When Privileged Guidance Misaligns: State-Matched Routing and Contextualized Self-Distillation for Multi-Turn Agents
The paper’s core claim is that privileged teacher guidance can hurt when the agent’s actual state no longer matches the reference path.
SMRC-SD checks each turn against the successful trajectory and distills only when the student is in a supported state. It also builds teacher context around the state the student actually reached, rather than treating the full reference path as universally valid. In ALFWorld and WebShop, the method beats unconditional full-path distillation, including gains from 0.746 to 0.865 and 0.574 to 0.693 with Qwen3-1.7B. HF Daily Papers' note
SMRC-SD checks each turn against the successful trajectory and distills only when the student is in a supported state. It also builds teacher context around the state the student actually reached, rather than treating the full reference path as universally valid. In ALFWorld and WebShop, the method beats unconditional full-path distillation, including gains from 0.746 to 0.865 and 0.574 to 0.693 with Qwen3-1.7B. HF Daily Papers' note
score 4