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.

In the upcoming sixth-generation (6G) communication system, to guarantee the realtime performance and high reliability of intelligent service, artificial intelligence (AI) is indispensable. Accurate prediction of cell traffic is essential in networking, such as energy-saving, load-balancing, handover, and ensuring the quality of service (QoS) for 6G everyone-centric customized services. In the scenario of cell traffic, inter-cell correlations, such as distances and handover operations between cells can directly affect the intra-cell traffic load concurrently. In this work, we proposed a multi-hierarchical spatial-temporal graph convolution network (MH-STGCN) to predict cell traffic indicators. The hierarchical graphs explicitly import different inter-cell correlations to improve the performance by associating these correlations with intra-cell indicators and multiple graphs allow to consider different correlations simultaneously. Therefore, MH-STGCN is capable to import correlations from multiple sources simultaneously. Experiments show that MH-STGCN is more reliable and improves MAE by 8.77% and RMSE by 26.14% on average from MTGNN. MTGNN outperforms the other baseline methods conducted on our dataset.

The global positioning system interferometric reflectometry (GPS-IR) method has the advantage of acquiring observations continuously in all weather conditions, which has great application potential in vegetation remote sensing. However, L-band electromagnetic wave signals are susceptible to various environmental factors, resulting in deviations in GPS-IR observation data at certain times. The accuracy and reliability of the vegetation index retrieving results are challenging to achieve by using single-frequency GPS data. This letter proposed a novel method to combine dual-frequency data for retrieving normalized difference vegetation index (NDVI). We integrated the multipath observations based on the theory of information entropy. Subsequently, a unary linear regression model was established to retrieve the NDVI from the calculated normalized microwave reflection index (NMRI). For validation purposes, the comparative analysis was conducted between the proposed model and the previous single-frequency NDVI retrieving model in terms of retrieval accuracy, based on the continuous observation data acquired by four GPS reference stations in the past five years. The experimental results indicated that the proposed model is available for retrieving the NDVI, with the correlation coefficient of 0.749–0.815 and the root mean square error (RMSE) of 0.056–0.081. Compared with the results acquired by the previous single-frequency NDVI retrieving model, the correlation coefficient of the retrieved NDVI was increased by an average of 18.5%, and the RMSE was reduced by 30.3%. The proposed method in this letter helps further improve the accuracy and continuity of NDVI observation data in some local areas, which contributes to grasping the growth status of vegetation comprehensively.