Megadose The AI curve, compressed.

PACT: Plan-Action Consistency Training by Preserving Correct Plans and Repairing Actions in GUI Agents

· ArXiv · AI/CL/LG ·
PACT raised AndroidWorld task success from 60.34% to 68.97% against an outcome-only training baseline.

The framework targets GUI agents that reason toward a correct immediate plan but then issue an action that fails to carry it out. It uses annotated reference actions to identify these mismatches, shields verified planning tokens from negative training updates, and favors the reference action while keeping the generated reasoning unchanged. Plan-action consistency on AndroidControl High rose from 83.71% to 89.74%. On the full 369-task OSWorld desktop benchmark, task success increased from 40.92% to 44.17%.

ArXiv · AI/CL/LG's note

score 5

Categories: Research