miHoYo scales embodied agents in virtual worlds: 120 configs map limits of continued pre-training
THUSI-Lab · hf · 2026-10-22
miHoYo's HonKai AI team with Tsinghua, HKU and others present a controlled scaling study of continued pre-training (CPT) for embodied agents in virtual worlds: 120 training configurations spanning 0.8B–27B models, 50–1,600 data hours, and five mixtures from Minecraft-only to diverse multi-world corpora.
- More data helps in-corpus performance; diversity improves transfer to related worlds; model size boosts data efficiency but generalizes worse to distant worlds under single-world training
- An early interface phase with broad cross-world gains gives way to a later world phase; an empirical scaling law relates validation loss to data volume, diversity and model size
- Models are competitive with Minecraft specialists and frontier models in 3D worlds, best in real-time settings
- Limits of imitation and persistent distant-world gaps suggest generalist embodied agents need richer training stages beyond CPT
More from Embodied
- Eric Jang: robots need open SDKs so the community can secure them at scale — ericjang11 · 2026-10-22
- NUS Paper Proposes PAPAV: A Unified Capability Framework from Computer-Use to Robot-Use Agents — showlab · 2026-10-22
- Cambridge's Rika Antonova on co-evolving robot bodies and brains, Co-design Gym expansion — contactrika · 2026-10-22
- Musk: Optimus V3 demo video in 1-2 months, AI5 chip coming April 2027 — elonmusk · 2026-10-22
- "Mad that this is a $10,000 car": 2nd-gen BYD Seagull stuns observers — alexmacgregor__ · 2026-10-22
- AFM scans compare Apple's AR coatings: iPhone Duo vs iPad nano-texture glass — jwt0625 · 2026-10-22