Zijun Gao

Zihao Luo, Xiangde Luo, Zijun Gao et al.

Deep learning models have exhibited remarkable efficacy in accurately delineating the prostate for diagnosis and treatment of prostate diseases, but challenges persist in achieving robust generalization across different medical centers. Source-free Domain Adaptation (SFDA) is a promising technique to adapt deep segmentation models to address privacy and security concerns while reducing domain shifts between source and target domains. However, recent literature indicates that the performance of SFDA remains far from satisfactory due to unpredictable domain gaps. Annotating a few target domain samples is acceptable, as it can lead to significant performance improvement with a low annotation cost. Nevertheless, due to extremely limited annotation budgets, careful consideration is needed in selecting samples for annotation. Inspired by this, our goal is to develop Active Source-free Domain Adaptation (ASFDA) for medical image segmentation. Specifically, we propose a novel Uncertainty-guided Tiered Self-training (UGTST) framework, consisting of efficient active sample selection via entropy-based primary local peak filtering to aggregate global uncertainty and diversity-aware redundancy filter, coupled with a tiered self-learning strategy, achieves stable domain adaptation. Experimental results on cross-center prostate MRI segmentation datasets revealed that our method yielded marked advancements, with a mere 5% annotation, exhibiting an average Dice score enhancement of 9.78% and 7.58% in two target domains compared with state-of-the-art methods, on par with fully supervised learning. Code is available at:https://github.com/HiLab-git/UGTST

Zijun Gao, Jun Wang, Guoxian Yu et al.

Existing Cross Modal Hashing (CMH) methods are mainly designed for balanced data, while imbalanced data with long-tail distribution is more general in real-world. Several long-tail hashing methods have been proposed but they can not adapt for multi-modal data, due to the complex interplay between labels and individuality and commonality information of multi-modal data. Furthermore, CMH methods mostly mine the commonality of multi-modal data to learn hash codes, which may override tail labels encoded by the individuality of respective modalities. In this paper, we propose LtCMH (Long-tail CMH) to handle imbalanced multi-modal data. LtCMH firstly adopts auto-encoders to mine the individuality and commonality of different modalities by minimizing the dependency between the individuality of respective modalities and by enhancing the commonality of these modalities. Then it dynamically combines the individuality and commonality with direct features extracted from respective modalities to create meta features that enrich the representation of tail labels, and binaries meta features to generate hash codes. LtCMH significantly outperforms state-of-the-art baselines on long-tail datasets and holds a better (or comparable) performance on datasets with balanced labels.

Jia Fu, Litingyu Wang, He Li et al.

Accurate delineation of Gross Tumor Volume (GTV), Lymph Node Clinical Target Volume (LN CTV), and Organ-at-Risk (OAR) from Computed Tomography (CT) scans is essential for precise radiotherapy planning in Nasopharyngeal Carcinoma (NPC). Building upon SegRap2023, which focused on OAR and GTV segmentation using single-center paired non-contrast CT (ncCT) and contrast-enhanced CT (ceCT) scans, the SegRap2025 challenge aims to enhance the generalizability and robustness of segmentation models across imaging centers and modalities. SegRap2025 comprises two tasks: Task01 addresses GTV segmentation using paired CT from the SegRap2023 dataset, with an additional external testing set to evaluate cross-center generalization, and Task02 focuses on LN CTV segmentation using multi-center training data and an unseen external testing set, where each case contains paired CT scans or a single modality, emphasizing both cross-center and cross-modality robustness. This paper presents the challenge setup and provides a comprehensive analysis of the solutions submitted by ten participating teams. For GTV segmentation task, the top-performing models achieved average Dice Similarity Coefficient (DSC) of 74.61% and 56.79% on the internal and external testing cohorts, respectively. For LN CTV segmentation task, the highest average DSC values reached 60.24%, 60.50%, and 57.23% on paired CT, ceCT-only, and ncCT-only subsets, respectively. SegRap2025 establishes a large-scale multi-center, multi-modality benchmark for evaluating the generalization and robustness in radiotherapy target segmentation, providing valuable insights toward clinically applicable automated radiotherapy planning systems. The benchmark is available at: https://hilab-git.github.io/SegRap2025_Challenge.

Jingjie Zhang, Hanqun Cao, Haosen Shi et al.

Cyclic peptides are attractive therapeutic modalities because their closed-ring topology can improve stability and target specificity. However, de novo cyclic peptide design remains challenging for diffusion generators, as macrocyclization requires satisfying sparse, non-smooth, and compositional geometric constraints. Existing constraint-conditioned methods largely rely on inference-time guidance, which can steer samples toward desired closures but does not directly change the learned generative distribution. We propose GeoCycler, a reward-weighted diffusion alignment framework for training conditional latent diffusion models toward macrocyclization feasibility. GeoCycler introduces a type-gated stair reward that activates distance-based shaping only when prerequisite residue or linker types are satisfied, providing dense geometric feedback while avoiding misleading signals from chemically incompatible anchors. Together with positive-only reward weighting and replay-based stabilization, GeoCycler aligns a single generator across multiple cyclization topologies. On the LNR benchmark, GeoCycler improves pass@5 closure success over strong guidance-based baselines across stapled, head-to-tail, disulfide, and bicyclic settings. In particular, it improves head-to-tail success by 20.8 percentage points over CP-Composer while maintaining comparable amino-acid and backbone-dihedral statistics. These results suggest that training-time alignment to sparse geometric constraints is a promising alternative to relying solely on post hoc sampling-time correction for cyclic peptide generation.

Qi Wang, Zijun Gao, Mingxiu Sui et al.

With the rise of deep learning technology in practical applications, Convolutional Neural Networks (CNNs) have been able to assist humans in solving many real-world problems. To enhance the performance of CNNs, numerous network architectures have been explored. Some of these architectures are designed based on the accumulated experience of researchers over time, while others are designed through neural architecture search methods. The improvements made to CNNs by the aforementioned methods are quite significant, but most of the improvement methods are limited in reality by model size and environmental constraints, making it difficult to fully realize the improved performance. In recent years, research has found that many CNN structures can be explained by the discretization of ordinary differential equations. This implies that we can design theoretically supported deep network structures using higher-order numerical difference methods. It should be noted that most of the previous CNN model structures are based on low-order numerical methods. Therefore, considering that the accuracy of linear multi-step numerical difference methods is higher than that of the forward Euler method, this paper proposes a stacking scheme based on the linear multi-step method. This scheme enhances the performance of ResNet without increasing the model size and compares it with the Runge-Kutta scheme. The experimental results show that the performance of the stacking scheme proposed in this paper is superior to existing stacking schemes (ResNet and HO-ResNet), and it has the capability to be extended to other types of neural networks.

Xiaohan Cheng, Taiyuan Mei, Yun Zi et al.

Zero sample learning is an effective method for data deficiency. The existing embedded zero sample learning methods only use the known classes to construct the embedded space, so there is an overfitting of the known classes in the testing process. This project uses category semantic similarity measures to classify multiple tags. This enables it to incorporate unknown classes that have the same meaning as currently known classes into the vector space when it is built. At the same time, most of the existing zero sample learning algorithms directly use the depth features of medical images as input, and the feature extraction process does not consider semantic information. This project intends to take ELMo-MCT as the main task and obtain multiple visual features related to the original image through self-attention mechanism. In this paper, a large number of experiments are carried out on three zero-shot learning reference datasets, and the best harmonic average accuracy is obtained compared with the most advanced algorithms.

Zijun Gao, Haibao Liu, Lingbo Li

Data Augmentation (DA) has become a critical approach in Time Series Classification (TSC), primarily for its capacity to expand training datasets, enhance model robustness, introduce diversity, and reduce overfitting. However, the current landscape of DA in TSC is plagued with fragmented literature reviews, nebulous methodological taxonomies, inadequate evaluative measures, and a dearth of accessible and user-oriented tools. This study addresses these challenges through a comprehensive examination of DA methodologies within the TSC domain.Our research began with an extensive literature review spanning a decade, revealing significant gaps in existing surveys and necessitating a detailed analysis of over 100 scholarly articles to identify more than 60 distinct DA techniques. This rigorous review led to the development of a novel taxonomy tailored to the specific needs of DA in TSC, categorizing techniques into five primary categories: Transformation-Based, Pattern-Based, Generative, Decomposition-Based, and Automated Data Augmentation. This taxonomy is intended to guide researchers in selecting appropriate methods with greater clarity. In response to the lack of comprehensive evaluations of foundational DA techniques, we conducted a thorough empirical study, testing nearly 20 DA strategies across 15 diverse datasets representing all types within the UCR time-series repository. Using ResNet and LSTM architectures, we employed a multifaceted evaluation approach, including metrics such as Accuracy, Method Ranking, and Residual Analysis, resulting in a benchmark accuracy of 84.98 +- 16.41% in ResNet and 82.41 +- 18.71% in LSTM. Our investigation underscored the inconsistent efficacies of DA techniques, for instance, methods like RGWs and Random Permutation significantly improved model performance, whereas others, like EMD, were less effective.

Wenjin Niu, Zijun Gao, Liyan Song et al.

Causal analysis has become an essential component in understanding the underlying causes of phenomena across various fields. Despite its significance, existing literature on causal discovery algorithms is fragmented, with inconsistent methodologies, i.e., there is no universal classification standard for existing methods, and a lack of comprehensive evaluations, i.e., data characteristics are often ignored to be jointly analyzed when benchmarking algorithms. This study addresses these gaps by conducting an exhaustive review and empirical evaluation for causal discovery methods on numerical data, aiming to provide a clearer and more structured understanding of the field. Our research begins with a comprehensive literature review spanning over two decades, analyzing over 200 academic articles and identifying more than 40 representative algorithms. This extensive analysis leads to the development of a structured taxonomy tailored to the complexities of causal discovery, categorizing methods into six main types. To address the lack of comprehensive evaluations, our study conducts an extensive empirical assessment of 29 causal discovery algorithms on multiple synthetic and real-world datasets. We categorize synthetic datasets based on size, linearity, and noise distribution, employing five evaluation metrics, and summarize the top-3 algorithm recommendations, providing guidelines for users in various data scenarios. Our results highlight a significant impact of dataset characteristics on algorithm performance. Moreover, a metadata extraction strategy with an accuracy exceeding 80% is developed to assist users in algorithm selection on unknown datasets. Based on these insights, we offer professional and practical guidelines to help users choose the most suitable causal discovery methods for their specific dataset.

Zhaotian Weng, Zijun Gao, Jerone Andrews et al.

Vision-language models (VLMs) pre-trained on extensive datasets can inadvertently learn biases by correlating gender information with specific objects or scenarios. Current methods, which focus on modifying inputs and monitoring changes in the model's output probability scores, often struggle to comprehensively understand bias from the perspective of model components. We propose a framework that incorporates causal mediation analysis to measure and map the pathways of bias generation and propagation within VLMs. This approach allows us to identify the direct effects of interventions on model bias and the indirect effects of interventions on bias mediated through different model components. Our results show that image features are the primary contributors to bias, with significantly higher impacts than text features, specifically accounting for 32.57% and 12.63% of the bias in the MSCOCO and PASCAL-SENTENCE datasets, respectively. Notably, the image encoder's contribution surpasses that of the text encoder and the deep fusion encoder. Further experimentation confirms that contributions from both language and vision modalities are aligned and non-conflicting. Consequently, focusing on blurring gender representations within the image encoder, which contributes most to the model bias, reduces bias efficiently by 22.03% and 9.04% in the MSCOCO and PASCAL-SENTENCE datasets, respectively, with minimal performance loss or increased computational demands.

Zijun Gao, Zhikun Xu, Xiao Ye et al.

Large language models (LLMs) often solve challenging math exercises yet fail to apply the concept right when the problem requires genuine understanding. Popular Reinforcement Learning with Verifiable Rewards (RLVR) pipelines reinforce final answers but provide little fine-grained conceptual signal, so models improve at pattern reuse rather than conceptual applications. We introduce CORE (Concept-Oriented REinforcement), an RL training framework that turns explicit concepts into a controllable supervision signal. Starting from a high-quality, low-contamination textbook resource that links verifiable exercises to concise concept descriptions, we run a sanity probe showing LLMs can restate definitions but fail concept-linked quizzes, quantifying the conceptual reasoning gap. CORE then (i) synthesizes concept-aligned quizzes, (ii) injects brief concept snippets during rollouts to elicit concept-primed trajectories, and (iii) reinforces conceptual reasoning via trajectory replacement after group failures, a lightweight forward-KL constraint that aligns unguided with concept-primed policies, or standard GRPO directly on concept-aligned quizzes. Across several models, CORE delivers consistent gains over vanilla and SFT baselines on both in-domain concept-exercise suites and diverse out-of-domain math benchmarks. CORE unifies direct training on concept-aligned quizzes and concept-injected rollouts under outcome regularization. It provides fine-grained conceptual supervision that bridges problem-solving competence and genuine conceptual reasoning, while remaining algorithm- and verifier-agnostic.

Zihao Luo, Zijun Gao, Wenjun Liao et al.

Accurate lymph node (LN) segmentation is critical in radiotherapy treatment and prognosis analysis, but is limited by the need for large annotated datasets. While deep learning-based segmentation foundation models show potential in developing high-performing models with fewer samples, their medical adaptation faces LN domain-specific prior deficiencies and inefficient few-shot fine-tuning for complex clinical practices, highlighting the necessity of an LN segmentation foundation model. In this work, we annotated 36,106 visible LNs from 3,346 publicly available head-and-neck CT scans to establish a robust LN segmentation model (nnUNetv2). Building on this, we propose Dynamic Gradient Sparsification Training (DGST), a few-shot fine-tuning approach that preserves foundational knowledge while dynamically updating the most critical parameters of the LN segmentation model with few annotations. We validate it on two publicly available LN segmentation datasets: SegRap2023 and LNQ2023. The results show that DGST outperforms existing few-shot fine-tuning methods, achieving satisfactory performance with limited labeled data. We release the dataset, models and all implementations to facilitate relevant research: https://github.com/Zihaoluoh/LN-Seg-FM.

Zhi Zhang, Qizhe Zhang, Zijun Gao et al.

With the growing size of pre-trained models, full fine-tuning and storing all the parameters for various downstream tasks is costly and infeasible. In this paper, we propose a new parameter-efficient fine-tuning method, Gradient-based Parameter Selection (GPS), demonstrating that only tuning a few selected parameters from the pre-trained model while keeping the remainder of the model frozen can generate similar or better performance compared with the full model fine-tuning method. Different from the existing popular and state-of-the-art parameter-efficient fine-tuning approaches, our method does not introduce any additional parameters and computational costs during both the training and inference stages. Another advantage is the model-agnostic and non-destructive property, which eliminates the need for any other design specific to a particular model. Compared with the full fine-tuning, GPS achieves 3.33% (91.78% vs. 88.45%, FGVC) and 9.61% (73.1% vs. 65.57%, VTAB) improvement of the accuracy with tuning only 0.36% parameters of the pre-trained model on average over 24 image classification tasks; it also demonstrates a significant improvement of 17% and 16.8% in mDice and mIoU, respectively, on medical image segmentation task. Moreover, GPS achieves state-of-the-art performance compared with existing PEFT methods.

Taiyuan Mei, Yun Zi, Xiaohan Cheng et al.

The internal structure and operation mechanism of large-scale language models are analyzed theoretically, especially how Transformer and its derivative architectures can restrict computing efficiency while capturing long-term dependencies. Further, we dig deep into the efficiency bottleneck of the training phase, and evaluate in detail the contribution of adaptive optimization algorithms (such as AdamW), massively parallel computing techniques, and mixed precision training strategies to accelerate convergence and reduce memory footprint. By analyzing the mathematical principles and implementation details of these algorithms, we reveal how they effectively improve training efficiency in practice. In terms of model deployment and inference optimization, this paper systematically reviews the latest advances in model compression techniques, focusing on strategies such as quantification, pruning, and knowledge distillation. By comparing the theoretical frameworks of these techniques and their effects in different application scenarios, we demonstrate their ability to significantly reduce model size and inference delay while maintaining model prediction accuracy. In addition, this paper critically examines the limitations of current efficiency optimization methods, such as the increased risk of overfitting, the control of performance loss after compression, and the problem of algorithm generality, and proposes some prospects for future research. In conclusion, this study provides a comprehensive theoretical framework for understanding the efficiency optimization of large-scale language models.

Ngoc Quach, Qi Wang, Zijun Gao et al.

The widespread use of knowledge graphs in various fields has brought about a challenge in effectively integrating and updating information within them. When it comes to incorporating contexts, conventional methods often rely on rules or basic machine learning models, which may not fully grasp the complexity and fluidity of context information. This research suggests an approach based on reinforcement learning (RL), specifically utilizing Deep Q Networks (DQN) to enhance the process of integrating contexts into knowledge graphs. By considering the state of the knowledge graph as environment states defining actions as operations for integrating contexts and using a reward function to gauge the improvement in knowledge graph quality post-integration, this method aims to automatically develop strategies for optimal context integration. Our DQN model utilizes networks as function approximators, continually updating Q values to estimate the action value function, thus enabling effective integration of intricate and dynamic context information. Initial experimental findings show that our RL method outperforms techniques in achieving precise context integration across various standard knowledge graph datasets, highlighting the potential and effectiveness of reinforcement learning in enhancing and managing knowledge graphs.

Hanqun Cao, Marcelo D. T. Torres, Jingjie Zhang et al.

Antimicrobial resistance (AMR) is projected to cause up to 10 million deaths annually by 2050, underscoring the urgent need for new antibiotics. Here we present ApexAmphion, a deep-learning framework for de novo design of antibiotics that couples a 6.4-billion-parameter protein language model with reinforcement learning. The model is first fine-tuned on curated peptide data to capture antimicrobial sequence regularities, then optimised with proximal policy optimization against a composite reward that combines predictions from a learned minimum inhibitory concentration (MIC) classifier with differentiable physicochemical objectives. In vitro evaluation of 100 designed peptides showed low MIC values (nanomolar range in some cases) for all candidates (100% hit rate). Moreover, 99 our of 100 compounds exhibited broad-spectrum antimicrobial activity against at least two clinically relevant bacteria. The lead molecules killed bacteria primarily by potently targeting the cytoplasmic membrane. By unifying generation, scoring and multi-objective optimization with deep reinforcement learning in a single pipeline, our approach rapidly produces diverse, potent candidates, offering a scalable route to peptide antibiotics and a platform for iterative steering toward potency and developability within hours.

Hanqun Cao, Xinyi Zhou, Zijun Gao et al.

Protein structure prediction often hinges on multiple sequence alignments (MSAs), which underperform on low-homology and orphan proteins. We introduce PLAME, a lightweight MSA design framework that leverages evolutionary embeddings from pretrained protein language models to generate MSAs that better support downstream folding. PLAME couples these embeddings with a conservation--diversity loss that balances agreement on conserved positions with coverage of plausible sequence variation. Beyond generation, we develop (i) an MSA selection strategy to filter high-quality candidates and (ii) a sequence-quality metric that is complementary to depth-based measures and predictive of folding gains. On AlphaFold2 low-homology/orphan benchmarks, PLAME delivers state-of-the-art improvements in structure accuracy (e.g., lDDT/TM-score), with consistent gains when paired with AlphaFold3. Ablations isolate the benefits of the selection strategy, and case studies elucidate how MSA characteristics shape AlphaFold confidence and error modes. Finally, we show PLAME functions as a lightweight adapter, enabling ESMFold to approach AlphaFold2-level accuracy while retaining ESMFold-like inference speed. PLAME thus provides a practical path to high-quality folding for proteins lacking strong evolutionary neighbors.

Junde Xu, Zijun Gao, Xinyi Zhou et al.

The inverse folding problem, aiming to design amino acid sequences that fold into desired three-dimensional structures, is pivotal for various biotechnological applications. Here, we introduce a novel approach leveraging Direct Preference Optimization (DPO) to fine-tune an inverse folding model using feedback from a protein folding model. Given a target protein structure, we begin by sampling candidate sequences from the inverse-folding model, then predict the three-dimensional structure of each sequence with the folding model to generate pairwise structural-preference labels. These labels are used to fine-tune the inverse-folding model under the DPO objective. Our results on the CATH 4.2 test set demonstrate that DPO fine-tuning not only improves sequence recovery of baseline models but also leads to a significant improvement in average TM-Score from 0.77 to 0.81, indicating enhanced structure similarity. Furthermore, iterative application of our DPO-based method on challenging protein structures yields substantial gains, with an average TM-Score increase of 79.5\% with regard to the baseline model. This work establishes a promising direction for enhancing protein sequence design ability from structure feedback by effectively utilizing preference optimization.

Zijun Gao, Yan Sun, Han Su

Generative models have achieved remarkable success across a range of applications, yet their evaluation still lacks principled uncertainty quantification. In this paper, we develop a method for comparing how close different generative models are to the underlying distribution of test samples. Particularly, our approach employs the Kullback-Leibler (KL) divergence to measure the distance between a generative model and the unknown test distribution, as KL requires no tuning parameters such as the kernels used by RKHS-based distances, and is the only $f$-divergence that admits a crucial cancellation to enable the uncertainty quantification. Furthermore, we extend our method to comparing conditional generative models and leverage Edgeworth expansions to address limited-data settings. On simulated datasets with known ground truth, we show that our approach realizes effective coverage rates, and has higher power compared to kernel-based methods. When applied to generative models on image and text datasets, our procedure yields conclusions consistent with benchmark metrics but with statistical confidence.

Jiayi Guo, Zijun Gao

We study the problem of selecting the best heterogeneous treatment effect (HTE) estimator from a collection of candidates in settings where the treatment effect is fundamentally unobserved. We cast estimator selection as a multiple testing problem and introduce a ground-truth-free procedure based on a cross-fitted, exponentially weighted test statistic. A key component of our method is a two-way sample splitting scheme that decouples nuisance estimation from weight learning and ensures the stability required for valid inference. Leveraging a stability-based central limit theorem, we establish asymptotic familywise error rate control under mild regularity conditions. Empirically, our procedure provides reliable error control while substantially reducing false selections compared with commonly used methods across ACIC 2016, IHDP, and Twins benchmarks, demonstrating that our method is feasible and powerful even without ground-truth treatment effects.

Zijun Gao, Qingyuan Zhao

Existing tools for explaining complex models and systems are associational rather than causal and do not provide mechanistic understanding. We propose a new notion called counterfactual explainability for causal attribution that is motivated by the concept of genetic heritability in twin studies. Counterfactual explainability extends methods for global sensitivity analysis (including the functional analysis of variance and Sobol's indices), which assumes independent explanatory variables, to dependent explanations by using a directed acyclic graphs to describe their causal relationship. Therefore, this explanability measure directly incorporates causal mechanisms by construction. Under a comonotonicity assumption, we discuss methods for estimating counterfactual explainability and apply them to a real dataset dataset to explain income inequality by gender, race, and educational attainment.

Heming Yao, Ryan W. Stidham, Zijun Gao et al.

Optical colonoscopy is an essential diagnostic and prognostic tool for many gastrointestinal diseases, including cancer screening and staging, intestinal bleeding, diarrhea, abdominal symptom evaluation, and inflammatory bowel disease assessment. Automated assessment of colonoscopy is of interest considering the subjectivity present in qualitative human interpretations of colonoscopy findings. Localization of the camera is essential to interpreting the meaning and context of findings for diseases evaluated by colonoscopy. In this study, we propose a camera localization system to estimate the relative location of the camera and classify the colon into anatomical segments. The camera localization system begins with non-informative frame detection and removal. Then a self-training end-to-end convolutional neural network is built to estimate the camera motion, where several strategies are proposed to improve its robustness and generalization on endoscopic videos. Using the estimated camera motion a camera trajectory can be derived and a relative location index calculated. Based on the estimated location index, anatomical colon segment classification is performed by constructing a colon template. The proposed motion estimation algorithm was evaluated on an external dataset containing the ground truth for camera pose. The experimental results show that the performance of the proposed method is superior to other published methods. The relative location index estimation and anatomical region classification were further validated using colonoscopy videos collected from routine clinical practice. This validation yielded an average accuracy in classification of 0.754, which is substantially higher than the performances obtained using location indices built from other methods.

Zijun Gao, Yanjun Han, Zhimei Ren et al.

In this paper, we study the multi-armed bandit problem in the batched setting where the employed policy must split data into a small number of batches. While the minimax regret for the two-armed stochastic bandits has been completely characterized in \cite{perchet2016batched}, the effect of the number of arms on the regret for the multi-armed case is still open. Moreover, the question whether adaptively chosen batch sizes will help to reduce the regret also remains underexplored. In this paper, we propose the BaSE (batched successive elimination) policy to achieve the rate-optimal regrets (within logarithmic factors) for batched multi-armed bandits, with matching lower bounds even if the batch sizes are determined in an adaptive manner.