State-Matched Routing and Contextualized Self-Distillation for Multi-Turn Agents

A new arXiv paper (2608.05219) introduces State-Matched Routing and Contextualized Self-Distillation (SMRC-SD), a method to improve privileged on-policy distillation for multi-turn agents. The authors identify a state-reference mismatch problem: when a student agent's actions deviate from the reference trajectory, the privileged guidance becomes unreliable. SMRC-SD explicitly determines when and how to apply privileged reference guidance, ensuring compatibility with the student's current execution state. The paper is announced as a new submission on arXiv, with the abstract outlining the motivation and the proposed solution. The work addresses a core challenge in interactive environments where student rollouts may reach states not covered by the reference, making indiscriminate distillation harmful. The method aims to provide dense supervision while maintaining alignment with the actual state, potentially improving training efficiency and performance for multi-turn agents. The paper does not specify authors, affiliations, or experimental results in the provided content, but it is a technical contribution to the field of AI and machine learning, specifically in the area of reinforcement learning and agent training.