From Interaction to Capability: Task Advancement as State and Supervision for GUI Agents
Tracking verified progress lifted GUI-agent success on MobileWorld from 12.0% to 28.2%, even before adding learned operations.
Task Advancement Memory (TAM) records how completed steps change an agent’s remaining goals across applications. After each task, it extracts reusable operations tied to required inputs, expected outcomes and verification conditions—even from locally successful steps in failed tasks. The authors report that full TAM improved task success by 17.6% relative to the strongest evaluated baseline while reducing lifecycle tokens per solved task by 38.7%. On SPA-Bench, it achieved state-of-the-art task success in the single-app English Level 3 setting.
ArXiv · AI/CL/LG's note
Task Advancement Memory (TAM) records how completed steps change an agent’s remaining goals across applications. After each task, it extracts reusable operations tied to required inputs, expected outcomes and verification conditions—even from locally successful steps in failed tasks. The authors report that full TAM improved task success by 17.6% relative to the strongest evaluated baseline while reducing lifecycle tokens per solved task by 38.7%. On SPA-Bench, it achieved state-of-the-art task success in the single-app English Level 3 setting.
ArXiv · AI/CL/LG's note
score 5