Search papers, labs, and topics across Lattice.
DAMODantong Zhu4 Jun Yang5 Conghui He2 Weijia Li2LiblibAIRUCShanghai Jiaotong UniversityJun 23, 2026arXiv:2606.24133This paper introduces the Holistic Data Scheduler (HDS), an innovative online data mixing framework for Large Language Model (LLM) pre-training that employs multi-objective reinforcement learning to optimize data composition dynamically. By leveraging the Soft Actor-Critic algorithm, HDS formulates the data scheduling challenge as a continuous control problem, integrating a novel holistic reward function that balances data quality, inter-domain influence, and model weight norms. Experimental results demonstrate that HDS achieves comparable performance to existing methods with 44% fewer training iterations and a 7.2% improvement on the MMLU 0-shot task, highlighting its potential to enhance both efficiency and model capability.
HDS achieves 44% faster training iterations while improving model performance, redefining efficiency in LLM pre-training.
The composition of training data, governed by the diversity of sources and their mixing strategy, is a cornerstone of Large Language Model (LLM) pre-training. Online Data Mixing (ODM), the technique of adaptively adjusting data mixtures during training, has emerged as a promising direction to improve efficiency. However, existing methods are constrained by their reliance on a singular optimization perspective, which fundamentally overlooks the need for complex LLM pre-training to consider the dynamic data composition from multiple dimensions. To overcome this limitation, we introduce the Holistic Data Scheduler (HDS), a novel online data mixing framework. HDS formulates the data scheduling challenge as a reinforcement learning problem in a continuous control space and leverages the Soft Actor-Critic (SAC) algorithm for its stability and sample efficiency in exploring the high-dimensional policy space. At the core of HDS lies a novel multi-objective, holistic reward function that integrates three critical perspectives: a data-driven reward for quality, a loss-driven reward capturing inter-domain influence, and a model-driven reward based on weight norms. To validate our design and determine its optimal configuration, we conducted systematic experiments on LLMs of various sizes. On The Pile benchmark, HDS reaches the final validation perplexity of the next best method with 44% fewer training iterations. Furthermore, it achieves a 7.2% improvement on the MMLU 0-shot task along with consistent gains on other benchmarks, showcasing its ability to enhance both training efficiency and final model capability.
