ECHOES: Seeing Beyond the Current Image via Recursive History Compression and Privileged On-Policy Distillation

Explain Like I'm Normal

Researchers developed ECHOES to help AI agents remember long sequences of GUI actions without clogging their brains with massive image files. By using a 'teacher' model that can see everything to train a 'student' model that summarizes efficiently, they've created a way for agents to act on long-term tasks without high inference costs. It essentially bridges the gap between seeing every screenshot and just keeping a smart text-based diary of what happened.