UnifiedPlayers: Enhance Tool-Integrated Reasoning in Agentic Reinforcement Learning
The paper’s claim is that training planning, execution, and verification together beats static or self-consistency-based evaluation for tool-using agents.
UnifiedPlayers splits the agent loop into three cooperative “players”: one makes tasks, one produces multi-turn Python-tool trajectories, and one builds executable verifiers. The authors say role-specific rewards under GRPO keep those moving parts aligned despite each changing the others’ training signal. Across two model backbones and twelve reasoning benchmarks, it tops the strongest prior baseline by at least 3.5% on math reasoning and 3.9% on general reasoning. Its learned verifier reports 84.2% adversarial detection accuracy and a more discriminating reward signal than self-consistency. ArXiv · AI/CL/LG's note
UnifiedPlayers splits the agent loop into three cooperative “players”: one makes tasks, one produces multi-turn Python-tool trajectories, and one builds executable verifiers. The authors say role-specific rewards under GRPO keep those moving parts aligned despite each changing the others’ training signal. Across two model backbones and twelve reasoning benchmarks, it tops the strongest prior baseline by at least 3.5% on math reasoning and 3.9% on general reasoning. Its learned verifier reports 84.2% adversarial detection accuracy and a more discriminating reward signal than self-consistency. ArXiv · AI/CL/LG's note
score 5