Addressing Intent Dissonance in Cross-Domain Distillation on YouTube

Modern recommender systems increasingly operate within mixed-interface ecosystems, transitioning users between high-intent browsing (e.g., Homepage) and low-intent browsing (e.g., infinite feeds). While zero-shot distillation offers a cost-efficient solution for some knowledge transfer scenarios, we demonstrate its critical limitations when source and target domains exhibit significant behavioral heterogeneity or "intent gaps". Extending our prior work on zero-shot cross distillation, this paper presents a case study on adapting a large-scale shared teacher optimized for high-intent (long-form) video ranking to low-intent (short-form) video feed. We hypothesize that zero-shot methods fail here due to both format differences, and severe intent dissonance where the active selection signals (clicks) from the teacher do not effectively map to the passive viewing signals (swipes/dwell) required by the student. To overcome this, we propose a domain-adaptive approach implemented on a multi-billion parameter model. First, we address label and covariate shift by enriching the teacher with Shorts signals; second, we tackle the practical challenge of increased distributional variance using an optimized backbone architecture to effectively disentangle these intent signals. The enhanced teacher yields significant gains in offline and online metrics, establishing a blueprint for successful cross-interface distillation in mixed-media platforms.