StraTA: Trajectory-Level Strategy Injection Improves Long-Horizon Agent RL

arXiv·medium signal

Strategic Trajectory Abstraction (StraTA) introduces explicit trajectory-level strategy into agentic RL by sampling a compact strategy from the initial task state, conditioning subsequent actions on it, and training strategy generation and action execution jointly with hierarchical GRPO-style rollouts. Addresses the core limitation that current agentic RL methods are purely reactive, weakening both exploration and credit assignment over extended trajectories.