Multi-task learning: shared representation trained jointly on multiple objectives captures more signal than separate models, especially when one task has sparse labels. Hard parameter sharing: shared layers + task-specific heads, single forward pass. Seesaw problem: conflicting objectives cause training instability. MMOE (Multi-gate Mixture of Experts) mitigates seesaw by letting tasks weight different expert sub-networks. At serving: combine task scores via weighted product or linear blend tuned by business policy.