Role-Agent Framework Enables LLMs to Self-Train Through Dual-Role Simulation

New framework allows single LLM to act as both agent and environment, achieving 4% performance gain over baselines through bootstrapped co-evolution.

Role-Agent Framework Enables LLMs to Self-Train Through Dual-Role Simulation

Researchers have introduced Role-Agent, a framework that enables a single Large Language Model to function simultaneously as both an agent and its training environment, according to a paper published on arxiv.org.

The framework addresses limitations in current LLM agent training, which the researchers describe as “often limited by inefficient interaction feedback and static training environments.” Role-Agent implements this through two components: World-In-Agent (WIA) and Agent-In-World (AIW).

According to the paper, in WIA, “the LLM acts as the agent and predicts future states after each action; the alignment between predicted and actual states is then used as a process reward, encouraging environment-aware reasoning.” In AIW, “the LLM analyzes failure modes from failed trajectories and retrieves tasks with similar failure patterns, thereby reshaping the training data distribution for targeted practice.”

Experiments on multiple benchmarks demonstrated that Role-Agent “consistently improves performance, yielding an average gain of over 4% over strong baselines,” according to arxiv.org. The paper spans 20 pages, including 12 pages of main text and 8 pages of appendix, and is described as “work in progress.” The research was submitted on June 9, 2026, by Xucong Wang and colleagues.

LLM agentsreinforcement learningmodel trainingarxiv