COLIPRI is generally superior to concurrent methods across all tasks. This is particularly clear when plugging an MLLM on top of our vision backbone. Our models are particularly stronger at clinical metrics, which are most relevant in practice.
🧵8/12
