Role-Agent: Bootstrapping LLM Agents via Dual-Role Evolution

The paper introduces Role-Agent, a framework that allows a single Large Language Model (LLM) to operate as both the agent and the environment, facilitating a bootstrapped co-evolution process. It consists of two components: World-In-Agent (WIA), which predicts future states to provide rewards based on alignment with actual states, and Agent-In-World (AIW), which analyzes failure modes to enhance training data distribution. Experimental results demonstrate that Role-Agent achieves an average performance improvement of over 4% on multiple benchmarks compared to strong baselines, highlighting its potential for enhancing generalization in LLM applications.