UnifiedPlayers: Enhance Tool-Integrated Reasoning in Agentic Reinforcement Learning

TL;DR - UnifiedPlayers jointly trains task planning, tool-based execution, and verifier generation as cooperating reinforcement-learning players. This coordination improves reasoning performance while producing more adaptive and discriminative evaluation signals.

• Uses specialized Planning, Execution, and Evaluation Players coordinated through role-specific rewards under GRPO.
• Generates tasks, multi-turn trajectories with Python tool calls, and executable verifiers within one self-evolving framework.
• Outperforms the strongest prior baseline by at least 3.5% on mathematical reasoning and 3.9% on general reasoning across two model backbones and 12 benchmarks.
• The learned verifier reaches 84.2% adversarial detection accuracy and yields 2.03× higher per-question reward variance than self-consistency.