demonstrations at test time