Task Alignment as a Data-Level Predictor of Inter-Task Dynamics in Linear MTL

Multi-task learning (MTL) learns several tasks jointly through shared representations, but negative interference between tasks can degrade performance. Methods designed to mitigate such interference introduce substantial computational overhead and are not consistently beneficial. We therefore investigate whether task interactions can instead be predicted from data-level properties available before training. By adapting the Riccati-based framework for linear network dynamics to the multi-task setting, we obtain exact learning dynamics and show how task alignment and magnitude imbalance govern task interactions. We further find empirically that commonly used training-dependent measures of task interference reflect this underlying data-level structure. This suggests that task alignment can help identify settings where costly mitigation strategies are unnecessary, while providing a basis for data-informed task weighting and grouping.