A scheduling method borrowed from graph theory may make multi-task AI training significantly less self-defeating.
Researchers introduced SON-GOKU, a training scheduler that measures how much different learning objectives pull a model in conflicting directions. It builds an interference graph from those measurements, then uses greedy graph coloring to partition tasks into groups that point roughly the same way. At each training step, only one color class of tasks runs - the partition recalculates continuously as task relationships shift. The approach was tested across six datasets and beat both standard baselines and current state-of-the-art multi-task optimizers.
Gradient conflict is a known, stubborn problem in multi-task learning: the more objectives you stack, the more they can cancel each other out, wasting compute and capping model quality. SON-GOKU sidesteps the usual fix - per-task loss weighting or gradient surgery - by never letting conflicting tasks share a mini-batch in the first place. That means it can drop in under any existing optimizer without retuning.
The graph coloring framing is clever, but the real test is scale: the paper reports results on six datasets, and whether the dynamic re-partitioning overhead stays manageable on large models with dozens of tasks remains an open question.