cs.AISep 23, 2026
SaveLearned Cross-Task Relationships in Multi-Task Models
Organizations: Google San Bruno, CA, USA
Abstract
We propose a framework that learns cross-task relationships in multi-task models by approximating the joint distribution of task labels through targeted pairwise relationships. This approach improves performance via transfer learning and enhances information extraction without the intractable complexity of modeling the full joint space. Although our framework applies to any multi-task system, we demonstrate its efficacy within YouTube's production recommendation systems. Experiments across the Notifications, Homepage, and Watch Next surfaces show improvements in both accuracy and user satisfaction metrics. Finally, we propose a workflow template to facilitate broader future implementation.
Figures & tables
| Save | No Save | |
|---|---|---|
| Click | 0.05 | 0.15 |
| No Click | 0.05 | 0.75 |
Table 1 . Joint Probability Distribution of Click vs. Save.
| Surface | Hypothesis | Auxiliary Heads | AUC Impact | Empirical Impact [95% CI] |
|---|---|---|---|---|
| Notifications (Sec. 6.1.1 ) | Combined: Transfer Learning + RL Input | 10 heads ( ): Cross-labels between direct actions (Open/Dismissal) and ambient activity (Visits). | 3–8% (Direct ); Neutral (Ambient ) | 0.17% [ 0.01, 0.32] Daily Active Users 0.19% [ 0.02, 0.37] Valued watch time 1.05% [ 1.84, 0.26] Opt-outs 7.21% [ 7.30, 7.11] Sends 7.67% [ 7.34, 7.99] Click-Through Rate |
| Homepage Phase 1 (Sec. 6.2.1 ) | Transfer Learning | 5 unconditioned heads: Watch Time Ratio, Save, Like, Subscribe, Inline Playback. | 0.2–0.5% (Primary: click , subscribe , inline playback , like , save ); Minor regressions (Unrelated: scroll , share ) | 0.05% [ 0.01, 0.09] Valued watch time 0.10% [ 0.05, 0.16] Shorts engagement 0.24% [ 0.42, 0.05] Low quality impressions |
| Phase 2 (Sec. 6.2.1 ) | Transfer Learning | 7 heads: Share (uncond.), Download (cond.), plus 5 heads: Save [Action] conditioned. | 1–2% ( share , download ); 0.5% ( like , dislike ) | 0.05% [ 0.00, 0.10] Valued watch time 0.06% [ 0.00, 0.12] Shorts engagement 0.16% [ 0.30, 0.03] Low quality impressions |
| Phase 3 (Sec. 6.2.2 ) | RL Input | 11 heads: Final refined set of unconditioned, conditioned, and cross-task predictions. | N/A; baseline contains MTL changes | 0.03% [ 0.00, 0.07] Home engagement 0.10% [ 0.05, 0.16] Shorts engagement 0.64% [ 0.86, 0.43] Low quality impressions |
| Watch Next (Sec. 6.3.1 ) | Transfer Learning | 4 unconditioned variants: Continuation, Like, Dislike, Subscribe. | 0.2–2.6% (Original heads) 0.25% (CTR), 0.22% (Bad Watch) | 0.19% [ 0.15, 0.23] Watch Next Valued Watch Time 0.08% [ 0.04, 0.12] Valued watch time 0.42% [ 0.53, 0.31] Low quality impressions |
Table 2 . Summary of Results. Note: Reported metrics may represent proprietary variants of the listed descriptors.