The authors studied continued pre-training, an additional training stage using recorded gameplay, for agents that perceive game images and issue keyboard and mouse controls.
Scaling gameplay training improves familiar worlds more than distant ones
Across 120 training configurations, Zhang and colleagues find that diverse gameplay helps agents transfer to related games, but offers limited gains in substantially different worlds.
Top University
Kuan Zhang · Yukun Chen · Zhihao Yang · Yue Su · Xiangnan Wu · Zirong Chen · +5 more
miHoYo Honkai AI R&D Team · Tsinghua University · University of Chinese Academy of Sciences · The University of Hong Kong · National University of Singapore
Research Digest··2 min read
Zhang and colleagues tested whether more gameplay data, broader game coverage and larger models make virtual agents more general.
§