Searcharxiv⌕ Search

arXiv subjects

Menghao Xu

Publications and source records attributed to Menghao Xu.

3 recordsLinked to original sources

TeleChat to Xing4: Advancing Frontier Agentic Models on Domestic Chips and Frameworks

We present Xing 4, a series of large language models that extends the TeleChat lineage through coordinated advances in model architecture and training infrastructure on Ascend. This report describes the design principles and systems engineering behind the series, spanning architecture selection, large-scale pretraining, and reasoning and agentic post-training. We study how sparse activation, attention design, model depth and width, and speculative decoding balance model capacity with computation, memory access, and execution overhead. Our pretraining infrastructure builds on MindSpore HyperParallel, combining topology-aware hybrid sharding, efficient distributed optimizer execution, and coordinated management of long-sequence workloads and activation memory. CPU affinity and core isolation further reduce host scheduling interference. Together, these mechanisms support the pretraining of a 285B-parameter model on 8,192 Ascend 910C cards, providing a concrete setting for the architectural and systems studies. For post-training, we introduce TeleGym, a unified agent execution platform that connects execution, evaluation, and trajectory construction to policy optimization. Disaggregated asynchronous execution, bounded oversampling, and partial rollout accommodate heterogeneous task durations while preserving trajectory continuity across policy updates. Together, these advances establish an Ascend-oriented training foundation for the continued development of Xing 4 across model scales and increasingly demanding reasoning and agent workloads.

cs.AI↗

Non-Gaussian reconciliation for continuous-variable quantum key distribution

Non-Gaussian modulation can improve the performance of continuous-variable quantum key distribution (CV-QKD). For Gaussian modulated coherent state CV-QKD, photon subtraction can realize non-Gaussian modulation, which can be equivalently implemented by non-Gaussian postselection. However, non-Gaussian reconciliation has not been deeply researched, which is one of the key technologies in CV-QKD. In this paper, we propose a non-Gaussian reconciliation method to obtain identical keys from non-Gaussian data. Multidimensional reconciliation and multi-edge type low density parity check codes (MET-LDPC) are used in non-Gaussian reconciliation scheme, where the layered belief propagation decoding algorithm of MET-LDPC codes is used to reduce the decoding complexity. Furthermore, we compare the error correction performance of Gaussian data and non-Gaussian data. The results show that the error correction performance of non-Gaussian data is better than Gaussian data, where the frame error rate can be reduced by 50% for code rate 0.1 at SNR of 0.1554 and the average iteration number can be reduced by 25%.

quant-ph↗

Sufficient variable screening via directional regression with censored response

We in this paper propose a directional regression based approach for ultrahigh dimensional sufficient variable screening with censored responses. The new method is designed in a model-free manner and thus can be adapted to various complex model structures. Under some commonly used assumptions, we show that the proposed method enjoys the sure screening property when the dimension p diverges at an exponential rate of the sample size n. To improve the marginal screening method, the corresponding iterative screening algorithm and stability screening algorithm are further equipped. We demonstrate the effectiveness of the proposed method through simulation studies and a real data analysis.

stat.ME↗