Meta's AVTR Teaches Multimodal Models Visual Imagination

Meta, KAUST, and Princeton introduced AVTR, an adaptive visual imagination routing mechanism for VLMs that boosts average scores by 4.5 points across 16 benchmarks while cutting latency by 33%.

2026-11-19 ~ 2026-11-21 · 2 related posts