The authors first used AndroidControl trajectories to test how evaluation affects apparent world-model gains.
Predicted GUI screens add little beyond direct action ranking
Controlled comparisons show that learned direct evaluation can match or exceed one-step world-model lookahead without generating future screens.
Big Tech
Sizhe Tang · Hanhan Zhou · Jeffrey Svajlenko · Rashmi Gangadharaiah
George Washington University · Amazon Web Services
Research Digest··2 min read
Thread:Agent Harness Optimization
Tang et al.
§