Institution profile

Xiaomi Corporation

Industry researchasia · cn
Official website
Research library313linked papers
Opportunities0open roles
Selected work

Representative Papers

MiMo-V2-Flash Technical Report

Jan 06, 2026arXiv.org

This work proposes a 309B-parameter sparse mixture-of-experts (MoE) language model with only 15B activated parameters per token, designed to enhance reasoning speed, capability, and agent-task performance while reducing computational costs. The architecture integrates sliding-window and global attention mechanisms and introduces a multi-token prediction (MTP) framework alongside a multi-teacher online policy distillation (MOPD) approach to enable efficient training and speculative decoding. Despite using merely one-half to one-third of the activated parameters compared to leading open-source models of similar scale, the proposed model achieves comparable or superior performance, accelerates inference by up to 2.6×, and supports context lengths of up to 3.6 million tokens.

11 citations1 influentialRead paper

WiCross: I Can Know When You Cross Using COTS WiFi Devices

Oct 08, 2023UbiComp/ISWC Adjunct

WiFi-based through-wall detection in smart homes often misclassifies turnaround motions as genuine wall crossings, resulting in high false-alarm rates. To address this, this paper proposes a behavior discrimination method leveraging statistical features of Channel State Information (CSI) phase measurements from commodity WiFi devices. Unlike conventional Received Signal Strength Indicator (RSSI)-based approaches, our method is the first to model the temporal distribution patterns of CSI phase to distinguish crossing from turnaround behaviors, and introduces a lightweight temporal statistical classifier. Experimental evaluation in real-world home environments demonstrates that the proposed method achieves over 95% detection accuracy and less than 5% false-alarm rate, significantly improving robustness and practicality. This work advances contactless, fine-grained human activity sensing by exploiting discriminative CSI phase dynamics, offering a novel and effective solution for reliable through-wall detection.

2 citationsRead paper

Federated Balanced Learning

Jan 20, 2026

This work addresses the issue of global model drift in federated learning caused by imbalanced client data under non-IID settings. To mitigate this, the authors propose a client-side sample balancing mechanism that leverages on-device generative models for knowledge infusion and resampling, achieving data equilibrium under a fixed sample budget. The approach integrates knowledge alignment and knowledge dropout regularization strategies to enhance model generalization. As the first framework to proactively prevent model drift at the client sample level, it is extensible to heterogeneous clients and compatible with diverse federated algorithms. Extensive experiments across multiple real-world, complex scenarios demonstrate its significant superiority over state-of-the-art methods, confirming its effectiveness and robustness.

1 citationsRead paper

Joint Training and Decoding for Multilingual End-to-End Simultaneous Speech Translation

Jun 04, 2023IEEE International Conference on Acoustics, Speech, and Signal Processing

This paper addresses the practical need for one-to-many end-to-end simultaneous speech translation (SimulST) in real-world multilingual scenarios. To this end, we propose the first unified modeling framework supporting joint multilingual training and real-time decoding. Methodologically, we introduce a novel synchronous/asynchronous hybrid training paradigm: asynchronous multilingual pretraining enhances cross-lingual knowledge transfer, while synchronous fine-tuning preserves low-latency constraints; additionally, we design a unified-separate hybrid decoder to balance decoding efficiency and translation quality. We further construct TED-MMST—the first publicly available, multi-party aligned, multilingual end-to-end SimulST benchmark dataset. Experiments demonstrate that our approach achieves superior trade-offs between translation quality (BLEU) and latency (Average Lagging Time, ALAT) on TED-MMST. Both the codebase and the TED-MMST dataset are open-sourced.

1 citationsRead paper
Recent publications

Latest Papers