Tool-Call Boundary Drift in Multi-Teacher Distillation
July 14, 2026
Research into multi-teacher on-policy distillation shows that specializing teachers for tool calls and direct responses can induce an invisible behavior shift toward over-calling tools. This drift occurs because aggregate loss metrics fail to capture the divergence in full-sequence per-token distribution.
HOW THIS AFFECTS YOU
●
builderYou should monitor per-token divergence during distillation to prevent your agents from calling tools unnecessarily.
●
researcherYou can adjust your distillation strategies to account for tool-call recall versus over-calling biases.