Zheng Zhong

ByteDance Seed, Jiaze Chen, Tiantian Fan et al. · bytedance

We introduce Seed1.5-Thinking, capable of reasoning through thinking before responding, resulting in improved performance on a wide range of benchmarks. Seed1.5-Thinking achieves 86.7 on AIME 2024, 55.0 on Codeforces and 77.3 on GPQA, demonstrating excellent reasoning abilities in STEM and coding. Beyond reasoning tasks, the method demonstrates notable generalization across diverse domains. For instance, it surpasses DeepSeek R1 by 8% in win rate on non-reasoning tasks, indicating its broader applicability. Compared to other state-of-the-art reasoning models, Seed1.5-Thinking is a Mixture-of-Experts (MoE) model with a relatively small size, featuring 20B activated and 200B total parameters. As part of our effort to assess generalized reasoning, we develop two internal benchmarks, BeyondAIME and Codeforces, both of which will be publicly released to support future research. Model trial link: https://www.volcengine.com/experience/ark.

Zhichen Zeng, Chi-Chih Chang, Jiayi Wang et al.

Mixture-of-experts (MoE) architectures enable trillion-parameter LLMs with sparsely activated experts. Expert parallelism (EP) is a widely adopted MoE training strategy, but it suffers from severe all-to-all communication bottlenecks, which is exaggerated by the limited inter-node network bandwidth as the growing model size requires distributing experts across GPU nodes. Prior work focused on overlapping these all-to-all communications with feed-forward network (FFN) and self-attention computations, which often leaves residual network-bound stalls due to inherent imbalance in attention and FFN layers' computation-communication ratios. We present DisagMoE, a disaggregated MoE training system that jointly optimizes model placement and scheduling for maximal efficiency. DisagMoE separates attention and FFN layers into disjoint GPU groups, introduces a multi-stage pipeline with uni-directional, many-to-many communications, and employs a computation-communication roofline model to balance GPU and network bandwidth allocation among the attention and FFN groups. DisagMoE is implemented on Megatron-LM, and evaluation shows that DisagMoE improves training efficiency across multiple MoE models with up to 1.8x speedup on 16-node 8xH800 clusters.

Zezhou Wang, Youjie Li, Zhiqi Lin et al.

Fully Sharded Data Parallel (FSDP), also known as ZeRO, is widely used for training large-scale models, featuring its flexibility and minimal intrusion on model code. However, current FSDP systems struggle with structure-aware training methods (e.g., block-wise quantized training) and with non-element-wise optimizers (e.g., Shampoo and Muon) used in cutting-edge models (e.g., Gemini, Kimi K2). FSDP's fixed element- or row-wise sharding formats conflict with the block-structured computations. In addition, today's implementations fall short in communication and memory efficiency, limiting scaling to tens of thousands of GPUs. We introduce veScale-FSDP, a redesigned FSDP system that couples a flexible sharding format, RaggedShard, with a structure-aware planning algorithm to deliver both flexibility and performance at scale. veScale-FSDP natively supports efficient data placement required by FSDP, empowering block-wise quantization and non-element-wise optimizers. As a result, veScale-FSDP achieves 5~66% higher throughput and 16~30% lower memory usage than existing FSDP systems, while scaling efficiently to tens of thousands of GPUs.

Bohong Wu, Mengzhao Chen, Xiang Luo et al.

Large Language Models (LLMs) are powerful but often too slow and costly for real-world use during inference. Looped transformers save on parameters by reusing the same weights for multiple computational steps, or "loops." However, this approach has a major flaw: the loops run one after another, causing inference latency and memory requirements to increase with each added loop. This makes them impractical for fast applications. To solve this problem, we introduce the Parallel Loop Transformer (PLT). PLT is a new architecture that delivers the performance benefits of a deep, looped model but with the low latency of a standard, non-looped model. PLT works using two key techniques. First, Cross-Loop Parallelism (CLP) breaks the sequential dependency by computing different loops for different tokens at the same time, all within a single pass. Second, to prevent memory costs from growing, we use an Efficient Representation Enhancement strategy. This method shares the memory (KV cache) from the first loop with all other loops. It then uses a Gated Sliding-Window Attention (G-SWA) to combine this shared global information with local information, maintaining high accuracy. Our experiments show that PLT achieves the high accuracy of a traditional looped model but with almost no extra latency or memory cost compared to a standard transformer.

Yichen Jiang, Shikha Bordia, Zheng Zhong et al.

We introduce HoVer (HOppy VERification), a dataset for many-hop evidence extraction and fact verification. It challenges models to extract facts from several Wikipedia articles that are relevant to a claim and classify whether the claim is Supported or Not-Supported by the facts. In HoVer, the claims require evidence to be extracted from as many as four English Wikipedia articles and embody reasoning graphs of diverse shapes. Moreover, most of the 3/4-hop claims are written in multiple sentences, which adds to the complexity of understanding long-range dependency relations such as coreference. We show that the performance of an existing state-of-the-art semantic-matching model degrades significantly on our dataset as the number of reasoning hops increases, hence demonstrating the necessity of many-hop reasoning to achieve strong results. We hope that the introduction of this challenging dataset and the accompanying evaluation task will encourage research in many-hop fact retrieval and information verification. We make the HoVer dataset publicly available at https://hover-nlp.github.io

Andras Rozsa, Zheng Zhong, Terrance E. Boult

Regarding image forensics, researchers have proposed various approaches to detect and/or localize manipulations, such as splices. Recent best performing image-forensics algorithms greatly benefit from the application of deep learning, but such tools can be vulnerable to adversarial attacks. Due to the fact that most of the proposed adversarial example generation techniques can be used only on end-to-end classifiers, the adversarial robustness of image-forensics methods that utilize deep learning only for feature extraction has not been studied yet. Using a novel algorithm capable of directly adjusting the underlying representations of patches we demonstrate on three non end-to-end deep learning-based splice localization tools that hiding manipulations of images is feasible via adversarial attacks. While the tested image-forensics methods, EXIF-SC, SpliceRadar, and Noiseprint, rely on feature extractors that were trained on different surrogate tasks, we find that the formed adversarial perturbations can be transferable among them regarding the deterioration of their localization performance.

Zhenyu Tang, Wei Zhao, Xingzhi Xie et al.

Background: Chest computed tomography (CT) is recognized as an important tool for COVID-19 severity assessment. As the number of affected patients increase rapidly, manual severity assessment becomes a labor-intensive task, and may lead to delayed treatment. Purpose: Using machine learning method to realize automatic severity assessment (non-severe or severe) of COVID-19 based on chest CT images, and to explore the severity-related features from the resulting assessment model. Materials and Method: Chest CT images of 176 patients (age 45.3$\pm$16.5 years, 96 male and 80 female) with confirmed COVID-19 are used, from which 63 quantitative features, e.g., the infection volume/ratio of the whole lung and the volume of ground-glass opacity (GGO) regions, are calculated. A random forest (RF) model is trained to assess the severity (non-severe or severe) based on quantitative features. Importance of each quantitative feature, which reflects the correlation to the severity of COVID-19, is calculated from the RF model. Results: Using three-fold cross validation, the RF model shows promising results, i.e., 0.933 of true positive rate, 0.745 of true negative rate, 0.875 of accuracy, and 0.91 of area under receiver operating characteristic curve (AUC). The resulting importance of quantitative features shows that the volume and its ratio (with respect to the whole lung volume) of ground glass opacity (GGO) regions are highly related to the severity of COVID-19, and the quantitative features calculated from the right lung are more related to the severity assessment than those of the left lung. Conclusion: The RF based model can achieve automatic severity assessment (non-severe or severe) of COVID-19 infection, and the performance is promising. Several quantitative features, which have the potential to reflect the severity of COVID-19, were revealed.

Aurobrata Ghosh, Zheng Zhong, Steve Cruz et al.

We consider an information theoretic approach to address the problem of identifying fake digital images. We propose an innovative method to formulate the issue of localizing manipulated regions in an image as a deep representation learning problem using the Information Bottleneck (IB), which has recently gained popularity as a framework for interpreting deep neural networks. Tampered images pose a serious predicament since digitized media is a ubiquitous part of our lives. These are facilitated by the easy availability of image editing software and aggravated by recent advances in deep generative models such as GANs. We propose InfoPrint, a computationally efficient solution to the IB formulation using approximate variational inference and compare it to a numerical solution that is computationally expensive. Testing on a number of standard datasets, we demonstrate that InfoPrint outperforms the state-of-the-art and the numerical solution. Additionally, it also has the ability to detect alterations made by inpainting GANs.

Aurobrata Ghosh, Zheng Zhong, Terrance E Boult et al.

Detection and localization of image manipulations like splices are gaining in importance with the easy accessibility of image editing softwares. While detection generates a verdict for an image it provides no insight into the manipulation. Localization helps explain a positive detection by identifying the pixels of the image which have been tampered. We propose a deep learning based method for splice localization without prior knowledge of a test image's camera-model. It comprises a novel approach for learning rich filters and for suppressing image-edges. Additionally, we train our model on a surrogate task of camera model identification, which allows us to leverage large and widely available, unmanipulated, camera-tagged image databases. During inference, we assume that the spliced and host regions come from different camera-models and we segment these regions using a Gaussian-mixture model. Experiments on three test databases demonstrate results on par with and above the state-of-the-art and a good generalization ability to unknown datasets.