RESEARCH · RESEARCH · #2191
ECHOES: privileged on-policy distillation and recursive textual history compression for long-horizon GUI agents
ECHOES is a method that trains GUI agents to compress visual history into a recursively updated textual state using a Compression-Think-Action (CTA) format. During training a multi-image-input teacher provides dense token-level privileged on-policy distillation to a student model; combined with a PMF-Gap-guided OPD-RL framework and periodic teacher updates, ECHOES improves task performance and memory fidelity without retaining screenshots at deployment and introduces the ECHOESBMK diagnostic benchmark.
KEY POINTS
- ECHOES is a method that trains GUI agents to compress visual history into a recursively updated textual state using a Compression-Think-Action (CTA) format.
- During training a multi-image-input teacher provides dense token-level privileged on-policy distillation to a student model; combined with a PMF-Gap-guided OPD-RL framework and periodic teacher updates, ECHOES improves task performance and memory fidelity without retaining screenshots at deployment and introduces the ECHOESBMK diagnostic benchmark.
- Reduces inference cost and improves long-horizon state retention for GUI agents by using training-time privileged inputs to teach compact textual histories, enabling better performance without storing screenshots.
WHY IT MATTERS
Reduces inference cost and improves long-horizon state retention for GUI agents by using training-time privileged inputs to teach compact textual histories, enabling better performance without storing screenshots.