arXiv subjects

Zibin Zheng

Publications and source records attributed to Zibin Zheng.

At least 109 records · Page 6Linked to original sources

arXiv2025

Measuring Diversity in Synthetic Datasets

Large language models (LLMs) are widely adopted to generate synthetic datasets for various natural language processing (NLP) tasks, such as text classification and summarization. However, accurately measuring the diversity of these synthetic datasets-an aspect crucial for robust model performance-remains a significant challenge. In this paper, we introduce DCScore, a novel method for measuring synthetic dataset diversity from a classification perspective. Specifically, DCScore formulates diversity evaluation as a sample classification task, leveraging mutual relationships among samples. We further provide theoretical verification of the diversity-related axioms satisfied by DCScore, highlighting its role as a principled diversity evaluation method. Experimental results on synthetic datasets reveal that DCScore enjoys a stronger correlation with multiple diversity pseudo-truths of evaluated datasets, underscoring its effectiveness. Moreover, both empirical and theoretical evidence demonstrate that DCScore substantially reduces computational costs compared to existing methods. Code is available at: https://github.com/bluewhalelab/dcscore.

cs.CL↗

arXiv2025

Blind Spot Navigation: Evolutionary Discovery of Sensitive Semantic Concepts for LVLMs

Adversarial attacks aim to generate malicious inputs that mislead deep models, but beyond causing model failure, they cannot provide certain interpretable information such as ``\textit{What content in inputs make models more likely to fail?}'' However, this information is crucial for researchers to specifically improve model robustness. Recent research suggests that models may be particularly sensitive to certain semantics in visual inputs (such as ``wet,'' ``foggy''), making them prone to errors. Inspired by this, in this paper we conducted the first exploration on large vision-language models (LVLMs) and found that LVLMs indeed are susceptible to hallucinations and various errors when facing specific semantic concepts in images. To efficiently search for these sensitive concepts, we integrated large language models (LLMs) and text-to-image (T2I) models to propose a novel semantic evolution framework. Randomly initialized semantic concepts undergo LLM-based crossover and mutation operations to form image descriptions, which are then converted by T2I models into visual inputs for LVLMs. The task-specific performance of LVLMs on each input is quantified as fitness scores for the involved semantics and serves as reward signals to further guide LLMs in exploring concepts that induce LVLMs. Extensive experiments on seven mainstream LVLMs and two multimodal tasks demonstrate the effectiveness of our method. Additionally, we provide interesting findings about the sensitive semantics of LVLMs, aiming to inspire further in-depth research.

cs.CV↗

arXiv2025

An Empirical Study on Embodied Artificial Intelligence Robot (EAIR) Software Bugs

Embodied Artificial Intelligence Robots (EAIR) is an emerging and rapidly evolving technological domain. Ensuring their program correctness is fundamental to their successful deployment. However, a general and in-depth understanding of EAIR system bugs remains lacking, which hinders the development of practices and techniques to tackle EAIR system bugs. To bridge this gap, we conducted the first systematic study of 885 EAIR system bugs collected from 80 EAIR system projects to investigate their symptoms, underlying causes, and module distribution. Our analysis takes considerable effort, which classifies these bugs into 18 underlying causes, 15 distinct symptoms, and identifies 13 affected modules. It reveals several new interesting findings and implications which help shed light on future research on tackling or repairing EAIR system bugs. First, among the 15 identified symptoms, our findings highlight 8 symptoms specific to EAIR systems, which is characterized by severe functional failures and potential physical hazards. Second, within the 18 underlying causes, we define 8 EAIR-specific causes, the majority of which stem from the intricate issues of AI- agent reasoning and decision making. Finally, to facilitate precise and efficient bug prediction, detection, and repair, we constructed a mapping between underlying causes and the modules in which they most frequently occur, which enables researchers to focus diagnostic efforts on the modules most susceptible to specific bug types.

cs.SE↗

arXiv2025

Accelerating Automatic Program Repair with Dual Retrieval-Augmented Fine-Tuning and Patch Generation on Large Language Models

Automated Program Repair (APR) is essential for ensuring software reliability and quality while enhancing efficiency and reducing developers' workload. Although rule-based and learning-based APR methods have demonstrated their effectiveness, their performance was constrained by the defect type of repair, the quality of training data, and the size of model parameters. Recently, Large Language Models (LLMs) combined with Retrieval-Augmented-Generation (RAG) have been increasingly adopted in APR tasks. However, current code LLMs and RAG designs neither fully address code repair tasks nor consider code-specific features. To overcome these limitations, we propose SelRepair, a novel APR approach with integration of a fine-tuned LLM with a newly-designed dual RAG module. This approach uses a bug-fix pair dataset for fine-tuning and incorporates semantic and syntactic/structural similarity information through an RAG selection gate. This design ensures relevant information is retrieved efficiently, thereby reducing token length and inference time. Evaluations on Java datasets show SelRepair outperforms other APR methods, achieving 26.29% and 17.64% in terms of exact match (EM) on different datasets while reducing inference time by at least 6.42% with controlled input lengths.

cs.SE↗

arXiv2025

An Empirical Study of Interaction Bugs in ROS-based Software

Modern robotic systems integrate multiple independent software and hardware components, each responsible for distinct functionalities such as perception, decision-making, and execution. These components interact extensively to accomplish complex end-to-end tasks. As a result, the overall system reliability depends not only on the correctness of individual components, but also on the correctness of their interactions. Failures often manifest at the boundaries between components, yet interaction-related reliability issues in robotics--referred to here as interaction bugs (iBugs)--remain underexplored. This work presents an empirical study of iBugs within robotic systems built using the Robot Operating System (ROS), a widely adopted open-source robotics framework. A total of 121 iBugs were analyzed across ten actively maintained and representative ROS projects. The identified iBugs are categorized into three major types: intra-system iBugs, hardware iBugs, and environmental iBugs, covering a broad range of interaction scenarios in robotics. The analysis includes an examination of root causes, fixing strategies, and the impact of these bugs. Several findingsa are derived that shed light on the nature of iBugs and suggest directions for improving their prevention and detection. These insights aim to inform the design of more robust and safer robotic systems.

cs.SE↗

arXiv2025

DaiFu: In-Situ Crash Recovery for Deep Learning Systems

Deep learning (DL) systems have been widely adopted in many areas, and are becoming even more popular with the emergence of large language models. However, due to the complex software stacks involved in their development and execution, crashes are unavoidable and common. Crashes severely waste computing resources and hinder development productivity, so efficient crash recovery is crucial. Existing solutions, such as checkpoint-retry, are too heavyweight for fast recovery from crashes caused by minor programming errors or transient runtime errors. Therefore, we present DaiFu, an in-situ recovery framework for DL systems. Through a lightweight code transformation to a given DL system, DaiFu augments it to intercept crashes in situ and enables dynamic and instant updates to its program running context (e.g., code, configurations, and other data) for agile crash recovery. Our evaluation shows that DaiFu helps reduce the restore time for crash recovery, achieving a 1372x speedup compared with state-of-the-art solutions. Meanwhile, the overhead of DaiFu is negligible (under 0.40%). We also construct a benchmark spanning 7 distinct crash scenarios in DL systems, and show the effectiveness of DaiFu in diverse situations.

cs.SE↗

arXiv2025

A Hierarchical and Evolvable Benchmark for Fine-Grained Code Instruction Following with Multi-Turn Feedback

Large language models (LLMs) have advanced significantly in code generation, yet their ability to follow complex programming instructions with layered and diverse constraints remains underexplored. Existing benchmarks often prioritize functional correctness, overlooking the nuanced requirements found in real-world development. We introduce MultiCodeIF, a comprehensive benchmark designed to evaluate instruction-following in code generation across multiple dimensions: constraint type, hierarchical levels, and iterative refinement. Built upon a structured taxonomy of 9 categories and 27 constraint types, MultiCodeIF enables granular assessment of both functional and non-functional instruction adherence. Using an automated pipeline, ConstraGen, we synthesize and evolve 2,021 code tasks sourced from 14 programming languages, supporting multi-turn evaluation through feedback-driven task variants. Empirical evaluation of six state-of-the-art LLMs uncovers substantial performance disparities. The top-performing model, Claude-3-7-Sonnet, achieves 63.0% average constraint satisfaction, while smaller models like Qwen3-1.7B fall to 44.8%. Models perform well on explicit constraints, but struggle with implicit or abstract constraints. Tasks with multiple hierarchical constraints significantly reduce model success rates, from 54.5% in single-level to just 18.8% in multi-level scenarios. However, structured feedback enables progressive improvement: average constraint satisfaction rises from 63.0% to 83.4% over four iterative refinement rounds. MultiCodeIF provides a scalable, constraint-aware, and feedback-sensitive framework to benchmark LLMs under realistic code generation scenarios, bridging the gap between synthetic evaluations and real-world instruction complexity. The full benchmark dataset, evaluation pipeline, and source code are available at https://github.com/SYSUSELab/MultiCodeIF.

cs.SE↗

arXiv2025

What Matters in LLM-generated Data: Diversity and Its Effect on Model Fine-Tuning

With the remarkable generative capabilities of large language models (LLMs), using LLM-generated data to train downstream models has emerged as a promising approach to mitigate data scarcity in specific domains and reduce time-consuming annotations. However, recent studies have highlighted a critical issue: iterative training on self-generated data results in model collapse, where model performance degrades over time. Despite extensive research on the implications of LLM-generated data, these works often neglect the importance of data diversity, a key factor in data quality. In this work, we aim to understand the implications of the diversity of LLM-generated data on downstream model performance. Specifically, we explore how varying levels of diversity in LLM-generated data affect downstream model performance. Additionally, we investigate the performance of models trained on data that mixes different proportions of LLM-generated data, which we refer to as synthetic data. Our experimental results show that, with minimal distribution shift, moderately diverse LLM-generated data can enhance model performance in scenarios with insufficient labeled data, whereas highly diverse generated data has a negative impact. We hope our empirical findings will offer valuable guidance for future studies on LLMs as data generators.

cs.CL↗

arXiv2025

SaGIF: Improving Individual Fairness in Graph Neural Networks via Similarity Encoding

Individual fairness (IF) in graph neural networks (GNNs), which emphasizes the need for similar individuals should receive similar outcomes from GNNs, has been a critical issue. Despite its importance, research in this area has been largely unexplored in terms of (1) a clear understanding of what induces individual unfairness in GNNs and (2) a comprehensive consideration of identifying similar individuals. To bridge these gaps, we conduct a preliminary analysis to explore the underlying reason for individual unfairness and observe correlations between IF and similarity consistency, a concept introduced to evaluate the discrepancy in identifying similar individuals based on graph structure versus node features. Inspired by our observations, we introduce two metrics to assess individual similarity from two distinct perspectives: topology fusion and feature fusion. Building upon these metrics, we propose Similarity-aware GNNs for Individual Fairness, named SaGIF. The key insight behind SaGIF is the integration of individual similarities by independently learning similarity representations, leading to an improvement of IF in GNNs. Our experiments on several real-world datasets validate the effectiveness of our proposed metrics and SaGIF. Specifically, SaGIF consistently outperforms state-of-the-art IF methods while maintaining utility performance. Code is available at: https://github.com/ZzoomD/SaGIF.

cs.LG↗

arXiv2025

FORGE: An LLM-driven Framework for Large-Scale Smart Contract Vulnerability Dataset Construction

High-quality smart contract vulnerability datasets are critical for evaluating security tools and advancing smart contract security research. Two major limitations of current manual dataset construction are (1) labor-intensive and error-prone annotation processes limiting the scale, quality, and evolution of the dataset, and (2) absence of standardized classification rules results in inconsistent vulnerability categories and labeling results across different datasets. To address these limitations, we present FORGE, the first automated approach for constructing smart contract vulnerability datasets. FORGE leverages an LLM-driven pipeline to extract high-quality vulnerabilities from real-world audit reports and classify them according to the CWE, the most widely recognized classification in software security. FORGE employs a divide-and-conquer strategy to extract structured and self-contained vulnerability information from these reports. Additionally, it uses a tree-of-thoughts technique to classify the vulnerability information into the hierarchical CWE classification. To evaluate FORGE's effectiveness, we run FORGE on 6,454 real-world audit reports and generate a dataset comprising 81,390 solidity files and 27,497 vulnerability findings across 296 CWE categories. Manual assessment of the dataset demonstrates high extraction precision and classification consistency with human experts (precision of 95.6% and inter-rater agreement k-$α$ of 0.87). We further validate the practicality of our dataset by benchmarking 13 existing security tools on our dataset. The results reveal the significant limitations in current detection capabilities. Furthermore, by analyzing the severity-frequency distribution patterns through a unified CWE perspective in our dataset, we highlight inconsistency between current smart contract research focus and priorities identified from real-world vulnerabilities...

cs.CR↗

arXiv2025

Beyond Functional Correctness: Investigating Coding Style Inconsistencies in Large Language Models

Large language models (LLMs) have brought a paradigm shift to the field of code generation, offering the potential to enhance the software development process. However, previous research mainly focuses on the accuracy of code generation, while coding style differences between LLMs and human developers remain under-explored. In this paper, we empirically analyze the differences in coding style between the code generated by mainstream Code LLMs and the code written by human developers, and summarize coding style inconsistency taxonomy. Specifically, we first summarize the types of coding style inconsistencies by manually analyzing a large number of generation results. We then compare the code generated by Code LLMs with the code written by human programmers in terms of readability, conciseness, and robustness. The results reveal that LLMs and developers have different coding styles. Additionally, we study the possible causes of these inconsistencies and provide some solutions to alleviate the problem.

cs.SE↗

arXiv2025

One Fits All: Learning Fair Graph Neural Networks for Various Sensitive Attributes

Recent studies have highlighted fairness issues in Graph Neural Networks (GNNs), where they produce discriminatory predictions against specific protected groups categorized by sensitive attributes such as race and age. While various efforts to enhance GNN fairness have made significant progress, these approaches are often tailored to specific sensitive attributes. Consequently, they necessitate retraining the model from scratch to accommodate changes in the sensitive attribute requirement, resulting in high computational costs. To gain deeper insights into this issue, we approach the graph fairness problem from a causal modeling perspective, where we identify the confounding effect induced by the sensitive attribute as the underlying reason. Motivated by this observation, we formulate the fairness problem in graphs from an invariant learning perspective, which aims to learn invariant representations across environments. Accordingly, we propose a graph fairness framework based on invariant learning, namely FairINV, which enables the training of fair GNNs to accommodate various sensitive attributes within a single training session. Specifically, FairINV incorporates sensitive attribute partition and trains fair GNNs by eliminating spurious correlations between the label and various sensitive attributes. Experimental results on several real-world datasets demonstrate that FairINV significantly outperforms state-of-the-art fairness approaches, underscoring its effectiveness. Our code is available via: https://github.com/ZzoomD/FairINV/.

cs.LG↗

arXiv2025

MTCMB: A Multi-Task Benchmark Framework for Evaluating LLMs on Knowledge, Reasoning, and Safety in Traditional Chinese Medicine

Traditional Chinese Medicine (TCM) is a holistic medical system with millennia of accumulated clinical experience, playing a vital role in global healthcare-particularly across East Asia. However, the implicit reasoning, diverse textual forms, and lack of standardization in TCM pose major challenges for computational modeling and evaluation. Large Language Models (LLMs) have demonstrated remarkable potential in processing natural language across diverse domains, including general medicine. Yet, their systematic evaluation in the TCM domain remains underdeveloped. Existing benchmarks either focus narrowly on factual question answering or lack domain-specific tasks and clinical realism. To fill this gap, we introduce MTCMB-a Multi-Task Benchmark for Evaluating LLMs on TCM Knowledge, Reasoning, and Safety. Developed in collaboration with certified TCM experts, MTCMB comprises 12 sub-datasets spanning five major categories: knowledge QA, language understanding, diagnostic reasoning, prescription generation, and safety evaluation. The benchmark integrates real-world case records, national licensing exams, and classical texts, providing an authentic and comprehensive testbed for TCM-capable models. Preliminary results indicate that current LLMs perform well on foundational knowledge but fall short in clinical reasoning, prescription planning, and safety compliance. These findings highlight the urgent need for domain-aligned benchmarks like MTCMB to guide the development of more competent and trustworthy medical AI systems. All datasets, code, and evaluation tools are publicly available at: https://github.com/Wayyuanyuan/MTCMB.

cs.CL↗

arXiv2025

SCA: Improve Semantic Consistent in Unrestricted Adversarial Attacks via DDPM Inversion

Systems based on deep neural networks are vulnerable to adversarial attacks. Unrestricted adversarial attacks typically manipulate the semantic content of an image (e.g., color or texture) to create adversarial examples that are both effective and photorealistic. Recent works have utilized the diffusion inversion process to map images into a latent space, where high-level semantics are manipulated by introducing perturbations. However, they often result in substantial semantic distortions in the denoised output and suffer from low efficiency. In this study, we propose a novel framework called Semantic-Consistent Unrestricted Adversarial Attacks (SCA), which employs an inversion method to extract edit-friendly noise maps and utilizes a Multimodal Large Language Model (MLLM) to provide semantic guidance throughout the process. Under the condition of rich semantic information provided by MLLM, we perform the DDPM denoising process of each step using a series of edit-friendly noise maps and leverage DPM Solver++ to accelerate this process, enabling efficient sampling with semantic consistency. Compared to existing methods, our framework enables the efficient generation of adversarial examples that exhibit minimal discernible semantic changes. Consequently, we for the first time introduce Semantic-Consistent Adversarial Examples (SCAE). Extensive experiments and visualizations have demonstrated the high efficiency of SCA, particularly in being on average 12 times faster than the state-of-the-art attacks. Our code can be found at https://github.com/Pan-Zihao/SCA.

cs.CV↗

arXiv2025

Fate: Fast Edge Inference of Mixture-of-Experts Models via Cross-Layer Gate

Large Language Models (LLMs) have demonstrated impressive performance across various tasks, and their application in edge scenarios has attracted significant attention. However, sparse-activated Mixture-of-Experts (MoE) models, which are well suited for edge scenarios, have received relatively little attention due to their high memory demands. Offload-based methods have been proposed to address this challenge, but they face difficulties with expert prediction. Inaccurate expert predictions can result in prolonged inference delays. To promote the application of MoE models in edge scenarios, we propose Fate, an offloading system designed for MoE models to enable efficient inference in resource-constrained environments. The key insight behind Fate is that gate inputs from adjacent layers can be effectively used for expert prefetching, achieving high prediction accuracy without additional GPU overhead. Furthermore, Fate employs a shallow-favoring expert caching strategy that increases the expert hit rate to 99\%. Additionally, Fate integrates tailored quantization strategies for cache optimization and IO efficiency. Experimental results show that, compared to Load on Demand and Expert Activation Path-based method, Fate achieves up to 4.5x and 1.9x speedups in prefill speed and up to 4.1x and 2.2x speedups in decoding speed, respectively, while maintaining inference quality. Moreover, Fate's performance improvements are scalable across different memory budgets.

cs.AI↗

arXiv2025

SolPhishHunter: Towards Detecting and Understanding Phishing on Solana

Solana is a rapidly evolving blockchain platform that has attracted an increasing number of users. However, this growth has also drawn the attention of malicious actors, with some phishers extending their reach into the Solana ecosystem. Unlike platforms such as Ethereum, Solana has distinct designs of accounts and transactions, leading to the emergence of new types of phishing transactions that we term SolPhish. We define three types of SolPhish and develop a detection tool called SolPhishHunter. Utilizing SolPhishHunter, we detect a total of 8,058 instances of SolPhish and conduct an empirical analysis of these detected cases. Our analysis explores the distribution and impact of SolPhish, the characteristics of the phishers, and the relationships among phishing gangs. Particularly, the detected SolPhish transactions have resulted in nearly \$1.1 million in losses for victims. We report our detection results to the community and construct SolPhishDataset, the \emph{first} Solana phishing-related dataset in academia.

cs.CR↗

arXiv2025

OmniGIRL: A Multilingual and Multimodal Benchmark for GitHub Issue Resolution

The GitHub issue resolution task aims to resolve issues reported in repositories automatically. With advances in large language models (LLMs), this task has gained increasing attention, and several benchmarks are proposed to evaluate the issue resolution ability of LLMs. However, existing benchmarks have three main limitations. First, current benchmarks focus on a single programming language, limiting the evaluation of issues from repositories across different languages. Second, they usually cover a narrow range of domains, which may fail to represent the diversity of real-world issues. Third, existing benchmarks rely solely on textual information in issue descriptions, overlooking multimodal information such as images in issues. In this paper, we propose OmniGIRL, a GitHub Issue ResoLution benchmark that is multilingual, multimodal, and multi-domain. OmniGIRL includes 959 task instances, which are collected from repositories across four programming languages (i.e., Python, JavaScript, TypeScript, and Java) and eight different domains. Our evaluation shows that current LLMs show limited performances on OmniGIRL. Notably, the best-performing model, GPT-4o, resolves only 8.6% of the issues. Besides, we find that current LLMs struggle to resolve issues requiring understanding images. The best performance is achieved by Claude-3.5-Sonnet, which resolves only 10.5% of the issues with image information. Finally, we analyze the reasons behind current LLMs' failure on OmniGIRL, providing insights for future improvements.

cs.SE↗

arXiv2025

Clients Collaborate: Flexible Differentially Private Federated Learning with Guaranteed Improvement of Utility-Privacy Trade-off

To defend against privacy leakage of user data, differential privacy is widely used in federated learning, but it is not free. The addition of noise randomly disrupts the semantic integrity of the model and this disturbance accumulates with increased communication rounds. In this paper, we introduce a novel federated learning framework with rigorous privacy guarantees, named FedCEO, designed to strike a trade-off between model utility and user privacy by letting clients ''Collaborate with Each Other''. Specifically, we perform efficient tensor low-rank proximal optimization on stacked local model parameters at the server, demonstrating its capability to flexibly truncate high-frequency components in spectral space. This capability implies that our FedCEO can effectively recover the disrupted semantic information by smoothing the global semantic space for different privacy settings and continuous training processes. Moreover, we improve the SOTA utility-privacy trade-off bound by order of $\sqrt{d}$, where $d$ is the input dimension. We illustrate our theoretical results with experiments on representative datasets and observe significant performance improvements and strict privacy guarantees under different privacy settings. The code is available at https://github.com/6lyc/FedCEO_Collaborate-with-Each-Other.

cs.LG↗