UnifiedPlayers trains three roles in one loop: a Planning Player creates tasks, an Execution Player solves them with multi-turn Python tool calls, and an Evaluation Player builds executable verifiers. Role-specific rewards coordinate all three under GRPO, so feedback adapts as the task distribution and trajectories change.
Across two model backbones and twelve reasoning benchmarks it beats the strongest prior baseline by at least 3.5% on math and 3.9% on general tasks. The learned verifier gets 84.2% adversarial detection accuracy, and its reward signal has 2.03x higher per-question variance than self-consistency, meaning more discriminative feedback. The key payoff is avoiding the degenerate loops that can occur when planning, execution, and evaluation drift independently.