🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

🚀Meet BLIP3-o, a new #VLM unifying image understanding and generation via CLIP-level feature alignment decoded by a Diffusion head. It is probably the best #GPT-4o type model you can find fully-open source (data, code, ckpt). TRY IT HERE👇

Zhongyang will present R2-T2 at #ICML2025 this Tue afternoon 2:00 - 4:30 pm📍East Exhibition Hall A-B #E-1605. He will also discuss our recent #COLM2025 paper C3PO. Both works reveal the poor routing in #MoE #llm and #vlm and develop test-time re-routing to bridge the gap.

🚨We found a simple but unexplored positional bias of in-context learning: the performance drastically changes if placing demos at different sessions of the prompt, and the best position varies across tasks and models. This analysis indicates a need for position-robust training.