agent-reasoning

Role-Agent introduces a framework where a single LLM acts as both agent and environment, enabling bootstrapped co-evolution via World-In-Agent and Agent-In-World components. It achieves over 4% average improvement on multiple benchmarks over strong baselines.