Learned Cross-Task Relationships in Multi-Task Models

Standard multi-task learning (MTL) models typically output marginal probabilities for individual tasks, effectively discarding the joint structural information present in the training data. The authors investigate whether explicitly modeling these cross-task relationships—specifically the joint distribution of task labels—can improve both the shared representation and the performance of downstream reinforcement learning (RL) policies in large-scale recommender systems.

Instead of attempting to model the full, intractable joint distribution of all tasks, the authors propose a targeted approximation using pairwise relationships. They introduce auxiliary heads into existing multi-task architectures to predict cross-labels (e.g., the probability of both a click and a save) or unconditional versions of conditional tasks. These auxiliary heads serve two purposes: they act as a form of covariance estimation and provide a focused gradient signal that forces the shared representation to encode discriminative features specific to task interactions. The authors validate this via a two-phase workflow: first, measuring transfer learning benefits by observing performance gains in existing tasks, and second, using these new predictions as inputs to an RL policy to improve decision-making.

Deployments across YouTube’s Notifications, Homepage, and Watch Next surfaces demonstrate that this framework consistently improves user engagement and satisfaction metrics. The authors find that connecting these auxiliary heads directly to the shared base layer of the model is more effective than routing them through task-specific towers, as it prevents gradient dilution. Furthermore, they show that these improvements are not merely the result of increased loss weighting, but rather the result of providing the model with explicit information about the covariance structure between tasks.