MAGE: Multi-Agent Graph-Guided Evolution for Language Models

A new framework called MAGE, which stands for Multi-Agent Graph-guided Evolution, has been introduced to create language-model agents that can evolve on their own. This system incorporates a co-evolutionary knowledge graph with four subgraphs, including one focused on experiences. This experience subgraph stores both corrections from teachers and the learner's previous successful reasoning paths, giving specific guidance for tasks. During its evolution, the graph is updated alongside a task-level search bandit and a skill-level routing bandit, all benefiting from a common reward stream, while the learner's main structure stays unchanged. A structural analysis in the paper shows how this approach maintains a static weak backbone during inference, addressing flaws seen in existing models. The research can be found on arXiv under the identifier 2605.10064.