Yujun Li

Tao Li, Zhengbao He, Yujun Li et al.

Fine-tuning large-scale pre-trained models is prohibitively expensive in terms of computation and memory costs. Low-Rank Adaptation (LoRA), a popular Parameter-Efficient Fine-Tuning (PEFT) method, offers an efficient solution by optimizing only low-rank matrices. Despite recent progress in improving LoRA's performance, the relationship between the LoRA optimization space and the full parameter space is often overlooked. A solution that appears flat in the loss landscape of the LoRA space may still exhibit sharp directions in the full parameter space, potentially compromising generalization. We introduce Flat-LoRA, which aims to identify a low-rank adaptation situated in a flat region of the full parameter space. Instead of adopting the well-established sharpness-aware minimization approach, which incurs significant computation and memory overheads, we employ a Bayesian expectation loss objective to preserve training efficiency. Further, we design a refined random perturbation generation strategy for improved performance and carefully manage memory overhead using random seeds. Experiments across diverse tasks-including mathematical reasoning, coding abilities, dialogue generation, instruction following, and text-to-image generation-demonstrate that Flat-LoRA improves both in-domain and out-of-domain generalization. Code is available at https://github.com/nblt/Flat-LoRA.

Yujun Li, Kaichun Mo, Haishan Ye

Kaczmarz algorithm is an efficient iterative algorithm to solve overdetermined consistent system of linear equations. During each updating step, Kaczmarz chooses a hyperplane based on an individual equation and projects the current estimate for the exact solution onto that space to get a new estimate. Many vairants of Kaczmarz algorithms are proposed on how to choose better hyperplanes. Using the property of randomly sampled data in high-dimensional space, we propose an accelerated algorithm based on clustering information to improve block Kaczmarz and Kaczmarz via Johnson-Lindenstrauss lemma. Additionally, we theoretically demonstrate convergence improvement on block Kaczmarz algorithm.

Haishan Ye, Yujun Li, Zhihua Zhang

Prior optimal CUR decomposition and near optimal column reconstruction methods have been established by combining BSS sampling and adaptive sampling. In this paper, we propose a new approach to the optimal CUR decomposition and near optimal column reconstruction by just using leverage score sampling. In our approach, both the BSS sampling and adaptive sampling are not needed. Moreover, our approach is the first $O(\mathrm{nnz}(\A))$ optimal CUR algorithm where $\A$ is a data matrix in question. We also extend our approach to the Nystr{ö}m method, obtaining a fast algorithm which runs $\tilde{O}(n^{2})$ or $O(\mathrm{\nnz}(\A))$

Songhao Yang, Ruixin Shen, Jin Shu et al.

Existing sub-/super-synchronous (SSO) suppression methods for the direct-drive permanent magnet synchronous generators (D-PMSG) integrated power systems are mainly achieved by external devices or sub-synchronous resonance damping controller (SSRDC) at the converters, facing challenges of considerable control costs, complex parameters tuning, or inadaptability to various operating conditions. To address these problems, this paper proposes an adaptive SSRDC based on the phase-locked loop (PLL) for D-PMSG integrated power systems. Firstly, the PLL parameter is found critical to SSO suppression by a comprehensive sensitivity analysis on the dominant poles of the impedance closed-loop transfer function. Motivated by this finding, this paper then designs a PLL-based SSRDC, which features a simple structure, easy parameter tuning, and flexible adaptability to various operating modes. The simplicity in structure is guaranteed by the avoidance of phase compensation. Benefiting from the simple structure, only one key parameter needs to be tuned. Moreover, two principles of parameter tuning are proposed to enhance the efficiency, robustness, and adaptability of the proposed SSRDC. The controller-hardware-in-the-loop (CHIL) tests verify the validity of the proposed SSRDC under various operating conditions. Finally, some concerns about this method such as frequency estimation, computational efficiency and potential impacts on PLL are thoroughly analyzed and clarified.

Zhou Liu, Yujun Li, Zhengying Liu et al.

Automatic theorem proving with deep learning methods has attracted attentions recently. In this paper, we construct an automatic proof system for trigonometric identities. We define the normalized form of trigonometric identities, design a set of rules for the proof and put forward a method which can generate theoretically infinite trigonometric identities. Our goal is not only to complete the proof, but to complete the proof in as few steps as possible. For this reason, we design a model to learn proof data generated by random BFS (rBFS), and it is proved theoretically and experimentally that the model can outperform rBFS after a simple imitation learning. After further improvement through reinforcement learning, we get AutoTrig, which can give proof steps for identities in almost as short steps as BFS (theoretically shortest method), with a time cost of only one-thousandth. In addition, AutoTrig also beats Sympy, Matlab and human in the synthetic dataset, and performs well in many generalization tasks.

Yujun Li, Hongyuan Zhang, Yuan Yuan

Graph contrastive learning (GCL) aims to learn representations from unlabeled graph data in a self-supervised manner and has developed rapidly in recent years. However, edgelevel contrasts are not well explored by most existing GCL methods. Most studies in GCL only regard edges as auxiliary information while updating node features. One of the primary obstacles of edge-based GCL is the heavy computation burden. To tackle this issue, we propose a model that can efficiently learn edge features for GCL, namely AugmentationFree Edge Contrastive Learning (AFECL) to achieve edgeedge contrast. AFECL depends on no augmentation consisting of two parts. Firstly, we design a novel edge feature generation method, where edge features are computed by embedding concatenation of their connected nodes. Secondly, an edge contrastive learning scheme is developed, where edges connecting the same nodes are defined as positive pairs, and other edges are defined as negative pairs. Experimental results show that compared with recent state-of-the-art GCL methods or even some supervised GNNs, AFECL achieves SOTA performance on link prediction and semi-supervised node classification of extremely scarce labels. The source code is available at https://github.com/YujunLi361/AFECL.

Yujun Li, Hongyuan Zhang, Yuan Yuan

Group Relative Policy Optimization (GRPO) has recently shown strong performance in post-training large language models and vision-language models. It raises a question of whether the GRPO also significantly promotes the test-time adaptation (TTA) of vision language models. In this paper, we propose Group Relative Policy Optimization for Test-Time Adaptation (GRPO-TTA), which adapts GRPO to the TTA setting by reformulating class-specific prompt prediction as a group-wise policy optimization problem. Specifically, we construct output groups by sampling top-K class candidates from CLIP similarity distributions, enabling probability-driven optimization without access to ground-truth labels. Moreover, we design reward functions tailored to test-time adaptation, including alignment rewards and dispersion rewards, to guide effective visual encoder tuning. Extensive experiments across diverse benchmarks demonstrate that GRPO-TTA consistently outperforms existing test-time adaptation methods, with notably larger performance gains under natural distribution shifts.

Yichun Yin, Wenyong Huang, Kaikai Song et al.

We present Pangu Ultra, a Large Language Model (LLM) with 135 billion parameters and dense Transformer modules trained on Ascend Neural Processing Units (NPUs). Although the field of LLM has been witnessing unprecedented advances in pushing the scale and capability of LLM in recent years, training such a large-scale model still involves significant optimization and system challenges. To stabilize the training process, we propose depth-scaled sandwich normalization, which effectively eliminates loss spikes during the training process of deep models. We pre-train our model on 13.2 trillion diverse and high-quality tokens and further enhance its reasoning capabilities during post-training. To perform such large-scale training efficiently, we utilize 8,192 Ascend NPUs with a series of system optimizations. Evaluations on multiple diverse benchmarks indicate that Pangu Ultra significantly advances the state-of-the-art capabilities of dense LLMs such as Llama 405B and Mistral Large 2, and even achieves competitive results with DeepSeek-R1, whose sparse model structure contains much more parameters. Our exploration demonstrates that Ascend NPUs are capable of efficiently and effectively training dense models with more than 100 billion parameters. Our model and system will be available for our commercial customers.

Yue Min, Ziyun Qiao, Ruining Chen et al.

LLM pre-training efficacy increasingly depends on data composition rather than sheer volume. Yet, optimal mixing is hindered by categorization flaws: human taxonomies suffer from ontological misalignment, and Euclidean clustering fails to address embedding anisotropy. We introduce GEM (Geometric Entropy Mixing), a framework reformulating data curation as a variational problem on the hypersphere augmented with a mixing-balance regularizer. By decoupling the generative prior and optimizing the objective via a provable MM (Minorize-Maximize) algorithm, GEM effectively counteracts the cluster collapse to discover balanced semantic structures invisible to Euclidean heuristics. We employ teacher-student distillation to scale this geometric fidelity to web-scale corpora and introduce the Geometric Influence Score (GIS) for interpretable taxonomy generation. Experiments with 1.1B-parameter models demonstrate that GEM establishes a new state-of-the-art when integrated into mixing strategies like DoReMi and RegMix, improving average downstream accuracy by up to 1.2% and offering a robust coordinate system for predictable data mixing.

Yehui Tang, Yichun Yin, Yaoyuan Wang et al.

Sparse large language models (LLMs) with Mixture of Experts (MoE) and close to a trillion parameters are dominating the realm of most capable language models. However, the massive model scale poses significant challenges for the underlying software and hardware systems. In this paper, we aim to uncover a recipe to harness such scale on Ascend NPUs. The key goals are better usage of the computing resources under the dynamic sparse model structures and materializing the expected performance gain on the actual hardware. To select model configurations suitable for Ascend NPUs without repeatedly running the expensive experiments, we leverage simulation to compare the trade-off of various model hyperparameters. This study led to Pangu Ultra MoE, a sparse LLM with 718 billion parameters, and we conducted experiments on the model to verify the simulation results. On the system side, we dig into Expert Parallelism to optimize the communication between NPU devices to reduce the synchronization overhead. We also optimize the memory efficiency within the devices to further reduce the parameter and activation management overhead. In the end, we achieve an MFU of 30.0% when training Pangu Ultra MoE, with performance comparable to that of DeepSeek R1, on 6K Ascend NPUs, and demonstrate that the Ascend system is capable of harnessing all the training stages of the state-of-the-art language models. Extensive experiments indicate that our recipe can lead to efficient training of large-scale sparse language models with MoE. We also study the behaviors of such models for future reference.

Xinwei Guo, Yujun Li, Yafeng Peng et al.

As AIGC has impacted our society profoundly in the past years, ethical issues have received tremendous attention. The most urgent one is the AIGC copyright dilemma, which can immensely stifle the development of AIGC and greatly cost the entire society. Given the complexity of AIGC copyright governance and the fact that no perfect solution currently exists, previous work advocated copyleft on AI governance but without substantive analysis. In this paper, we take a step further to explore the feasibility of copyleft to alleviate the AIGC copyright dilemma. We conduct a mixed-methods study from two aspects: qualitatively, we use a formal what-if analysis to clarify the dilemma and provide case studies to show the feasibility of copyleft; quantitatively, we perform a carefully designed survey to find out how the public feels about copylefting AIGC. The key findings include: a) people generally perceive the dilemma, b) they prefer to use authorized AIGC under loose restriction, and c) they are positive to copyleft in AIGC and willing to use it in the future.

Haowei Lin, Haotian Ye, Wenzheng Feng et al.

Discovering scaling laws for predicting model performance at scale is a fundamental and open-ended challenge, mostly reliant on slow, case specific human experimentation. To investigate the potential for LLMs to automate this process, we collect over 5,000 experiments from existing literature and curate seven diverse scaling law discovery tasks. While existing agents struggle to produce accurate law formulas, this paper introduces SLDAgent, an evolution-based agent that co-optimize the scaling law model and the parameters, enabling it to autonomously explore complex relationships between variables. For the first time, we demonstrates that SLDAgent can automatically discover laws that exhibit consistently more accurate extrapolation than their established, human-derived counterparts across all tasks. Through comprehensive analysis, we elucidate why these discovered laws are superior and verify their practical utility in both pretraining and finetuning applications. This work establishes a new paradigm for agentic scientific discovery, showing that AI systems can understand their own scaling behavior, and can contribute novel and practical knowledge back to the research community.

Jieyu Yuan, Yujun Li, Yuanlin Zhang et al.

Novel view synthesis for underwater scene reconstruction presents unique challenges due to complex light-media interactions. Optical scattering and absorption in water body bring inhomogeneous medium attenuation interference that disrupts conventional volume rendering assumptions of uniform propagation medium. While 3D Gaussian Splatting (3DGS) offers real-time rendering capabilities, it struggles with underwater inhomogeneous environments where scattering media introduce artifacts and inconsistent appearance. In this study, we propose a physics-based framework that disentangles object appearance from water medium effects through tailored Gaussian modeling. Our approach introduces appearance embeddings, which are explicit medium representations for backscatter and attenuation, enhancing scene consistency. In addition, we propose a distance-guided optimization strategy that leverages pseudo-depth maps as supervision with depth regularization and scale penalty terms to improve geometric fidelity. By integrating the proposed appearance and medium modeling components via an underwater imaging model, our approach achieves both high-quality novel view synthesis and physically accurate scene restoration. Experiments demonstrate our significant improvements in rendering quality and restoration accuracy over existing methods. The project page is available at https://bilityniu.github.io/3D-UIR.

Li Zhou, Minhuan Huang, Yujun Li et al.

With the continuous extension of the Industrial Internet, cyber incidents caused by software vulnerabilities have been increasing in recent years. However, software vulnerabilities detection is still heavily relying on code review done by experts, and how to automatedly detect software vulnerabilities is an open problem so far. In this paper, we propose a novel solution named GraphEye to identify whether a function of C/C++ code has vulnerabilities, which can greatly alleviate the burden of code auditors. GraphEye is originated from the observation that the code property graph of a non-vulnerable function naturally differs from the code property graph of a vulnerable function with the same functionality. Hence, detecting vulnerable functions is attributed to the graph classification problem.GraphEye is comprised of VecCPG and GcGAT. VecCPG is a vectorization for the code property graph, which is proposed to characterize the key syntax and semantic features of the corresponding source code. GcGAT is a deep learning model based on the graph attention graph, which is proposed to solve the graph classification problem according to VecCPG. Finally, GraphEye is verified by the SARD Stack-based Buffer Overflow, Divide-Zero, Null Pointer Deference, Buffer Error, and Resource Error datasets, the corresponding F1 scores are 95.6%, 95.6%,96.1%,92.6%, and 96.1% respectively, which validate the effectiveness of the proposed solution.

Luo Luo, Yujun Li, Cheng Chen

We study the smooth minimax optimization problem $\min_{\bf x}\max_{\bf y} f({\bf x},{\bf y})$, where $f$ is $\ell$-smooth, strongly-concave in ${\bf y}$ but possibly nonconvex in ${\bf x}$. Most of existing works focus on finding the first-order stationary points of the function $f({\bf x},{\bf y})$ or its primal function $P({\bf x})\triangleq \max_{\bf y} f({\bf x},{\bf y})$, but few of them focus on achieving second-order stationary points. In this paper, we propose a novel approach for minimax optimization, called Minimax Cubic Newton (MCN), which could find an $\big(\varepsilon,κ^{1.5}\sqrt{ρ\varepsilon}\,\big)$-second-order stationary point of $P({\bf x})$ with calling ${\mathcal O}\big(κ^{1.5}\sqrtρ\varepsilon^{-1.5}\big)$ times of second-order oracles and $\tilde{\mathcal O}\big(κ^{2}\sqrtρ\varepsilon^{-1.5}\big)$ times of first-order oracles, where $κ$ is the condition number and $ρ$ is the Lipschitz continuous constant for the Hessian of $f({\bf x},{\bf y})$. In addition, we propose an inexact variant of MCN for high-dimensional problems to avoid calling expensive second-order oracles. Instead, our method solves the cubic sub-problem inexactly via gradient descent and matrix Chebyshev expansion. This strategy still obtains the desired approximate second-order stationary point with high probability but only requires $\tilde{\mathcal O}\big(κ^{1.5}\ell\varepsilon^{-2}\big)$ Hessian-vector oracle calls and $\tilde{\mathcal O}\big(κ^{2}\sqrtρ\varepsilon^{-1.5}\big)$ first-order oracle calls. To the best of our knowledge, this is the first work that considers the non-asymptotic convergence behavior of finding second-order stationary points for minimax problems without the convex-concave assumptions.

Yimin Huang, Yujun Li, Hanrong Ye et al.

The evaluation of hyperparameters, neural architectures, or data augmentation policies becomes a critical model selection problem in advanced deep learning with a large hyperparameter search space. In this paper, we propose an efficient and robust bandit-based algorithm called Sub-Sampling (SS) in the scenario of hyperparameter search evaluation. It evaluates the potential of hyperparameters by the sub-samples of observations and is theoretically proved to be optimal under the criterion of cumulative regret. We further combine SS with Bayesian Optimization and develop a novel hyperparameter optimization algorithm called BOSS. Empirical studies validate our theoretical arguments of SS and demonstrate the superior performance of BOSS on a number of applications, including Neural Architecture Search (NAS), Data Augmentation (DA), Object Detection (OD), and Reinforcement Learning (RL).

Luo Luo, Cheng Chen, Yujun Li et al.

We consider saddle point problems which objective functions are the average of $n$ strongly convex-concave individual components. Recently, researchers exploit variance reduction methods to solve such problems and achieve linear-convergence guarantees. However, these methods have a slow convergence when the condition number of the problem is very large. In this paper, we propose a stochastic proximal point algorithm, which accelerates the variance reduction method SAGA for saddle point problems. Compared with the catalyst framework, our algorithm reduces a logarithmic term of condition number for the iteration complexity. We adopt our algorithm to policy evaluation and the empirical results show that our method is much more efficient than state-of-the-art methods.