Game-diverse AI training beats Minecraft-only models, preprint finds

A research team that includes members of miHoYo's Honkai AI R&D Team, Tsinghua University, the University of Chinese Academy of Sciences, the University of Hong Kong, the National University of Singapore and Peking University published a preprint on October 20 examining how AI agents trained on games transfer to other games, according to automaton-media.com.

The study, titled “Scaling in Games: Continued Pre-Training For Embodied Agents In Diverse Virtual Worlds,” used Alibaba's multimodal AI Qwen3.5 and continued pre-training on gameplay data spanning 50 to 1600 hours across 79 game datasets. The team ran 120 controlled conditions, varying model size, data amount and the share of Minecraft data across five settings.

Models trained on Minecraft alone stopped improving or got worse on combined validation loss, and scored lowest of the five conditions on benchmarks including ViZDoom. Models trained on a diverse set of games reached an average OGA score about 2.8 times that of Minecraft-only models at 1600 hours of training data.

Performance also improved on Doom, which made up only a small part of the training data, while models failed to consistently beat the pre-training baseline on 2D and 2.5D GameWorld benchmarks. The co-first authors include Kuan Zhang, Yukun Chen and Zhihao Yang.

Real-time tests

In real-time tests, the team's model outperformed general-purpose AIs such as GPT-5.6 Sol, Claude Opus 5.0 and Gemini-3.8-Flash on several games including Minecraft, though the general AIs scored higher under non-real-time conditions.

Earlier work on continued pre-training for embodied agents changed data amount, diversity and model size at the same time, leaving the effect of each factor unclear, the research team said.