Self‑evolving methods let tool‑using agents generate their own training data, reducing reliance on human‑annotated trajectories. Existing approaches usually separate trajectory generation from evaluation, using static verifiers that cannot adapt to emerging failure modes or self‑consistency signals, which may amplify shared errors across trajectories. Jointly adapting planning, execution and evaluation can mitigate these issues, but introduces a coordination challenge: each component continuously alters the data or feedback used to train the others. UnifiedPlayers addresses this challenge with a cooperative framework that includes a Planning Player that creates tasks, an Execution Player that produces multi‑turn trajectories with Python tool calls, and an Evaluation Player that builds executable verifiers. Role‑specific rewards are designed and unified under GRPO, guiding the three players toward a common learning objective. Across two model backbones and twelve reasoning benchmarks, UnifiedPlayers improves mathematical reasoning by at least 3.5% and general reasoning by at least 3.9% over the strongest baseline. The learned verifier attains 84.2% adversarial detection accuracy, and its reward signal shows $2.03\times$ higher per‑question variance than a self‑consistency baseline, offering more discriminative verification. These findings highlight cooperation among specialized players as a promising route to self‑enhanced tool‑integrated agents.
Review