15/08/2026
Interesting contextual Vision AI research. Vision Language Models are scoring poorly on inference accuracy and one the main challenges for current AI Models is Visual Perception: Humans continue to be much better at visual accuracy perception. This project: ActiveVision is a AI approach trying to get this wide accuracy gap closer, and it is an eye opening for the sometimes excessive optimistic hype around AI. Yes, without proper accuracy scores and better visual reasoning models, current Frontier Models continue to be quite far from any type of General Artificial Intelligence:
ActiveVision
An Exam for Active Observers
Jiarui Zhang*, Muzi Tao*, Shangshang Wang*, Ollie Liu, Xuezhe Ma, Willie Neiswanger University of Southern California · *Equal contribution