EvolvingBrain Gives Multimodal Agents an Adaptive Skill System for Open-World Self-Evolution

Summary

The paper introduces EvolvingBrain, an adaptive skill system designed to address the limited self-evolution of multimodal agents in open worlds. The authors argue that supervised fine-tuning mainly supports imitation and outcome-based reinforcement learning underuses information from intermediate steps. EvolvingBrain's Stratified Skill Self-Evolution module extracts procedural and problematic patterns from model-specific trajectories, turns experience into skills, and organizes them into Domain, Plan, and Action/Trap levels. Its State-Aware Adaptive Execution module compresses long interactions into Structured Abstractive Memory and uses a state-aware skill-fusion strategy to select knowledge relevant to the agent's current situation. The resulting skill repository is intended to support repeated evolution through interactions with an open world. Experiments on VAB-Minecraft and DiscoveryWorld report higher success rates and generalization on open-world tasks. The abstract says the code is included in the supplementary materials and is planned for public release in the camera-ready version.