Hongwei Chen

Peng Di, Jianguo Li, Hang Yu et al.

Code Large Language Models (Code LLMs) have gained significant attention in the industry due to their wide applications in the full lifecycle of software engineering. However, the effectiveness of existing models in understanding non-English inputs for multi-lingual code-related tasks is still far from well studied. This paper introduces CodeFuse-13B, an open-sourced pre-trained code LLM. It is specifically designed for code-related tasks with both English and Chinese prompts and supports over 40 programming languages. CodeFuse achieves its effectiveness by utilizing a high quality pre-training dataset that is carefully filtered by program analyzers and optimized during the training process. Extensive experiments are conducted using real-world usage scenarios, the industry-standard benchmark HumanEval-x, and the specially designed CodeFuseEval for Chinese prompts. To assess the effectiveness of CodeFuse, we actively collected valuable human feedback from the AntGroup's software development process where CodeFuse has been successfully deployed. The results demonstrate that CodeFuse-13B achieves a HumanEval pass@1 score of 37.10%, positioning it as one of the top multi-lingual code LLMs with similar parameter sizes. In practical scenarios, such as code generation, code translation, code comments, and testcase generation, CodeFuse performs better than other models when confronted with Chinese prompts.

Sathya Chitturi, Zhurun Ji, Alexander Petsch et al.

The observation and description of collective excitations in solids is a fundamental issue when seeking to understand the physics of a many-body system. Analysis of these excitations is usually carried out by measuring the dynamical structure factor, S(Q, $ω$), with inelastic neutron or x-ray scattering techniques and comparing this against a calculated dynamical model. Here, we develop an artificial intelligence framework which combines a neural network trained to mimic simulated data from a model Hamiltonian with automatic differentiation to recover unknown parameters from experimental data. We benchmark this approach on a Linear Spin Wave Theory (LSWT) simulator and advanced inelastic neutron scattering data from the square-lattice spin-1 antiferromagnet La$_2$NiO$_4$. We find that the model predicts the unknown parameters with excellent agreement relative to analytical fitting. In doing so, we illustrate the ability to build and train a differentiable model only once, which then can be applied in real-time to multi-dimensional scattering data, without the need for human-guided peak finding and fitting algorithms. This prototypical approach promises a new technology for this field to automatically detect and refine more advanced models for ordered quantum systems.

Yuyao Huang, Tingzhao Fu, Honghao Huang et al.

The ever-growing deep learning technologies are making revolutionary changes for modern life. However, conventional computing architectures are designed to process sequential and digital programs, being extremely burdened with performing massive parallel and adaptive deep learning applications. Photonic integrated circuits provide an efficient approach to mitigate bandwidth limitations and power-wall brought by its electronic counterparts, showing great potential in ultrafast and energy-free high-performance computing. Here, we propose an optical computing architecture enabled by on-chip diffraction to implement convolutional acceleration, termed optical convolution unit (OCU). We demonstrate that any real-valued convolution kernels can be exploited by OCU with a prominent computational throughput boosting via the concept of structral re-parameterization. With OCU as the fundamental unit, we build an optical convolutional neural network (oCNN) to implement two popular deep learning tasks: classification and regression. For classification, Fashion-MNIST and CIFAR-4 datasets are tested with accuracy of 91.63% and 86.25%, respectively. For regression, we build an optical denoising convolutional neural network (oDnCNN) to handle Gaussian noise in gray scale images with noise level σ = 10, 15, 20, resulting clean images with average PSNR of 31.70dB, 29.39dB and 27.72dB, respectively. The proposed OCU presents remarkable performance of low energy consumption and high information density due to its fully passive nature and compact footprint, providing a highly parallel while lightweight solution for future computing architecture to handle high dimensional tensors in deep learning.

Haifeng Wang, Hua Wu, Tian Wu et al.

In this report, we introduce ERNIE 5.0, a natively autoregressive foundation model desinged for unified multimodal understanding and generation across text, image, video, and audio. All modalities are trained from scratch under a unified next-group-of-tokens prediction objective, based on an ultra-sparse mixture-of-experts (MoE) architecture with modality-agnostic expert routing. To address practical challenges in large-scale deployment under diverse resource constraints, ERNIE 5.0 adopts a novel elastic training paradigm. Within a single pre-training run, the model learns a family of sub-models with varying depths, expert capacities, and routing sparsity, enabling flexible trade-offs among performance, model size, and inference latency in memory- or time-constrained scenarios. Moreover, we systematically address the challenges of scaling reinforcement learning to unified foundation models, thereby guaranteeing efficient and stable post-training under ultra-sparse MoE architectures and diverse multimodal settings. Extensive experiments demonstrate that ERNIE 5.0 achieves strong and balanced performance across multiple modalities. To the best of our knowledge, among publicly disclosed models, ERNIE 5.0 represents the first production-scale realization of a trillion-parameter unified autoregressive model that supports both multimodal understanding and generation. To facilitate further research, we present detailed visualizations of modality-agnostic expert routing in the unified model, alongside comprehensive empirical analysis of elastic training, aiming to offer profound insights to the community.

Yubin Zang, Boyu Hua, Zhenzhou Tang et al.

In cater the need of Beyond 5G communications, large numbers of data driven artificial intelligence based fiber models has been put forward as to utilize artificial intelligence's regression ability to predict pulse evolution in fiber transmission at a much faster speed compared with the traditional split step Fourier method. In order to increase the physical interpretabiliy, principle driven fiber models have been proposed which inserts the Nonlinear Schodinger Equation into their loss functions. However, regardless of either principle driven or data driven models, they need to be re-trained the whole model under different transmission conditions. Unfortunately, this situation can be unavoidable when conducting the fiber communication optimization work. If the scale of different transmission conditions is large, then the whole model needs to be retrained large numbers of time with relatively large scale of parameters which may consume higher time costs. Computing efficiency will be dragged down as well. In order to address this problem, we propose the principle driven parameterized fiber model in this manuscript. This model breaks down the predicted NLSE solution with respect to one set of transmission condition into the linear combination of several eigen solutions which were outputted by each pre-trained principle driven fiber model via the reduced basis method. Therefore, the model can greatly alleviate the heavy burden of re-training since only the linear combination coefficients need to be found when changing the transmission condition. Not only strong physical interpretability can the model posses, but also higher computing efficiency can be obtained. Under the demonstration, the model's computational complexity is 0.0113% of split step Fourier method and 1% of the previously proposed principle driven fiber model.

Conghe Wang, Yutong He, Xia Wang et al.

Non-line-of-sight (NLOS) imaging is an emerging technique for detecting objects behind obstacles or around corners. Recent studies on passive NLOS mainly focus on steady-state measurement and reconstruction methods, which show limitations in recognition of moving targets. To the best of our knowledge, we propose a novel event-based passive NLOS imaging method. We acquire asynchronous event-based data which contains detailed dynamic information of the NLOS target, and efficiently ease the degradation of speckle caused by movement. Besides, we create the first event-based NLOS imaging dataset, NLOS-ES, and the event-based feature is extracted by time-surface representation. We compare the reconstructions through event-based data with frame-based data. The event-based method performs well on PSNR and LPIPS, which is 20% and 10% better than frame-based method, while the data volume takes only 2% of traditional method.

Yubin Zang, Boyu Hua, Zhipeng Lin et al.

In this manuscript, a novelty principle driven fiber transmission model for short-distance transmission with parameterized inputs is put forward. By taking into the account of the previously proposed principle driven fiber model, the reduced basis expansion method and transforming the parameterized inputs into parameterized coefficients of the Nonlinear Schrodinger Equations, universal solutions with respect to inputs corresponding to different bit rates can all be obtained without the need of re-training the whole model. This model, once adopted, can have prominent advantages in both computation efficiency and physical background. Besides, this model can still be effectively trained without the needs of transmitted signals collected in advance. Tasks of on-off keying signals with bit rates ranging from 2Gbps to 50Gbps are adopted to demonstrate the fidelity of the model.

Yubin Zang, Zuxing Zhang, Simin Li et al.

Optical neural networks have long cast attention nowadays. Like other optical structured neural networks, fiber neural networks which utilize the mechanism of light transmission to compute can take great advantages in both computing efficiency and power cost. Though the potential ability of optical fiber was demonstrated via the establishing of fiber neural networks, it will be of great significance of combining both fiber transmission and computing functions so as to cater the needs of future beyond 5G intelligent communication signal processing. Thus, in this letter, the fiber neural networks and their related optical signal processing methods will be both developed. In this way, information derived from the transmitted signals can be directly processed in the optical domain rather than being converted to the electronic domain. As a result, both prominent gains in processing efficiency and power cost can be further obtained. The fidelity of the whole structure and related methods is demonstrated by the task of modulation format recognition which plays important role in fiber optical communications without losing the generality.

Hongwei Chen, Yishu Lei, Dan Zhang et al.

Test-time scaling has emerged as a promising paradigm in language modeling, wherein additional computational resources are allocated during inference to enhance model performance. Recent approaches, such as DeepConf, have demonstrated the efficacy of this strategy, however, they often incur substantial computational overhead to achieve competitive results. In this work, we propose MatryoshkaThinking, a novel method that significantly reduces computational cost while maintaining state-of-the-art performance. Specifically, MatryoshkaThinking attains a score of 99.79 on AIME2025 using only 4% of the computation required by DeepConf. The core of our approach lies in the recursive exploitation of the model's intrinsic capabilities in reasoning, verification, and summarization, which collectively enhance the retention of correct solutions and reduce the disparity between Pass@k and Pass@1. Comprehensive evaluations across multiple open-source models and challenging multi-modal reasoning benchmarks validate the effectiveness and generality of our method. These findings offer new insights into the design of efficient and scalable test-time inference strategies for advanced language models.

Ming Liang, Xiaoheng Xie, Gehao Zhang et al.

The success of language models in code assistance has spurred the proposal of repository-level code completion as a means to enhance prediction accuracy, utilizing the context from the entire codebase. However, this amplified context can inadvertently increase inference latency, potentially undermining the developer experience and deterring tool adoption - a challenge we termed the Context-Latency Conundrum. This paper introduces REPOFUSE, a pioneering solution designed to enhance repository-level code completion without the latency trade-off. REPOFUSE uniquely fuses two types of context: the analogy context, rooted in code analogies, and the rationale context, which encompasses in-depth semantic relationships. We propose a novel rank truncated generation (RTG) technique that efficiently condenses these contexts into prompts with restricted size. This enables REPOFUSE to deliver precise code completions while maintaining inference efficiency. Through testing with the CrossCodeEval suite, REPOFUSE has demonstrated a significant leap over existing models, achieving a 40.90% to 59.75% increase in exact match (EM) accuracy for code completions and a 26.8% enhancement in inference speed. Beyond experimental validation, REPOFUSE has been integrated into the workflow of a large enterprise, where it actively supports various coding tasks.

Shuo Yang, Zirui Shang, Yongqi Wang et al.

This paper proposes a novel framework for multi-label image recognition without any training data, called data-free framework, which uses knowledge of pre-trained Large Language Model (LLM) to learn prompts to adapt pretrained Vision-Language Model (VLM) like CLIP to multilabel classification. Through asking LLM by well-designed questions, we acquire comprehensive knowledge about characteristics and contexts of objects, which provides valuable text descriptions for learning prompts. Then we propose a hierarchical prompt learning method by taking the multi-label dependency into consideration, wherein a subset of category-specific prompt tokens are shared when the corresponding objects exhibit similar attributes or are more likely to co-occur. Benefiting from the remarkable alignment between visual and linguistic semantics of CLIP, the hierarchical prompts learned from text descriptions are applied to perform classification of images during inference. Our framework presents a new way to explore the synergies between multiple pre-trained models for novel category recognition. Extensive experiments on three public datasets (MS-COCO, VOC2007, and NUS-WIDE) demonstrate that our method achieves better results than the state-of-the-art methods, especially outperforming the zero-shot multi-label recognition methods by 4.7% in mAP on MS-COCO.

Linghan Zheng, Hui Liu, Xiaojun Lin et al.

In previous studies, code-based models have consistently outperformed text-based models in reasoning-intensive scenarios. When generating our knowledge base for Retrieval-Augmented Generation (RAG), we observed that code-based models also perform exceptionally well in Chinese QA Pair Extraction task. Further, our experiments and the metrics we designed discovered that code-based models containing a certain amount of Chinese data achieve even better performance. Additionally, the capabilities of code-based English models in specified Chinese tasks offer a distinct perspective for discussion on the philosophical "Chinese Room" thought experiment.

Chaojie Ji, Hongwei Chen, Ruxin Wang et al.

Clustering techniques attempt to group objects with similar properties into a cluster. Clustering the nodes of an attributed graph, in which each node is associated with a set of feature attributes, has attracted significant attention. Graph convolutional networks (GCNs) represent an effective approach for integrating the two complementary factors of node attributes and structural information for attributed graph clustering. However, oversmoothing of GCNs produces indistinguishable representations of nodes, such that the nodes in a graph tend to be grouped into fewer clusters, and poses a challenge due to the resulting performance drop. In this study, we propose a smoothness sensor for attributed graph clustering based on adaptive smoothness-transition graph convolutions, which senses the smoothness of a graph and adaptively terminates the current convolution once the smoothness is saturated to prevent oversmoothing. Furthermore, as an alternative to graph-level smoothness, a novel fine-gained node-wise level assessment of smoothness is proposed, in which smoothness is computed in accordance with the neighborhood conditions of a given node at a certain order of graph convolution. In addition, a self-supervision criterion is designed considering both the tightness within clusters and the separation between clusters to guide the whole neural network training process. Experiments show that the proposed methods significantly outperform 12 other state-of-the-art baselines in terms of three different metrics across four benchmark datasets. In addition, an extensive study reveals the reasons for their effectiveness and efficiency.

Yubin Zang, Minghua Chen, Sigang Yang et al.

In this paper, a novel architecture of electro-optical neural networks based on the time-stretch method is proposed and numerically simulated. By stretching time-domain ultrashort pulses, multiplications of large scale weight matrices and vectors can be implemented on light and multiple-layer of feedforward neural network operations can be easily implemented with fiber loops. Via simulation, the performance of a three-layer electro-optical neural network is tested by the handwriting digit recognition task and the accuracy reaches 88% under considerable noise.

Qiang Guo, Hongwei Chen, Yuxi Wang et al.

Single-pixel cameras based on the concepts of compressed sensing (CS) leverage the inherent structure of images to retrieve them with far fewer measurements and operate efficiently over a significantly broader spectral range than conventional silicon-based cameras. Recently, photonic time-stretch (PTS) technique facilitates the emergence of high-speed single-pixel cameras. A significant breakthrough in imaging speed of single-pixel cameras enables observation of fast dynamic phenomena. However, according to CS theory, image reconstruction is an iterative process that consumes enormous amounts of computational time and cannot be performed in real time. To address this challenge, we propose a novel single-pixel imaging technique that can produce high-quality images through rapid acquisition of their effective spatial Fourier spectrum. We employ phase-shifting sinusoidal structured illumination instead of random illumination for spectrum acquisition and apply inverse Fourier transform to the obtained spectrum for image restoration. We evaluate the performance of our prototype system by recognizing quick response (QR) codes and flow cytometric screening of cells. A frame rate of 625 kHz and a compression ratio of 10% are experimentally demonstrated in accordance with the recognition rate of the QR code. An imaging flow cytometer enabling high-content screening with an unprecedented throughput of 100,000 cells/s is also demonstrated. For real-time imaging applications, the proposed single-pixel microscope can significantly reduce the time required for image reconstruction by two orders of magnitude, which can be widely applied in industrial quality control and label-free biomedical imaging.