GUI Agents +28% Success, 80% Less History: Match Screens Not Tasks
Summary
Exposes a systematic flaw in how personal GUI agents use interaction history—retrieved task-relevant experience helps almost exclusively at step one, leaving ~90% of decisions without sustained benefit—and introduces a training-free fix that cuts context size by 80% while improving per-step success by 28%; directly affects anyone deploying enterprise desktop or mobile agents.
Originally reported by paper
Read the original article →Original headline: GUI Agents +28% Success, 80% Less History: Match Screens Not Tasks