Skip to content
arXiv Game AI· Runxuan Liu·· 2026-10-15AI score27

EvolvingBrain: An adaptive skill system for multi-modal agents self-evolution in open worlds

EvolvingBrain: An Adaptive Skill System for Multi-modal Agents Self-Evolution in Open Worlds

AI summary

EvolvingBrain proposes an adaptive skill system as the core of self-evolution for multimodal agents in open worlds. Its hierarchical skill self-evolution extracts procedural and problem patterns from the model's own trajectories, forming a hierarchical skill architecture with Domain, Plan, and Action/Trap Skills. Experiments on VAB-Minecraft and DiscoveryWorld show that the method improves task success rates in open-world tasks and demonstrates generalization ability.

Source: arXiv Game AI · arxiv.org