UnifiedPlayers trains the task generator, the executor and the verifier together instead of freezing two of them

arXiv·medium signalXBlueskyLinkedInCopy link

arXiv 2609.20089 (submitted 2026-09-17) attacks a failure mode in self-evolving tool-using agents: static verifiers cannot adapt to new failure modes, and self-consistency signals reinforce errors shared across trajectories. UnifiedPlayers runs a Planning Player that generates tasks, an Execution Player that produces multi-turn Python tool-call trajectories, and an Evaluation Player that writes executable verifiers, with role-specific rewards coordinating all three as each keeps changing the data the others train on. It is a concrete answer to who checks the checker when nobody is labeling trajectories by hand.