SkillGraph Framework Enables Self-Evolving Multi-Agent Collaboration with Dynamic Topology

A new research paper introduces SkillGraph, a framework addressing limitations in scaling vision-language models into Visual Multiagent Systems (VMAS). The approach tackles two interconnected problems: fixed communication topologies that ignore visual content and query context, and static agent reasoning abilities during deployment. These issues reinforce each other—rigid topologies can't utilize richer agent expertise, while static agents lack motivation to specialize for specific queries. SkillGraph jointly evolves both agent expertise and communication topology. The framework employs a Multimodal Graph Transformer (MMGT) that encodes visual tokens, instruction semantics, and active skill embeddings to predict query-conditioned collaboration graphs. This replaces hand-crafted routing with dynamic, content-aware information flow. Additionally, a Skill Designer component distills and refines reasoning heuristics from failure cases, creating a self-evolving system. The research was published on arXiv with identifier 2604.17503v1, categorized as a new announcement. The work focuses on overcoming current bottlenecks in multi-agent visual systems by enabling adaptive collaboration structures that respond to both visual inputs and task requirements.