One AI system trains its own task-maker, doer and grader

UnifiedPlayers trains three AI roles, a task planner, an executor and a verifier, together, according to a paper posted to arXiv on September 20.

That matters for anyone scaling tool-agent training without human labels. A fixed verifier can't keep up as an agent finds new ways to fail. Voting across an agent's own repeated attempts breaks when every attempt shares the same mistake.

The Planning Player writes new tasks. The Execution Player runs multi-turn Python tool calls to attempt them. The Evaluation Player writes executable code that checks whether the attempt worked, rather than relying on a checklist written once and left alone.

Each role gets its own reward signal, tuned to what that role is supposed to get right. Because all three train at once, each one's output becomes training data for the other two. A task the planner invents shapes what the executor practices, and how well the executor performs shapes what the evaluator has to check.

All three train together, so the grader keeps changing right along with the work it grades.