Jin Yang

Quzhe Huang, Zhenwei An, Nan Zhuang et al. · pku

In this paper, we introduce a novel dynamic expert selection framework for Mixture of Experts (MoE) models, aiming to enhance computational efficiency and model performance by adjusting the number of activated experts based on input difficulty. Unlike traditional MoE approaches that rely on fixed Top-K routing, which activates a predetermined number of experts regardless of the input's complexity, our method dynamically selects experts based on the confidence level in expert selection for each input. This allows for a more efficient utilization of computational resources, activating more experts for complex tasks requiring advanced reasoning and fewer for simpler tasks. Through extensive evaluations, our dynamic routing method demonstrates substantial improvements over conventional Top-2 routing across various benchmarks, achieving an average improvement of 0.7% with less than 90% activated parameters. Further analysis shows our model dispatches more experts to tasks requiring complex reasoning skills, like BBH, confirming its ability to dynamically allocate computational resources in alignment with the input's complexity. Our findings also highlight a variation in the number of experts needed across different layers of the transformer model, offering insights into the potential for designing heterogeneous MoE frameworks. The code and models are available at https://github.com/ZhenweiAn/Dynamic_MoE.

Josep Crego, Jungi Kim, Guillaume Klein et al.

Since the first online demonstration of Neural Machine Translation (NMT) by LISA, NMT development has recently moved from laboratory to production systems as demonstrated by several entities announcing roll-out of NMT engines to replace their existing technologies. NMT systems have a large number of training configurations and the training process of such systems is usually very long, often a few weeks, so role of experimentation is critical and important to share. In this work, we present our approach to production-ready systems simultaneously with release of online demonstrators covering a large variety of languages (12 languages, for 32 language pairs). We explore different practical choices: an efficient and evolutive open-source framework; data preparation; network architecture; additional implemented features; tuning for production; etc. We discuss about evaluation methodology, present our first findings and we finally outline further work. Our ultimate goal is to share our expertise to build competitive production systems for "generic" translation. We aim at contributing to set up a collaborative framework to speed-up adoption of the technology, foster further research efforts and enable the delivery and adoption to/by industry of use-case specific engines integrated in real production workflows. Mastering of the technology would allow us to build translation engines suited for particular needs, outperforming current simplest/uniform systems.

Letian Kang, Shixian Luo, Yiqiang Li et al.

Vision-Language Models (VLMs) have demonstrated strong capabilities in multimodal understanding and generation tasks. However, their application to long video understanding remains hindered by the quadratic complexity of standard attention mechanisms. In this work, we introduce \textbf{PEVLM}, a fine-tuning-free parallel encoding method designed to enhance the prefilling efficiency of VLMs in long video scenarios. PEVLM partitions the input video into context blocks with a shared sink block, while preserving sequential position embeddings to align the attention weight distribution with that of Full-Attention. This design reduces attention complexity from $O((T \times N)^2)$ to $O(T \times N)$ where $T$ is the number of frames and $N$ the number of tokens per frame, without sacrificing accuracy. Extensive experiments across multiple state-of-the-art models and benchmarks demonstrate that PEVLM consistently outperforms existing parallel encoding approaches, achieving up to \textbf{7.47x} speedup in attention computation and reducing end-to-end latency by \textbf{40\%}. Remarkably, PEVLM not only maintains high accuracy, but in some settings even surpasses Full-Attention performance. Under strict latency constraints, it achieves substantial gains, improving accuracy from \textbf{23.26\%} to \textbf{61.03\%}. These results underscore the effectiveness of PEVLM for low-latency, long-context video understanding, making it a promising solution for real-world applications.