Effect of the distillation objective in the confidence-based trigger

Determine whether the task–distillation objective used by the confidence-based AKD trigger can produce successful transfer when the selector, transfer budget, and other trigger components are held distinct from the objective itself.

Background

The distillation trigger selects queries using disagreement, retrieval-context similarity, and intrinsic parametric confidence, then applies a task–distillation objective to transfer retrieval-conditioned behavior into the parametric model. The registered three-step transfer budget was ablated on one combination, where increasing the budget made performance worse, but the objective itself was not independently removed or replaced.

Consequently, the negative trigger result does not identify whether the task–distillation objective is responsible for the failure. The authors explicitly characterize this objective as still untested, leaving its effect unresolved.

References

The transfer objective is unablated. The trigger's registered transfer budget is 32 queries, one LoRA step per cycle, three cycles; the budget was ablated on one combination (Section~\ref{sec:mech_select}, exploratory) and made the outcome worse, so the R-2 negative is not a negative for this budget alone --- the distillation objective itself remains untested.