PiKE: Adaptive Data Mixing for Large-Scale Multi-Task Learning Under Low Gradient Conflicts

Modern foundation models are trained on diverse datasets to enhance generalization across tasks and domains. A central challenge in this process is determining how to effectively mix and sample data from multiple sources. This naturally leads to a multi-task learning (MTL) perspective.

While prior work in MTL has emphasized mitigating gradient conflicts, we observe that large-scale pretraining scenarios—such as multilingual or multi-domain training—often exhibit little to no gradient conflict. Motivated by this observation, we propose $\text{PiKE}$ ($\text{P}$ositive gradient $\text{i}$nteraction-based $\text{K}$-task weights $\text{E}$stimator), an adaptive data mixing algorithm that dynamically adjusts sampling weights during training. PiKE exploits non-conflicting gradient interactions to minimize a near-tight upper bound on the average loss decrease at each step, while incurring negligible computational overhead.

We provide theoretical convergence guarantees and show that PiKE outperforms static and non-adaptive mixing baselines. Furthermore, we extend PiKE to promote balanced learning across tasks. Extensive experiments on large-scale language model pretraining confirm that PiKE achieves faster convergence and improved downstream performance compared to existing approaches.