Yubo Zhang

Yubo Zhang, Xingxing Zhang, Xun Wang et al. · microsoft-research

Prompts with different control signals (e.g., length, keywords, etc.) can be used to control text summarization. When control signals are available, they can control the properties of generated summaries and potentially improve summarization quality (since more information are given). Unfortunately, control signals are not already available during inference time. In this paper, we propose Lotus (shorthand for Latent Prompt Tuning for Summarization), which is a single model that can be applied in both controlled and uncontrolled (without control signals) modes. During training, Lotus learns latent prompt representations from prompts with gold control signals using a contrastive learning objective. Experiments show Lotus in uncontrolled mode consistently improves upon strong (uncontrollable) summarization models across four different summarization datasets. We also demonstrate generated summaries can be controlled using prompts with user specified control tokens.

Cheng Cui, Ting Sun, Suyin Liang et al.

We introduce PaddleOCR-VL-1.5, an upgraded model achieving a new state-of-the-art (SOTA) accuracy of 94.5% on OmniDocBench v1.5. To rigorously evaluate robustness against real-world physical distortions, including scanning, skew, warping, screen-photography, and illumination, we propose the Real5-OmniDocBench benchmark. Experimental results demonstrate that this enhanced model attains SOTA performance on the newly curated benchmark. Furthermore, we extend the model's capabilities by incorporating seal recognition and text spotting tasks, while remaining a 0.9B ultra-compact VLM with high efficiency. Code: https://github.com/PaddlePaddle/PaddleOCR

Zelun Zhang, Hongen Liu, Suyin Liang et al.

We introduce PaddleOCR-VL-1.6, an upgraded compact document parsing model built upon PaddleOCR-VL-1.5. Although PaddleOCR-VL-1.5 establishes a strong 0.9B baseline, its remaining errors concentrate in under-optimized regions where model behavior is unstable, data coverage is sparse, or supervision is unreliable. Rather than expanding the training corpus indiscriminately, PaddleOCR-VL-1.6 introduces a region-aware data optimization framework that identifies weak regions from the previous model, applies targeted enhancement to these regions, and improves the reliability of supervision signals. It further adopts a progressive post-training recipe based on curated data selection and reinforcement learning, pushing model performance to a higher level through staged optimization. PaddleOCR-VL-1.6 achieves a new state-of-the-art score of 96.33% on OmniDocBench v1.6, demonstrates strong competitiveness against top-tier VLMs, and provides a practical post-training recipe for the PaddleOCR-VL series.

Shangchen Han, Po-chen Wu, Yubo Zhang et al.

Real-time tracking of 3D hand pose in world space is a challenging problem and plays an important role in VR interaction. Existing work in this space are limited to either producing root-relative (versus world space) 3D pose or rely on multiple stages such as generating heatmaps and kinematic optimization to obtain 3D pose. Moreover, the typical VR scenario, which involves multi-view tracking from wide \ac{fov} cameras is seldom addressed by these methods. In this paper, we present a unified end-to-end differentiable framework for multi-view, multi-frame hand tracking that directly predicts 3D hand pose in world space. We demonstrate the benefits of end-to-end differentiabilty by extending our framework with downstream tasks such as jitter reduction and pinch prediction. To demonstrate the efficacy of our model, we further present a new large-scale egocentric hand pose dataset that consists of both real and synthetic data. Experiments show that our system trained on this dataset handles various challenging interactive motions, and has been successfully applied to real-time VR applications.

Zhihu Wang, Shiwan Zhao, Yu Wang et al.

The Chain-of-Thought (CoT) paradigm has become a pivotal method for solving complex problems with large language models (LLMs). However, its application to domain-specific tasks remains challenging, as LLMs often fail to decompose tasks accurately or execute subtasks effectively. This paper introduces the Re-TASK framework, a novel theoretical model that revisits LLM tasks from capability, skill, and knowledge perspectives, drawing on the principles of Bloom's Taxonomy and Knowledge Space Theory. While CoT provides a workflow-centric perspective on tasks, Re-TASK introduces a Chain-of-Learning (CoL) paradigm that highlights task dependencies on specific capability items, further broken down into their constituent knowledge and skill components. To address CoT failures, we propose a Re-TASK prompting strategy, which strengthens task-relevant capabilities through targeted knowledge injection and skill adaptation. Experiments across diverse domains demonstrate the effectiveness of Re-TASK. In particular, we achieve improvements of 45.00% on Yi-1.5-9B and 24.50% on Llama3-Chinese-8B for legal tasks. These results highlight the potential of Re-TASK to significantly enhance LLM performance and its applicability in specialized domains. We release our code and data at https://github.com/Uylee/Re-TASK.

Shuxian Wang, Yubo Zhang, Sarah K. McGill et al.

Reconstructing a 3D surface from colonoscopy video is challenging due to illumination and reflectivity variation in the video frame that can cause defective shape predictions. Aiming to overcome this challenge, we utilize the characteristics of surface normal vectors and develop a two-step neural framework that significantly improves the colonoscopy reconstruction quality. The normal-based depth initialization network trained with self-supervised normal consistency loss provides depth map initialization to the normal-depth refinement module, which utilizes the relationship between illumination and surface normals to refine the frame-wise normal and depth predictions recursively. Our framework's depth accuracy performance on phantom colonoscopy data demonstrates the value of exploiting the surface normals in colonoscopy reconstruction, especially on en face views. Due to its low depth error, the prediction result from our framework will require limited post-processing to be clinically applicable for real-time colonoscopy reconstruction.

Cheng Cui, Ting Sun, Suyin Liang et al.

Document parsing is a fine-grained task where image resolution significantly impacts performance. While advanced research leveraging vision-language models benefits from high-resolution input to boost model performance, this often leads to a quadratic increase in the number of vision tokens and significantly raises computational costs. We attribute this inefficiency to substantial visual regions redundancy in document images, like background. To tackle this, we propose PaddleOCR-VL, a novel coarse-to-fine architecture that focuses on semantically relevant regions while suppressing redundant ones, thereby improving both efficiency and performance. Specifically, we introduce a lightweight Valid Region Focus Module (VRFM) which leverages localization and contextual relationship prediction capabilities to identify valid vision tokens. Subsequently, we design and train a compact yet powerful 0.9B vision-language model (PaddleOCR-VL-0.9B) to perform detailed recognition, guided by VRFM outputs to avoid direct processing of the entire large image. Extensive experiments demonstrate that PaddleOCR-VL achieves state-of-the-art performance in both page-level parsing and element-level recognition. It significantly outperforms existing solutions, exhibits strong competitiveness against top-tier VLMs, and delivers fast inference while utilizing substantially fewer vision tokens and parameters, highlighting the effectiveness of targeted coarse-to-fine parsing for accurate and efficient document understanding. The source code and models are publicly available at https://github.com/PaddlePaddle/PaddleOCR.

Yubo Zhang, Shudi Hou, Mingyu Derek Ma et al.

Large language models (LLMs) are increasingly applied to clinical decision-making. However, their potential to exhibit bias poses significant risks to clinical equity. Currently, there is a lack of benchmarks that systematically evaluate such clinical bias in LLMs. While in downstream tasks, some biases of LLMs can be avoided such as by instructing the model to answer "I'm not sure...", the internal bias hidden within the model still lacks deep studies. We introduce CLIMB (shorthand for A Benchmark of Clinical Bias in Large Language Models), a pioneering comprehensive benchmark to evaluate both intrinsic (within LLMs) and extrinsic (on downstream tasks) bias in LLMs for clinical decision tasks. Notably, for intrinsic bias, we introduce a novel metric, AssocMAD, to assess the disparities of LLMs across multiple demographic groups. Additionally, we leverage counterfactual intervention to evaluate extrinsic bias in a task of clinical diagnosis prediction. Our experiments across popular and medically adapted LLMs, particularly from the Mistral and LLaMA families, unveil prevalent behaviors with both intrinsic and extrinsic bias. This work underscores the critical need to mitigate clinical bias and sets a new standard for future evaluations of LLMs' clinical bias.

Cheng Cui, Yubo Zhang, Ting Sun et al.

The advent of "OCR 2.0" and large-scale vision-language models (VLMs) has set new benchmarks in text recognition. However, these unified architectures often come with significant computational demands, challenges in precise text localization within complex layouts, and a propensity for textual hallucinations. Revisiting the prevailing notion that model scale is the sole path to high accuracy, this paper introduces PP-OCRv5, a meticulously optimized, lightweight OCR system with merely 5 million parameters. We demonstrate that PP-OCRv5 achieves performance competitive with many billion-parameter VLMs on standard OCR benchmarks, while offering superior localization precision and reduced hallucinations. The cornerstone of our success lies not in architectural expansion but in a data-centric investigation. We systematically dissect the role of training data by quantifying three critical dimensions: data difficulty, data accuracy, and data diversity. Our extensive experiments reveal that with a sufficient volume of high-quality, accurately labeled, and diverse data, the performance ceiling for traditional, efficient two-stage OCR pipelines is far higher than commonly assumed. This work provides compelling evidence for the viability of lightweight, specialized models in the large-model era and offers practical insights into data curation for OCR. The source code and models are publicly available at https://github.com/PaddlePaddle/PaddleOCR.

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.

Yubo Zhang, Feiyang Niu, Qing Ping et al.

To solve video-and-language grounding tasks, the key is for the network to understand the connection between the two modalities. For a pair of video and language description, their semantic relation is reflected by their encodings' similarity. A good multi-modality encoder should be able to well capture both inputs' semantics and encode them in the shared feature space where embedding distance gets properly translated into their semantic similarity. In this work, we focused on this semantic connection between video and language, and developed a multi-level alignment training scheme to directly shape the encoding process. Global and segment levels of video-language alignment pairs were designed, based on the information similarity ranging from high-level context to fine-grained semantics. The contrastive loss was used to contrast the encodings' similarities between the positive and negative alignment pairs, and to ensure the network is trained in such a way that similar information is encoded closely in the shared feature space while information of different semantics is kept apart. Our multi-level alignment training can be applied to various video-and-language grounding tasks. Together with the task-specific training loss, our framework achieved comparable performance to previous state-of-the-arts on multiple video QA and retrieval datasets.

Yu Wang, Shiwan Zhao, Zhihu Wang et al.

The Chain-of-Thought (CoT) paradigm has emerged as a critical approach for enhancing the reasoning capabilities of large language models (LLMs). However, despite their widespread adoption and success, CoT methods often exhibit instability due to their inability to consistently ensure the quality of generated reasoning paths, leading to sub-optimal reasoning performance. To address this challenge, we propose the \textbf{Strategic Chain-of-Thought} (SCoT), a novel methodology designed to refine LLM performance by integrating strategic knowledge prior to generating intermediate reasoning steps. SCoT employs a two-stage approach within a single prompt: first eliciting an effective problem-solving strategy, which is then used to guide the generation of high-quality CoT paths and final answers. Our experiments across eight challenging reasoning datasets demonstrate significant improvements, including a 21.05\% increase on the GSM8K dataset and 24.13\% on the Tracking\_Objects dataset, respectively, using the Llama3-8b model. Additionally, we extend the SCoT framework to develop a few-shot method with automatically matched demonstrations, yielding even stronger results. These findings underscore the efficacy of SCoT, highlighting its potential to substantially enhance LLM performance in complex reasoning tasks.

Yubo Zhang, Zhiguo Hao, Songhao Yang et al.

Additional active power control (AAPC) of wind turbines (WTs) is essential to improve the transient frequency stability of low-inertia power systems. Most of the existing research has focused on imitating the frequency response of the synchronous generator (SG), known as virtual inertia control (VIC), but are such control laws optimal for the power systems? Inspired by this question, this paper proposes an optimal AAPC of WTs to maximize the frequency nadir post a major power deficit. By decoupling the WT response and the frequency dynamics, the optimal frequency trajectory is solved based on the trajectory model, and its universality is strictly proven. Then the optimal AAPC of WTs is constructed reversely based on the average system frequency (ASF) model with the optimal frequency trajectory as the desired control results. The proposed method can significantly improve the system frequency nadir. Meanwhile, the event insensitivity makes it can be deployed based on the on-line rolling update under a hypothetic disturbance, avoiding the heavy post-event computational burden. Finally, simulation results in a two-machine power system and the IEEE 39 bus power system verify the effectiveness of the optimal AAPC of WTs.

Songhao Yang, Yubo Zhang, Zhiguo Hao et al.

Current transformer (CT) saturation is one of the dominant causes of relay protection devices' malfunctions, which pose a threat to the safe operation of the power system. To address this problem, we propose a hybrid physical model- and data-driven method. The method firstly detects the CT saturation and then compensates it to reproduce the real waveform. Considering the multi-factor and strong nonlinearity of CT saturation, a data-driven model, namely the Fully Convolutional Network (FCN), is built to detect the operation status of CT. As for the compensation, a physical model of short-circuit current is used for its conciseness and universality. Through tactfully integrating the data model and the physical model, the proposed method is endowed with two major merits: the arduous adjustment of universal thresholds and parameters in existing methods is avoided, and the deficiency in generalization and interpretability of the data-driven method is assuaged. Simulation and experimental results verify the effectiveness of the proposed method. Furthermore, its application potential to future protection is explored.

Yuan Liao, Jiang Bian, Yuhui Yun et al.

While the field of NL2SQL has made significant advancements in translating natural language instructions into executable SQL scripts for data querying and processing, achieving full automation within the broader data science pipeline - encompassing data querying, analysis, visualization, and reporting - remains a complex challenge. This study introduces SageCopilot, an advanced, industry-grade system system that automates the data science pipeline by integrating Large Language Models (LLMs), Autonomous Agents (AutoAgents), and Language User Interfaces (LUIs). Specifically, SageCopilot incorporates a two-phase design: an online component refining users' inputs into executable scripts through In-Context Learning (ICL) and running the scripts for results reporting & visualization, and an offline preparing demonstrations requested by ICL in the online phase. A list of trending strategies such as Chain-of-Thought and prompt-tuning have been used to augment SageCopilot for enhanced performance. Through rigorous testing and comparative analysis against prompt-based solutions, SageCopilot has been empirically validated to achieve superior end-to-end performance in generating or executing scripts and offering results with visualization, backed by real-world datasets. Our in-depth ablation studies highlight the individual contributions of various components and strategies used by SageCopilot to the end-to-end correctness for data sciences.

Xiaoxin Lu, Yubo Zhang, Jing Li et al.

Huge volumes of patient queries are daily generated on online health forums, rendering manual doctor allocation a labor-intensive task. To better help patients, this paper studies a novel task of doctor recommendation to enable automatic pairing of a patient to a doctor with relevant expertise. While most prior work in recommendation focuses on modeling target users from their past behavior, we can only rely on the limited words in a query to infer a patient's needs for privacy reasons. For doctor modeling, we study the joint effects of their profiles and previous dialogues with other patients and explore their interactions via self-learning. The learned doctor embeddings are further employed to estimate their capabilities of handling a patient query with a multi-head attention mechanism. For experiments, a large-scale dataset is collected from Chunyu Yisheng, a Chinese online health forum, where our model exhibits the state-of-the-art results, outperforming baselines only consider profiles and past dialogues to characterize a doctor.

Zexuan Lin, Songhao Yang, Yubo Zhang et al.

When a slightly faulty transformer closes without load, the current waveform presents the coexistence of inrush and fault current. At this time, the inrush blocking module will block the relay, which may delay the removal of the slight fault and lead to more serious faults. To address this problem, this paper proposes a data-aided power transformer differential protection without inrush blocking module. The key to eliminating the negative influence of inrush current is to extract the fundamental component from the non-inrush part of the current waveform, which corresponds to the unsaturation period of the transformer core. Firstly, a data-aided module, namely an Attention module embedded Fully Convolutional Network (A-FCN), is built to distinguish the inrush and non-inrush parts of the current waveform. Then, a physical model of the current waveform is built for the non-inrush part, and the fundamental component is extracted by the nonlinear least square (NLS) algorithm. The proposed method can avoid the block of differential protections when inrush current occurs, which improves the sensitivity and rapidity of the relay, especially in the case of a weak internal fault hidden in inrush current. Finally, simulation and experimental data verify the effectiveness and generalization of the proposed method.

Cheng Cui, Ting Sun, Manhui Lin et al.

This technical report introduces PaddleOCR 3.0, an Apache-licensed open-source toolkit for OCR and document parsing. To address the growing demand for document understanding in the era of large language models, PaddleOCR 3.0 presents three major solutions: (1) PP-OCRv5 for multilingual text recognition, (2) PP-StructureV3 for hierarchical document parsing, and (3) PP-ChatOCRv4 for key information extraction. Compared to mainstream vision-language models (VLMs), these models with fewer than 100 million parameters achieve competitive accuracy and efficiency, rivaling billion-parameter VLMs. In addition to offering a high-quality OCR model library, PaddleOCR 3.0 provides efficient tools for training, inference, and deployment, supports heterogeneous hardware acceleration, and enables developers to easily build intelligent document applications.

Cheng Cui, Ting Sun, Suyin Liang et al.

In this report, we propose PaddleOCR-VL, a SOTA and resource-efficient model tailored for document parsing. Its core component is PaddleOCR-VL-0.9B, a compact yet powerful vision-language model (VLM) that integrates a NaViT-style dynamic resolution visual encoder with the ERNIE-4.5-0.3B language model to enable accurate element recognition. This innovative model efficiently supports 109 languages and excels in recognizing complex elements (e.g., text, tables, formulas, and charts), while maintaining minimal resource consumption. Through comprehensive evaluations on widely used public benchmarks and in-house benchmarks, PaddleOCR-VL achieves SOTA performance in both page-level document parsing and element-level recognition. It significantly outperforms existing solutions, exhibits strong competitiveness against top-tier VLMs, and delivers fast inference speeds. These strengths make it highly suitable for practical deployment in real-world scenarios. Code is available at https://github.com/PaddlePaddle/PaddleOCR .

Yubo Zhang, Yanfang Liu, Xinxin Fan et al.

Code search aims to retrieve the code snippet that highly matches the given query described in natural language. Recently, many code pre-training approaches have demonstrated impressive performance on code search. However, existing code search methods still suffer from two performance constraints: inadequate semantic representation and the semantic gap between natural language (NL) and programming language (PL). In this paper, we propose CPLCS, a contrastive prompt learning-based code search method based on the cross-modal interaction mechanism. CPLCS comprises:(1) PL-NL contrastive learning, which learns the semantic matching relationship between PL and NL representations; (2) a prompt learning design for a dual-encoder structure that can alleviate the problem of inadequate semantic representation; (3) a cross-modal interaction mechanism to enhance the fine-grained mapping between NL and PL. We conduct extensive experiments to evaluate the effectiveness of our approach on a real-world dataset across six programming languages. The experiment results demonstrate the efficacy of our approach in improving semantic representation quality and mapping ability between PL and NL.

Yubo Zhang, Hao Tan, Mohit Bansal

Vision-and-Language Navigation (VLN) requires an agent to follow natural-language instructions, explore the given environments, and reach the desired target locations. These step-by-step navigational instructions are crucial when the agent is navigating new environments about which it has no prior knowledge. Most recent works that study VLN observe a significant performance drop when tested on unseen environments (i.e., environments not used in training), indicating that the neural agent models are highly biased towards training environments. Although this issue is considered as one of the major challenges in VLN research, it is still under-studied and needs a clearer explanation. In this work, we design novel diagnosis experiments via environment re-splitting and feature replacement, looking into possible reasons for this environment bias. We observe that neither the language nor the underlying navigational graph, but the low-level visual appearance conveyed by ResNet features directly affects the agent model and contributes to this environment bias in results. According to this observation, we explore several kinds of semantic representations that contain less low-level visual information, hence the agent learned with these features could be better generalized to unseen testing environments. Without modifying the baseline agent model and its training method, our explored semantic features significantly decrease the performance gaps between seen and unseen on multiple datasets (i.e. R2R, R4R, and CVDN) and achieve competitive unseen results to previous state-of-the-art models. Our code and features are available at: https://github.com/zhangybzbo/EnvBiasVLN

Longxuan Wei, Yubo Zhang, Zijiao Zhang et al.

Large language models achieve strong reasoning performance, yet existing decoding strategies either explore blindly (random sampling) or redundantly (independent multi-sampling). We propose Entropy-Tree, a tree-based decoding method that exploits entropy as a signal for branching decisions--expanding the search tree only at positions where the model exhibits genuine uncertainty. Entropy-Tree shows superior accuracy and calibration in reasoning tasks: it achieves better pass@k than Multi-chain across multiple models and datasets, and its predictive entropy demonstrates better AUROC compared to several traditional metrics. Entropy-Tree unifies efficient structured exploration and reliable uncertainty estimation within a single decoding procedure.

Yubo Zhang, Jeremy Johnston, Xiaodong Wang

We develop an end-to-end deep learning framework for downlink beamforming in large-scale sparse MIMO channels. The core is a deep EDN architecture with three modules: (i) an encoder NN, deployed at each user end, that compresses estimated downlink channels into low-dimensional latent vectors. The latent vector from each user is compressed and then fed back to the BS. (ii) a beamformer decoder NN at the BS that maps recovered latent vectors to beamformers, and (iii) a channel decoder NN at the BS that reconstructs downlink channels from recovered latent vectors to further refine the beamformers. The training of EDN leverages two key strategies: (a) semi-amortized learning, where the beamformer decoder NN contains an analytical gradient ascent during both training and inference stages, and (b) knowledge distillation, where the loss function consists of a supervised term and an unsupervised term, and starting from supervised training with MMSE beamformers, over the epochs, the model training gradually shifts toward unsupervised using the sum-rate objective. The proposed EDN beamforming framework is extended to both far-field and near-field hybrid beamforming scenarios. Extensive simulations validate its effectiveness under diverse network and channel conditions.

Yu Wang, Shiwan Zhao, Zhihu Wang et al.

The integration of external knowledge through Retrieval-Augmented Generation (RAG) has become foundational in enhancing large language models (LLMs) for knowledge-intensive tasks. However, existing RAG paradigms often overlook the cognitive step of applying knowledge, leaving a gap between retrieved facts and task-specific reasoning. In this work, we introduce RAG+, a principled and modular extension that explicitly incorporates application-aware reasoning into the RAG pipeline. RAG+ constructs a dual corpus consisting of knowledge and aligned application examples, created either manually or automatically, and retrieves both jointly during inference. This design enables LLMs not only to access relevant information but also to apply it within structured, goal-oriented reasoning processes. Experiments across mathematical, legal, and medical domains, conducted on multiple models, demonstrate that RAG+ consistently outperforms standard RAG variants, achieving average improvements of 3-5%, and peak gains up to 13.5% in complex scenarios. By bridging retrieval with actionable application, RAG+ advances a more cognitively grounded framework for knowledge integration, representing a step toward more interpretable and capable LLMs.

Yubo Zhang, Songhao Yang, Zhiguo Hao et al.

The fast frequency support (FFS) towards frequency trajectory optimization provides a system view for the frequency regulation of wind farms (WFs). However, the existing frequency trajectory optimization-based FFS generally relies on the accurate governor dynamics model of synchronous generators (SGs), which aggrandizes the difficulty of controller implementation. In this paper, a proportional-integral (PI) based FFS of WFs is designed for tracking the optimal frequency trajectory, which gets rid of the dependence on the governor model. Firstly, the prototypical PI-based FFS of WFs is proposed and its feasibility for tracking the optimal frequency trajectory is analyzed and demonstrated. Then, based on the "frequency-RoCoF" form of the optimal frequency trajectory, a more practical PI controller is constructed, avoiding the time dependence of the prototypical PI controller. Besides, an adaptive gain associated with PI parameters is designed for multi-WF coordination. Finally, the validity of the proposed method is verified in both the single-WF system and the multi-WF system.

Yubo Zhang, Xiao-Yang Liu, Xiaodong Wang

We develop an unsupervised deep learning framework for downlink beamforming in large-scale MU-MISO channels. The model is trained offline, allowing real-time inference through lightweight feedforward computations in dynamic communication environments. Following the learning-to-optimize (L2O) paradigm, a multi-layer Transformer iteratively refines both channel and beamformer features via residual connections. To enhance training, three strategies are introduced: (i) curriculum learning (CL) to improve early-stage convergence and avoid local optima, (ii) semi-amortized learning to refine each Transformer block with a few gradient ascent steps, and (iii) sliding-window training to stabilize optimization by training only a subset of Transformer blocks at a time. Extensive simulations show that the proposed scheme outperforms existing baselines at low-to-medium SNRs and closely approaches WMMSE performance at high SNRs, while achieving substantially faster inference than iterative and online learning approaches.

Jinlin Xiang, Minho Choi, Yubo Zhang et al.

Hybrid Optical Neural Networks (ONNs, typically consisting of an optical frontend and a digital backend) offer an energy-efficient alternative to fully digital deep networks for real-time, power-constrained systems. However, their adoption is limited by two main challenges: the accuracy gap compared to large-scale networks during training, and discrepancies between simulated and fabricated systems that further degrade accuracy. While previous work has proposed end-to-end optimizations for specific datasets (e.g., MNIST) and optical systems, these approaches typically lack generalization across tasks and hardware designs. To address these limitations, we propose a task-agnostic and hardware-agnostic pipeline that supports image classification and segmentation across diverse optical systems. To assist optical system design before training, we estimate achievable model accuracy based on user-specified constraints such as physical size and the dataset. For training, we introduce Neural Tangent Knowledge Distillation (NTKD), which aligns optical models with electronic teacher networks, thereby narrowing the accuracy gap. After fabrication, NTKD also guides fine-tuning of the digital backend to compensate for implementation errors. Experiments on multiple datasets (e.g., MNIST, CIFAR, Carvana Masking) and hardware configurations show that our pipeline consistently improves ONN performance and enables practical deployment in both pre-fabrication simulations and physical implementations.

Yubo Zhang, Pedro Botelho, Trevor Gordon et al.

We consider a decentralized wireless network with several source-destination pairs sharing a limited number of orthogonal frequency bands. Sources learn to adapt their transmissions (specifically, their band selection strategy) over time, in a decentralized manner, without sharing information with each other. Sources can only observe the outcome of their own transmissions (i.e., success or collision), having no prior knowledge of the network size or of the transmission strategy of other sources. The goal of each source is to maximize their own throughput while striving for network-wide fairness. We propose a novel fully decentralized Reinforcement Learning (RL)-based solution that achieves fairness without coordination. The proposed Fair Share RL (FSRL) solution combines: (i) state augmentation with a semi-adaptive time reference; (ii) an architecture that leverages risk control and time difference likelihood; and (iii) a fairness-driven reward structure. We evaluate FSRL in more than 50 network settings with different number of agents, different amounts of available spectrum, in the presence of jammers, and in an ad-hoc setting. Simulation results suggest that, when we compare FSRL with a common baseline RL algorithm from the literature, FSRL can be up to 89.0% fairer (as measured by Jain's fairness index) in stringent settings with several sources and a single frequency band, and 48.1% fairer on average.

Yubo Zhang, Jan-Michael Frahm, Samuel Ehrenstein et al.

One of the key elements of reconstructing a 3D mesh from a monocular video is generating every frame's depth map. However, in the application of colonoscopy video reconstruction, producing good-quality depth estimation is challenging. Neural networks can be easily fooled by photometric distractions or fail to capture the complex shape of the colon surface, predicting defective shapes that result in broken meshes. Aiming to fundamentally improve the depth estimation quality for colonoscopy 3D reconstruction, in this work we have designed a set of training losses to deal with the special challenges of colonoscopy data. For better training, a set of geometric consistency objectives was developed, using both depth and surface normal information. Also, the classic photometric loss was extended with feature matching to compensate for illumination noise. With the training losses powerful enough, our self-supervised framework named ColDE is able to produce better depth maps of colonoscopy data as compared to the previous work utilizing prior depth knowledge. Used in reconstruction, our network is able to reconstruct good-quality colon meshes in real-time without any post-processing, making it the first to be clinically applicable.

Yubo Zhang, Shuxian Wang, Ruibin Ma et al.

High screening coverage during colonoscopy is crucial to effectively prevent colon cancer. Previous work has allowed alerting the doctor to unsurveyed regions by reconstructing the 3D colonoscopic surface from colonoscopy videos in real-time. However, the lighting inconsistency of colonoscopy videos can cause a key component of the colonoscopic reconstruction system, the SLAM optimization, to fail. In this work we focus on the lighting problem in colonoscopy videos. To successfully improve the lighting consistency of colonoscopy videos, we have found necessary a lighting correction that adapts to the intensity distribution of recent video frames. To achieve this in real-time, we have designed and trained an RNN network. This network adapts the gamma value in a gamma-correction process. Applied in the colonoscopic surface reconstruction system, our light-weight model significantly boosts the reconstruction success rate, making a larger proportion of colonoscopy video segments reconstructable and improving the reconstruction quality of the already reconstructed segments.

Peirong Liu, Lin Tian, Yubo Zhang et al.

Transport processes are ubiquitous. They are, for example, at the heart of optical flow approaches; or of perfusion imaging, where blood transport is assessed, most commonly by injecting a tracer. An advection-diffusion equation is widely used to describe these transport phenomena. Our goal is estimating the underlying physics of advection-diffusion equations, expressed as velocity and diffusion tensor fields. We propose a learning framework (YETI) building on an auto-encoder structure between 2D and 3D image time-series, which incorporates the advection-diffusion model. To help with identifiability, we develop an advection-diffusion simulator which allows pre-training of our model by supervised learning using the velocity and diffusion tensor fields. Instead of directly learning these velocity and diffusion tensor fields, we introduce representations that assure incompressible flow and symmetric positive semi-definite diffusion fields and demonstrate the additional benefits of these representations on improving estimation accuracy. We further use transfer learning to apply YETI on a public brain magnetic resonance (MR) perfusion dataset of stroke patients and show its ability to successfully distinguish stroke lesions from normal brain regions via the estimated velocity and diffusion tensor fields.

Hengrui Wang, Yubo Zhang, Mingzhi Chen et al.

Spectral clustering is one of the most prominent clustering approaches. The distance-based similarity is the most widely used method for spectral clustering. However, people have already noticed that this is not suitable for multi-scale data, as the distance varies a lot for clusters with different densities. State of the art(ROSC and CAST ) addresses this limitation by taking the reachability similarity of objects into account. However, we observe that in real-world scenarios, data in the same cluster tend to present in a smooth manner, and previous algorithms never take this into account. Based on this observation, we propose a novel clustering algorithm, which con-siders the smoothness of data for the first time. We first divide objects into a great many tiny clusters. Our key idea is to cluster tiny clusters, whose centers constitute smooth graphs. Theoretical analysis and experimental results show that our clustering algorithm significantly outperforms state of the art. Although in this paper, we singly focus on multi-scale situations, the idea of data smoothness can certainly be extended to any clustering algorithms

Hengyu Zhao, Yubo Zhang, Pingfan Meng et al.

Recently, autonomous driving development ignited competition among car makers and technical corporations. Low-level automation cars are already commercially available. But high automated vehicles where the vehicle drives by itself without human monitoring is still at infancy. Such autonomous vehicles (AVs) rely on the computing system in the car to to interpret the environment and make driving decisions. Therefore, computing system design is essential particularly in enhancing the attainment of driving safety. However, to our knowledge, no clear guideline exists so far regarding safety-aware AV computing system and architecture design. To understand the safety requirement of AV computing system, we performed a field study by running industrial Level-4 autonomous driving fleets in various locations, road conditions, and traffic patterns. The field study indicates that traditional computing system performance metrics, such as tail latency, average latency, maximum latency, and timeout, cannot fully satisfy the safety requirement for AV computing system design. To address this issue, we propose a `safety score' as a primary metric for measuring the level of safety in AV computing system design. Furthermore, we propose a perception latency model, which helps architects estimate the safety score of given architecture and system design without physically testing them in an AV. We demonstrate the use of our safety score and latency model, by developing and evaluating a safety-aware AV computing system computation hardware resource management scheme.

Yubo Zhang, Pavel Tokmakov, Martial Hebert et al.

The recent introduction of the AVA dataset for action detection has caused a renewed interest to this problem. Several approaches have been recently proposed that improved the performance. However, all of them have ignored the main difficulty of the AVA dataset - its realistic distribution of training and test examples. This dataset was collected by exhaustive annotation of human action in uncurated videos. As a result, the most common categories, such as `stand' or `sit', contain tens of thousands of examples, whereas rare ones have only dozens. In this work we study the problem of action detection in a highly-imbalanced dataset. Differently from previous work on handling long-tail category distributions, we begin by analyzing the imbalance in the test set. We demonstrate that the standard AP metric is not informative for the categories in the tail, and propose an alternative one - Sampled AP. Armed with this new measure, we study the problem of transferring representations from the data-rich head to the rare tail categories and propose a simple but effective approach.

Yubo Zhang, Pavel Tokmakov, Martial Hebert et al.

A dominant paradigm for learning-based approaches in computer vision is training generic models, such as ResNet for image recognition, or I3D for video understanding, on large datasets and allowing them to discover the optimal representation for the problem at hand. While this is an obviously attractive approach, it is not applicable in all scenarios. We claim that action detection is one such challenging problem - the models that need to be trained are large, and labeled data is expensive to obtain. To address this limitation, we propose to incorporate domain knowledge into the structure of the model, simplifying optimization. In particular, we augment a standard I3D network with a tracking module to aggregate long term motion patterns, and use a graph convolutional network to reason about interactions between actors and objects. Evaluated on the challenging AVA dataset, the proposed approach improves over the I3D baseline by 5.5% mAP and over the state-of-the-art by 4.8% mAP.

Yubo Zhang, Vishnu Naresh Boddeti, Kris M. Kitani

Accurately identifying hands in images is a key sub-task for human activity understanding with wearable first-person point-of-view cameras. Traditional hand segmentation approaches rely on a large corpus of manually labeled data to generate robust hand detectors. However, these approaches still face challenges as the appearance of the hand varies greatly across users, tasks, environments or illumination conditions. A key observation in the case of many wearable applications and interfaces is that, it is only necessary to accurately detect the user's hands in a specific situational context. Based on this observation, we introduce an interactive approach to learn a person-specific hand segmentation model that does not require any manually labeled training data. Our approach proceeds in two steps, an interactive bootstrapping step for identifying moving hand regions, followed by learning a personalized user specific hand appearance model. Concretely, our approach uses two convolutional neural networks: (1) a gesture network that uses pre-defined motion information to detect the hand region; and (2) an appearance network that learns a person specific model of the hand region based on the output of the gesture network. During training, to make the appearance network robust to errors in the gesture network, the loss function of the former network incorporates the confidence of the gesture network while learning. Experiments demonstrate the robustness of our approach with an F1 score over 0.8 on all challenging datasets across a wide range of illumination and hand appearance variations, improving over a baseline approach by over 10%.