A new preprint on arXiv introduces ToolOPD, a credit-aware approach to on-policy distillation (OPD) for tool-use agents. The authors note that standard OPD improves small language agents by providing dense, token-level supervision from a teacher model on trajectories the student itself generates.
The paper argues that standard OPD implicitly assumes something that may not hold in tool-use settings. ToolOPD modifies the distillation process to be credit-aware, though the abstract does not specify the exact mechanism or experimental results.
As an arXiv announcement, the work has not yet been peer-reviewed. The abstract is truncated, so the full details of the proposed method and its evaluation are not available from this source alone.