arXiv subjects

Lijun Zhang

Publications and source records attributed to Lijun Zhang.

At least 73 records · Page 4Linked to original sources

arXiv2025

Continuous Subspace Optimization for Continual Learning

Continual learning aims to learn multiple tasks sequentially while preserving prior knowledge, but faces the challenge of catastrophic forgetting when adapting to new tasks. Recently, approaches leveraging pre-trained models have gained increasing popularity in mitigating this issue, due to the strong generalization ability of foundation models. To adjust pre-trained models for new tasks, existing methods usually employ low-rank adaptation, which restricts parameter updates to a fixed low-rank subspace. However, constraining the optimization space inherently compromises the model's learning capacity, resulting in inferior performance. To address this limitation, we propose Continuous Subspace Optimization for Continual Learning (CoSO) to fine-tune the model in a series of subspaces rather than a single one. These sequential subspaces are dynamically determined through the singular value decomposition of the gradients. CoSO updates the model by projecting gradients onto these subspaces, ensuring memory-efficient optimization. To mitigate forgetting, the optimization subspace of each task is constrained to be orthogonal to the historical task subspace. During task learning, CoSO maintains a task-specific component that captures the critical update directions for the current task. Upon completing a task, this component is used to update the historical task subspace, laying the groundwork for subsequent learning. Extensive experiments on multiple datasets demonstrate that CoSO significantly outperforms state-of-the-art methods, especially in challenging scenarios with long task sequences.

cs.CV↗

arXiv2025

Non-stationary Delayed Online Convex Optimization: From Full-information to Bandit Setting

Although online convex optimization (OCO) under arbitrary delays has received increasing attention recently, previous studies focus on stationary environments with the goal of minimizing static regret. In this paper, we investigate the delayed OCO in non-stationary environments, and choose dynamic regret with respect to any sequence of comparators as the performance metric. To this end, we first propose an algorithm called Mild-OGD for the full-information case, where delayed gradients are available. The basic idea is to maintain multiple experts in parallel, each performing a gradient descent step with different learning rates for every delayed gradient according to their arrival order, and utilize a meta-algorithm to track the best one based on their delayed performance. Despite the simplicity of this idea, our novel analysis shows that the dynamic regret of Mild-OGD can be automatically bounded by $O(\sqrt{\bar{d}T(P_T+1)})$ under the in-order assumption and $O(\sqrt{dT(P_T+1)})$ in the worst case, where $\bar{d}$ and $d$ denote the average and maximum delay respectively, $T$ is the time horizon, and $P_T$ is the path-length of comparators. Moreover, we demonstrate that the result in the worst case is optimal by deriving a matching lower bound. Finally, we develop a bandit variant of Mild-OGD for a more challenging case with only delayed loss values. Interestingly, we prove that under a relatively large amount of delay, our bandit algorithm even enjoys the best dynamic regret bound of existing non-delayed bandit algorithms.

cs.LG↗

arXiv2025

SelaVPR++: Towards Seamless Adaptation of Foundation Models for Efficient Place Recognition

Recent studies show that the visual place recognition (VPR) method using pre-trained visual foundation models can achieve promising performance. In our previous work, we propose a novel method to realize seamless adaptation of foundation models to VPR (SelaVPR). This method can produce both global and local features that focus on discriminative landmarks to recognize places for two-stage VPR by a parameter-efficient adaptation approach. Although SelaVPR has achieved competitive results, we argue that the previous adaptation is inefficient in training time and GPU memory usage, and the re-ranking paradigm is also costly in retrieval latency and storage usage. In pursuit of higher efficiency and better performance, we propose an extension of the SelaVPR, called SelaVPR++. Concretely, we first design a parameter-, time-, and memory-efficient adaptation method that uses lightweight multi-scale convolution (MultiConv) adapters to refine intermediate features from the frozen foundation backbone. This adaptation method does not back-propagate gradients through the backbone during training, and the MultiConv adapter facilitates feature interactions along the spatial axes and introduces proper local priors, thus achieving higher efficiency and better performance. Moreover, we propose an innovative re-ranking paradigm for more efficient VPR. Instead of relying on local features for re-ranking, which incurs huge overhead in latency and storage, we employ compact binary features for initial retrieval and robust floating-point (global) features for re-ranking. To obtain such binary features, we propose a similarity-constrained deep hashing method, which can be easily integrated into the VPR pipeline. Finally, we improve our training strategy and unify the training protocol of several common training datasets to merge them for better training of VPR models. Extensive experiments show that ......

cs.CV↗

arXiv2025

Analyzing the Power of Chain of Thought through Memorization Capabilities

It has been shown that the chain of thought (CoT) can enhance the power of large language models (LLMs) to solve certain mathematical reasoning problems. However, the capacity of CoT is still not fully explored. As an important instance, the following basic question has not yet been answered: Does CoT expand the capability of transformers across all reasoning tasks? We demonstrate that reasoning with transformers is essentially a memorization problem for reasoning datasets. Thus, examining the power of CoT across all reasoning tasks amounts to analyzing the memorization capabilities of CoT transformers. In this paper, we give a complete description of the memorization capabilities of fixed-precision transformers with or without CoT and give a negative answer to the above-mentioned question. Precisely, we first give necessary and sufficient conditions for fixed-precision transformers with and without CoT to memorize a finite reasoning dataset and show that these two conditions do not imply each other. Then, we give lower and upper bounds for the number of parameters needed for transformers with or without CoT to memorize a finite reasoning dataset with $N$ elements, which are $\overlineΘ(N)$ in all cases. This implies that there exist reasoning tasks for which CoT does not enhance the reasoning power of transformers, leading to a negative answer to the above-mentioned question. Finally, we give the first results on memorizing infinite reasoning datasets by CoT transformers and show that some simple infinite datasets cannot be memorized by transformers with or without CoT.

cs.LG↗

arXiv2025

A Comprehensive Empirical Evaluation of Agent Frameworks on Code-centric Software Engineering Tasks

Unlike traditional automation tools or static LLM-based systems, agents combine decision-making and tool utilization to accomplish complex tasks, showing great potential in software engineering. However, existing studies largely focus on specific tasks or isolated aspects, providing an incomplete picture of agents' practical capabilities. To address this, we conduct a comprehensive empirical study evaluating seven general-purpose agent frameworks across three representative code-centric tasks: software development, vulnerability detection, and program repair. Each task is assessed using standard, widely adopted benchmarks to ensure objective and comparable evaluation. Agent performance is systematically analyzed from three complementary perspectives: effectiveness (task success), efficiency (execution process), and overhead (token consumption). Our findings reveal distinct capability patterns and trade-offs among the evaluated frameworks. In terms of effectiveness, agents achieve moderate overall performance. Regarding efficiency, AgentOrchestra tends to exhibit the longest trajectories and the most correction attempts due to coordination overhead, whereas OpenHands demonstrate stronger reflective reasoning abilities. For overhead, software development incurs the highest monetary cost, while GPTswarm remains the most cost-efficient. Furthermore, we conduct an in-depth cross-analysis of the relationship between effectiveness and efficiency, exploring the underlying reasons behind their interplay. These findings guide both practical adoption and future research toward more efficient software engineering agents.

cs.SE↗

arXiv2025

SeCon-RAG: A Two-Stage Semantic Filtering and Conflict-Free Framework for Trustworthy RAG

Retrieval-augmented generation (RAG) systems enhance large language models (LLMs) with external knowledge but are vulnerable to corpus poisoning and contamination attacks, which can compromise output integrity. Existing defenses often apply aggressive filtering, leading to unnecessary loss of valuable information and reduced reliability in generation. To address this problem, we propose a two-stage semantic filtering and conflict-free framework for trustworthy RAG. In the first stage, we perform a joint filter with semantic and cluster-based filtering which is guided by the Entity-intent-relation extractor (EIRE). EIRE extracts entities, latent objectives, and entity relations from both the user query and filtered documents, scores their semantic relevance, and selectively adds valuable documents into the clean retrieval database. In the second stage, we proposed an EIRE-guided conflict-aware filtering module, which analyzes semantic consistency between the query, candidate answers, and retrieved knowledge before final answer generation, filtering out internal and external contradictions that could mislead the model. Through this two-stage process, SeCon-RAG effectively preserves useful knowledge while mitigating conflict contamination, achieving significant improvements in both generation robustness and output trustworthiness. Extensive experiments across various LLMs and datasets demonstrate that the proposed SeCon-RAG markedly outperforms state-of-the-art defense methods.

cs.CL↗

arXiv2025

Goal-Guided Efficient Exploration via Large Language Model in Reinforcement Learning

Real-world decision-making tasks typically occur in complex and open environments, posing significant challenges to reinforcement learning (RL) agents' exploration efficiency and long-horizon planning capabilities. A promising approach is LLM-enhanced RL, which leverages the rich prior knowledge and strong planning capabilities of LLMs to guide RL agents in efficient exploration. However, existing methods mostly rely on frequent and costly LLM invocations and suffer from limited performance due to the semantic mismatch. In this paper, we introduce a Structured Goal-guided Reinforcement Learning (SGRL) method that integrates a structured goal planner and a goal-conditioned action pruner to guide RL agents toward efficient exploration. Specifically, the structured goal planner utilizes LLMs to generate a reusable, structured function for goal generation, in which goals are prioritized. Furthermore, by utilizing LLMs to determine goals' priority weights, it dynamically generates forward-looking goals to guide the agent's policy toward more promising decision-making trajectories. The goal-conditioned action pruner employs an action masking mechanism that filters out actions misaligned with the current goal, thereby constraining the RL agent to select goal-consistent policies. We evaluate the proposed method on Crafter and Craftax-Classic, and experimental results demonstrate that SGRL achieves superior performance compared to existing state-of-the-art methods.

cs.LG↗

arXiv2025

Efficient Methods for Non-stationary Online Learning

Non-stationary online learning has drawn much attention in recent years. In particular, dynamic regret and adaptive regret are proposed as two principled performance measures for online convex optimization in non-stationary environments. To optimize them, a two-layer online ensemble is usually deployed due to the inherent uncertainty of non-stationarity, in which multiple base-learners are maintained and a meta-algorithm is employed to track the best one on the fly. However, the two-layer structure raises concerns about computational complexity -- such methods typically maintain $O(\log T)$ base-learners simultaneously for a $T$-round online game and thus perform multiple projections onto the feasible domain per round, which becomes the computational bottleneck when the domain is complicated. In this paper, we present efficient methods for optimizing dynamic regret and adaptive regret that reduce the number of projections per round from $O(\log T)$ to $1$. The proposed algorithms require only one gradient query and one function evaluation at each round. Our technique hinges on the reduction mechanism developed in parameter-free online learning and requires non-trivial modifications for non-stationary online methods. Furthermore, we study an even stronger measure, namely "interval dynamic regret", and reduce the number of projections per round from $O(\log^2 T)$ to $1$ for minimizing it. Our reduction demonstrates broad generality and applies to two important applications: online stochastic control and online principal component analysis, resulting in methods that are both efficient and optimal. Finally, empirical studies verify our theoretical findings.

cs.LG↗

arXiv2025

Convergence Analysis of the Lion Optimizer in Centralized and Distributed Settings

In this paper, we analyze the convergence properties of the Lion optimizer. First, we establish that the Lion optimizer attains a convergence rate of $\mathcal{O}(d^{1/2}T^{-1/4})$ under standard assumptions, where $d$ denotes the problem dimension and $T$ is the iteration number. To further improve this rate, we introduce the Lion optimizer with variance reduction, resulting in an enhanced convergence rate of $\mathcal{O}(d^{1/2}T^{-1/3})$. We then analyze in distributed settings, where the standard and variance reduced version of the distributed Lion can obtain the convergence rates of $\mathcal{O}(d^{1/2}(nT)^{-1/4})$ and $\mathcal{O}(d^{1/2}(nT)^{-1/3})$, with $n$ denoting the number of nodes. Furthermore, we investigate a communication-efficient variant of the distributed Lion that ensures sign compression in both communication directions. By employing the unbiased sign operations, the proposed Lion variant and its variance reduction counterpart, achieve convergence rates of $\mathcal{O}\left( \max \left\{\frac{d^{1/4}}{T^{1/4}}, \frac{d^{1/10}}{n^{1/5}T^{1/5}} \right\} \right)$ and $\mathcal{O}\left( \frac{d^{1/4}}{T^{1/4}} \right)$, respectively.

cs.LG↗

arXiv2025

A Differentiated Reward Method for Reinforcement Learning based Multi-Vehicle Cooperative Decision-Making Algorithms

Reinforcement learning (RL) shows great potential for optimizing multi-vehicle cooperative driving strategies through the state-action-reward feedback loop, but it still faces challenges such as low sample efficiency. This paper proposes a differentiated reward method based on steady-state transition systems, which incorporates state transition gradient information into the reward design by analyzing traffic flow characteristics, aiming to optimize action selection and policy learning in multi-vehicle cooperative decision-making. The performance of the proposed method is validated in RL algorithms such as MAPPO, MADQN, and QMIX under varying autonomous vehicle penetration. The results show that the differentiated reward method significantly accelerates training convergence and outperforms centering reward and others in terms of traffic efficiency, safety, and action rationality. Additionally, the method demonstrates strong scalability and environmental adaptability, providing a novel approach for multi-agent cooperative decision-making in complex traffic scenarios.

cs.AI↗

arXiv2025

A Value Based Parallel Update MCTS Method for Multi-Agent Cooperative Decision Making of Connected and Automated Vehicles

To solve the problem of lateral and logitudinal joint decision-making of multi-vehicle cooperative driving for connected and automated vehicles (CAVs), this paper proposes a Monte Carlo tree search (MCTS) method with parallel update for multi-agent Markov game with limited horizon and time discounted setting. By analyzing the parallel actions in the multi-vehicle joint action space in the partial-steady-state traffic flow, the parallel update method can quickly exclude potential dangerous actions, thereby increasing the search depth without sacrificing the search breadth. The proposed method is tested in a large number of randomly generated traffic flow. The experiment results show that the algorithm has good robustness and better performance than the SOTA reinforcement learning algorithms and heuristic methods. The vehicle driving strategy using the proposed algorithm shows rationality beyond human drivers, and has advantages in traffic efficiency and safety in the coordinating zone.

cs.MA↗

arXiv2025

Dual Adaptivity: Universal Algorithms for Minimizing the Adaptive Regret of Convex Functions

To deal with changing environments, a new performance measure -- adaptive regret, defined as the maximum static regret over any interval, was proposed in online learning. Under the setting of online convex optimization, several algorithms have been successfully developed to minimize the adaptive regret. However, existing algorithms lack universality in the sense that they can only handle one type of convex functions and need apriori knowledge of parameters, which hinders their application in real-world scenarios. To address this limitation, this paper investigates universal algorithms with dual adaptivity, which automatically adapt to the property of functions (convex, exponentially concave, or strongly convex), as well as the nature of environments (stationary or changing). Specifically, we propose a meta-expert framework for dual adaptive algorithms, where multiple experts are created dynamically and aggregated by a meta-algorithm. The meta-algorithm is required to yield a second-order bound, which can accommodate unknown function types. We further incorporate the technique of sleeping experts to capture the changing environments. For the construction of experts, we introduce two strategies (increasing the number of experts or enhancing the capabilities of experts) to achieve universality. Theoretical analysis shows that our algorithms are able to minimize the adaptive regret for multiple types of convex functions simultaneously, and also allow the type of functions to switch between rounds. Moreover, we extend our meta-expert framework to online composite optimization, and develop a universal algorithm for minimizing the adaptive regret of composite functions.

cs.LG↗

arXiv2025

Online Nonsubmodular Optimization with Delayed Feedback in the Bandit Setting

We investigate the online nonsubmodular optimization with delayed feedback in the bandit setting, where the loss function is $α$-weakly DR-submodular and $β$-weakly DR-supermodular. Previous work has established an $(α,β)$-regret bound of $\mathcal{O}(nd^{1/3}T^{2/3})$, where $n$ is the dimensionality and $d$ is the maximum delay. However, its regret bound relies on the maximum delay and is thus sensitive to irregular delays. Additionally, it couples the effects of delays and bandit feedback as its bound is the product of the delay term and the $\mathcal{O}(nT^{2/3})$ regret bound in the bandit setting without delayed feedback. In this paper, we develop two algorithms to address these limitations, respectively. Firstly, we propose a novel method, namely DBGD-NF, which employs the one-point gradient estimator and utilizes all the available estimated gradients in each round to update the decision. It achieves a better $\mathcal{O}(n\bar{d}^{1/3}T^{2/3})$ regret bound, which is relevant to the average delay $\bar{d} = \frac{1}{T}\sum_{t=1}^T d_t\leq d$. Secondly, we extend DBGD-NF by employing a blocking update mechanism to decouple the joint effect of the delays and bandit feedback, which enjoys an $\mathcal{O}(n(T^{2/3} + \sqrt{dT}))$ regret bound. When $d = \mathcal{O}(T^{1/3})$, our regret bound matches the $\mathcal{O}(nT^{2/3})$ bound in the bandit setting without delayed feedback. Compared to our first $\mathcal{O}(n\bar{d}^{1/3}T^{2/3})$ bound, it is more advantageous when the maximum delay $d = o(\bar{d}^{2/3}T^{1/3})$. Finally, we conduct experiments on structured sparse learning to demonstrate the superiority of our methods.

cs.LG↗

arXiv2025

Shapley Uncertainty in Natural Language Generation

In question-answering tasks, determining when to trust the outputs is crucial to the alignment of large language models (LLMs). Kuhn et al. (2023) introduces semantic entropy as a measure of uncertainty, by incorporating linguistic invariances from the same meaning. It primarily relies on setting threshold to measure the level of semantic equivalence relation. We propose a more nuanced framework that extends beyond such thresholding by developing a Shapley-based uncertainty metric that captures the continuous nature of semantic relationships. We establish three fundamental properties that characterize valid uncertainty metrics and prove that our Shapley uncertainty satisfies these criteria. Through extensive experiments, we demonstrate that our Shapley uncertainty more accurately predicts LLM performance in question-answering and other datasets, compared to similar baseline measures.

cs.AI↗

arXiv2025

External Knowledge Injection for CLIP-Based Class-Incremental Learning

Class-Incremental Learning (CIL) enables learning systems to continuously adapt to evolving data streams. With the advancement of pre-training, leveraging pre-trained vision-language models (e.g., CLIP) offers a promising starting point for CIL. However, CLIP makes decisions by matching visual embeddings to class names, overlooking the rich contextual information conveyed through language. For instance, the concept of ``cat'' can be decomposed into features like tail, fur, and face for recognition. Besides, since the model is continually updated, these detailed features are overwritten in CIL, requiring external knowledge for compensation. In this paper, we introduce ExterNal knowledGe INjEction (ENGINE) for CLIP-based CIL. To enhance knowledge transfer from outside the dataset, we propose a dual-branch injection tuning framework that encodes informative knowledge from both visual and textual modalities. The visual branch is enhanced with data augmentation to enrich the visual features, while the textual branch leverages GPT-4 to rewrite discriminative descriptors. In addition to this on-the-fly knowledge injection, we also implement post-tuning knowledge by re-ranking the prediction results during inference. With the injected knowledge, the model can better capture informative features for downstream tasks as data evolves. Extensive experiments demonstrate the state-of-the-art performance of ENGINE. Code is available at: https://github.com/LAMDA-CL/ICCV25-ENGINE

cs.CV↗

arXiv2025

Scheduzz: Constraint-based Fuzz Driver Generation with Dual Scheduling

Fuzzing a library requires experts to understand the library usage well and craft high-quality fuzz drivers, which is tricky and tedious. Therefore, many techniques have been proposed to automatically generate fuzz drivers. However, they fail to generate rational fuzz drivers due to the lack of adherence to proper library usage conventions, such as ensuring a resource is closed after being opened. To make things worse, existing library fuzzing techniques unconditionally execute each driver, resulting in numerous irrational drivers that waste computational resources while contributing little coverage and generating false positive bug reports. To tackle these challenges, we propose a novel automatic library fuzzing technique, Scheduzz, an LLM-based library fuzzing technique. It leverages LLMs to understand rational usage of libraries and extract API combination constraints. To optimize computational resource utilization, a dual scheduling framework is implemented to efficiently manage API combinations and fuzz drivers. The framework models driver generation and the corresponding fuzzing campaign as an online optimization problem. Within the scheduling loop, multiple API combinations are selected to generate fuzz drivers, while simultaneously, various optimized fuzz drivers are scheduled for execution or suspension. We implemented Scheduzz and evaluated it in 33 real-world libraries. Compared to baseline approaches, Scheduzz significantly reduces computational overhead and outperforms UTopia on 16 out of 21 libraries. It achieves 1.62x, 1.50x, and 1.89x higher overall coverage than the state-of-the-art techniques CKGFuzzer, Promptfuzz, and the handcrafted project OSS-Fuzz, respectively. In addition, Scheduzz discovered 33 previously unknown bugs in these well-tested libraries, 3 of which have been assigned CVEs.

cs.SE↗

arXiv2025

Revisiting Physically Realizable Adversarial Object Attack against LiDAR-based Detection: Clarifying Problem Formulation and Experimental Protocols

Adversarial robustness in LiDAR-based 3D object detection is a critical research area due to its widespread application in real-world scenarios. While many digital attacks manipulate point clouds or meshes, they often lack physical realizability, limiting their practical impact. Physical adversarial object attacks remain underexplored and suffer from poor reproducibility due to inconsistent setups and hardware differences. To address this, we propose a device-agnostic, standardized framework that abstracts key elements of physical adversarial object attacks, supports diverse methods, and provides open-source code with benchmarking protocols in simulation and real-world settings. Our framework enables fair comparison, accelerates research, and is validated by successfully transferring simulated attacks to a physical LiDAR system. Beyond the framework, we offer insights into factors influencing attack success and advance understanding of adversarial robustness in real-world LiDAR perception.

cs.CV↗

arXiv2025

Robust Incomplete-Modality Alignment for Ophthalmic Disease Grading and Diagnosis via Labeled Optimal Transport

Multimodal ophthalmic imaging-based diagnosis integrates color fundus image with optical coherence tomography (OCT) to provide a comprehensive view of ocular pathologies. However, the uneven global distribution of healthcare resources often results in real-world clinical scenarios encountering incomplete multimodal data, which significantly compromises diagnostic accuracy. Existing commonly used pipelines, such as modality imputation and distillation methods, face notable limitations: 1)Imputation methods struggle with accurately reconstructing key lesion features, since OCT lesions are localized, while fundus images vary in style. 2)distillation methods rely heavily on fully paired multimodal training data. To address these challenges, we propose a novel multimodal alignment and fusion framework capable of robustly handling missing modalities in the task of ophthalmic diagnostics. By considering the distinctive feature characteristics of OCT and fundus images, we emphasize the alignment of semantic features within the same category and explicitly learn soft matching between modalities, allowing the missing modality to utilize existing modality information, achieving robust cross-modal feature alignment under the missing modality. Specifically, we leverage the Optimal Transport for multi-scale modality feature alignment: class-wise alignment through predicted class prototypes and feature-wise alignment via cross-modal shared feature transport. Furthermore, we propose an asymmetric fusion strategy that effectively exploits the distinct characteristics of OCT and fundus modalities. Extensive evaluations on three large ophthalmic multimodal datasets demonstrate our model's superior performance under various modality-incomplete scenarios, achieving Sota performance in both complete modality and inter-modality incompleteness conditions. Code is available at https://github.com/Qinkaiyu/RIMA

cs.CV↗