Yuhang Chen, Zhen Tan, Ajay Kumar Jaiswal, Huaizhi Qu, Xinyu Zhao, Qi Lin, Yu Cheng, Andrew Kwong, Zhichao Cao, Tianlong Chen. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Holographic displays are gaining increasing popularity, particularly as a holy grail solution to augmented and virtual reality (AR/VR) wearable displays. However, the generation of holograms is computationally intensive for AR/VR edge devices, which are expected to be compact and lightweight with low power consumption and heat dissipation to last all day. To address this, we propose a distributed hologram generation framework, dubbed HoloZip, to jointly generate, compress, transmit, and decode computer-generated holograms across the cloud-edge devices. Specifically, we perform compute intensive hologram generation via a vision transformer based backbone on the cloud, compress and transmit before decoding on a local edge device by a lightweight model. Our HoloZip framework supports both both 2D/3D holograms as well as holographic videos, achieving a reconstruction quality of over 30 dB in PSNR (lossy compression standard) with bit rates as low as 0.6 bits per pixel as validated by experiments.

With advancements in reasoning capabilities, Large Language Models (LLMs) are increasingly employed for automated judgment tasks. While LLMs-as-Judges offer promise in automating evaluations, current approaches often rely on simplistic aggregation methods (e.g., majority voting), which can fail even when individual agents provide correct answers. To address this, we propose a multi-agent debate judge framework where agents collaboratively reason and iteratively refine their responses. We formalize the debate process mathematically, analyzing agent interactions and proving that debate amplifies correctness compared to static ensembles. To enhance efficiency, we introduce a stability detection mechanism that models judge consensus dynamics via a time-varying Beta-Binomial mixture, with adaptive stopping based on distributional similarity (Kolmogorov-Smirnov test). This mechanism models the judges' collective correct rate dynamics using a time-varying mixture of Beta-Binomial distributions and employs an adaptive stopping criterion based on distributional similarity (Kolmogorov-Smirnov statistic). Experiments across multiple benchmarks and models demonstrate that our framework improves judgment accuracy over majority voting while maintaining computational efficiency.

Over the past few years, deep learning has been rapidly adopted in many fields. Among the various hardware accelerators specifically for deep learning computation, graphics processing units (GPUs) are mainly used. GPU occupancy—the average ratio of active warps to maximum supported warps on all streaming multiprocessors—is an essential indicator of how well GPUs are utilized. Predicting the GPU occupancy of deep learning models is critical for boosting both job runtime performance and platform resource efficiency. However, GPU occupancy prediction is challenging due to the complex factors hidden in framework runtimes and diverse architectures and hyperparameters of models. In this paper, we propose DNN-occu to predict the GPU occupancy of deep learning models. Our key observation is that models can be represented as directed acyclic computation graphs. DNN-occu extracts a set of occupancy-related features from the computational semantics of the graph nodes and edges. It also employs a novel graph neural network for better feature encoding and prediction generalization. The experiments on various configurations of real-world deep learning models show that DNN-occu achieves high accuracy for occupancy prediction (with an overall error of 9.271%) and has a strong generalization ability for unseen models. In addition, we apply DNN-occu in a trace-driven simulation of deep learning workload scheduling and achieve up to a 31.45% increase in overall GPU utilization and a 19.71% reduction in makespan.