Isolate the effect of teacher-sourced versus self-rephrased preference pairs

Determine whether teacher-sourced or self-rephrased preferred responses produce different effects in preference optimization for FiMI Banking, since the reported evaluation does not isolate the choice of preferred-response source.

Background

FiMI Banking constructs preference pairs from failures of the base model and considers two sources for the preferred response: a substantially larger teacher model or a self-rephrased revision generated by the policy itself. The paper reports that these constructions differ conceptually, but the evaluation does not attribute the reported checkpoint to either construction separately.

Because the reported results do not isolate the preferred-response source, the relative effect of teacher-sourced and self-rephrased preference data remains unresolved. Establishing this effect would require a controlled comparison in which the two preference-pair constructions are evaluated independently under otherwise matched training and evaluation conditions.

References

Any difference between teacher-sourced and self-rephrased preference pairs therefore remains an open question rather than a measured effect.

— FiMI Banking: A Sovereign Model for Indian Retail Banking  (2609.03960 - Team et al., 3 Sep 2026) in Section Discussion and Analysis, paragraph “Three caveats qualify that reading,” first bullet, labeled “Mechanism evidence”