arXiv subjects

Yuming Zhou

Publications and source records attributed to Yuming Zhou.

At least 19 recordsLinked to original sources

arXiv2026

Characterizing Large Language Model Agentic Workflows: A Study on N8n Ecosystem

Large Language Models (LLMs) are rapidly being adopted in low-code and no-code automation platforms, where non-expert users design workflows that combine natural language understanding with external services and APIs. LLM agents are LLM systems that use LLMs as a core "brain" to reason, plan, and autonomously execute complex, multi-step tasks. In this paper, we present the first large-scale empirical study of LLM agentic workflows in low-code automation platforms. We analyze more than 6,000 publicly available n8n workflows and examine four aspects of their design: task distribution, structural and tool use patterns, reliability mechanisms, and autonomy levels. Our analysis shows that LLM workflows are not merely prompt response pipelines. Instead, LLMs are commonly embedded within broader automation structures involving control logic, external tools, communication services, storage systems, and human review points. We further find that while many workflows include lightweight post-processing or routing logic after LLM execution, explicit reliability mechanisms such as structured fallback paths, repair loops, failure-specific alerts, and human approval gates remain relatively uncommon. These results reveal a gap between the increasing deployment of LLM agents in practical automation ecosystems and the limited engineering support for reliability, safety, and governance. Overall, our study provides ten empirical findings and five research takeaways for researchers, platform developers, and practitioners seeking to understand and improve real-world LLM agentic workflows.

cs.AI↗

arXiv2026

Harnessing Routing Foresight for Micro-step-level MoE load balancing in RL Post-training

Mixture-of-Experts (MoE) and reinforcement learning (RL) post-training now dominate large language model (LLM) development, yet expert load imbalance remains a critical challenge. Existing load-balancing systems target pre-training by relying on historical step-level statistics. However, these methods fail under the unique workload dynamics of RL post-training: the step-level load is stable, but the tiny batch sizes processed during micro-steps cause severe, high-frequency load fluctuations. We introduce ForeMoE, a micro-step-level load balancing system for MoE RL post-training. Instead of relying on historical statistics, ForeMoE exploits the multi-stage RL pipeline (rollout, recompute, policy update) by using foreseeable routing information from the rollout stage to proactively guide load balancing in the remaining stages. To support frequent per-micro-step reconfiguration, ForeMoE employs a hierarchical planner that decomposes the NP-hard load balancing problem into tractable sub-components, alongside a transfer engine that leverages complementary hardware paths (CPU-assisted and GPU-direct) for overlapped expert transfer. Evaluations on 64 GPUs demonstrate that ForeMoE achieves up to a 1.45$\times$ speedup over state-of-the-art RL post-training systems.

cs.DC↗

arXiv2026

ClozeMaster: Fuzzing Rust Compiler by Harnessing LLMs for Infilling Masked Real Programs

Ensuring the reliability of the Rust compiler is of paramount importance, given increasing adoption of Rust for critical systems development, due to its emphasis on memory and thread safety. However, generating valid test programs for the Rust compiler poses significant challenges, given Rust's complex syntax and strict requirements. With the growing popularity of large language models (LLMs), much research in software testing has explored using LLMs to generate test cases. Still, directly using LLMs to generate Rust programs often results in a large number of invalid test cases. Existing studies have indicated that test cases triggering historical compiler bugs can assist in software testing. Our investigation into Rust compiler bug issues supports this observation. Inspired by existing work and our empirical research, we introduce a bracket-based masking and filling strategy called clozeMask. The clozeMask strategy involves extracting test code from historical issue reports, identifying and masking code snippets with specific structures, and using an LLM to fill in the masked portions for synthesizing new test programs. This approach harnesses the generative capabilities of LLMs while retaining the ability to trigger Rust compiler bugs. It enables comprehensive testing of the compiler's behavior, particularly exploring edge cases. We implemented our approach as a prototype CLOZEMASTER. CLOZEMASTER has identified 27 confirmed bugs for rustc and mrustc, of which 10 have been fixed by developers. Furthermore, our experimental results indicate that CLOZEMASTER outperforms existing fuzzers in terms of code coverage and effectiveness.

cs.SE↗

arXiv2026

DebugHarness: Emulating Human Dynamic Debugging for Autonomous Program Repair

Patching severe security flaws in complex software remains a major challenge. While automated tools like fuzzers efficiently discover bugs, fixing deep-rooted low-level faults (e.g., use-after-free and memory corruption) still requires labor-intensive manual analysis by experts. Emerging Large Language Model (LLM) agents attempt to automate this pipeline, but they typically treat bug fixing as a purely static code-generation task. Relying solely on static artifacts, these methods miss the dynamic execution context strictly necessary for diagnosing intricate memory safety violations. To overcome these limitations, we introduce DebugHarness, an autonomous LLM-powered debugging agent harness that resolves complex vulnerabilities by emulating the interactive debugging practices of human systems engineers. Instead of merely examining static code, DebugHarness actively queries the live runtime environment. Driven by a reproducible crash, it utilizes a pattern-guided investigation strategy to formulate hypotheses, interactively probes program memory states and execution paths, and synthesizes patches via a closed-loop validation cycle. We evaluate DebugHarness on SEC-bench, a rigorous dataset of real-world C/C++ security vulnerabilities. DebugHarness successfully patches approximately 90% of the evaluated bugs. This yields a relative improvement of over 30% compared to state-of-the-art baselines, demonstrating that dynamic debugging significantly enhances LLM diagnostic capabilities. Overall, DebugHarness establishes a novel paradigm for automated program repair, bridging the gap between static LLM reasoning and the dynamic intricacies of low-level systems programming.

cs.SE↗

arXiv2026

Can Adjusting Hyperparameters Lead to Green Deep Learning: An Empirical Study on Correlations between Hyperparameters and Energy Consumption of Deep Learning Models

Context: Along with developing Deep learning (DL) models, larger datasets and more complex model structures are applied, leading to rising computing resources and energy consumption, which is an alert that green DL models should receive more attention. Objective: This paper focuses on a novel view to analyze DL energy consumption: the effect of hyperparameters on the energy cost of DL models. Method: Our approach involves using mutation operators to simulate how practitioners adjust hyperparameters, such as epochs and learning rates. We train the original and mutated models separately and gather energy information and run-time performance metrics. Moreover, we focus on the parallel scenario where multiple DL models are trained in parallel. Results: To examine the effect of hyperparameters on energy consumption, we conducted extensive experiments on five real-world DL models. The results show that (1) many hyperparameters studied have a (positive or negative) correlation with energy consumption, (2) adjusting hyperparameters can make DL models greener, i.e., lead to less energy consumption without performance damage, and (3) in a parallel environment, energy consumption becomes more susceptible to change. Conclusions: We suggest that hyperparameters need more attention in developing DL models, as appropriately adjusting hyperparameters would cause green DL models.

cs.SE↗

arXiv2026

StaleFlow: Staleness-Aware Data Management for Mitigating Data Skewness in Fully Disaggregated RL Post-Training

Reinforcement learning (RL) post-training has become pivotal for enhancing the capabilities of modern large models. A recent trend is to develop RL systems with a fully disaggregated architecture, which decouples the three RL phases (rollout, reward, and training) onto separate resources and executes them asynchronously. However, two critical data-level concerns arise: (1) asynchronous execution leads to data staleness in trajectories (the data generated by rollout) as the model parameters used in rollout may not be up to date, which impairs RL convergence; and (2) the length variation of trajectories introduces severe data skewness, leading to workload imbalance and degraded system performance. Existing systems fail to address these two concerns in a unified manner. Techniques that tightly control data staleness often constrain effective data skewness mitigation, while aggressive data skewness mitigation tends to exacerbate data staleness. As a result, systems are forced to trade off convergence for performance, or vice versa. To address this, we propose StaleFlow, an RL post-training system that jointly tackles data staleness and skewness. First, to control staleness, StaleFlow introduces a global consistency protocol that tracks the full lifecycle of each trajectory and constrains staleness. Second, to mitigate skewness, StaleFlow re-designs the RL system architecture by constructing data servers for trajectories and parameters to achieve flexible rollout coordination. Subsequently, we develop a suite of staleness-aware, throughput-oriented strategies to enhance system performance. Evaluations show that StaleFlow achieves up to 1.42-2.68$\times$ (1.18-1.91$\times$ on average) higher throughput than state-of-the-art systems, without compromising convergence. Our source code is available: https://github.com/psrl-project/psrl.

cs.DC↗

arXiv2025

Isolating Compiler Faults via Multiple Pairs of Adversarial Compilation Configurations

Compilers are fundamental to modern software development, making the effective identification and resolution of compiler faults essential. However, localizing these faults to specific source files remains highly challenging due to the complexity and scale of modern compiler infrastructures. In this study, we propose MultiConf, a novel approach that automatically isolates compiler faults by constructing multiple pairs of adversarial compilation configurations. Each adversarial compilation configuration pair consists of a failing configuration and its corresponding passing configuration, which differ in only a small number of fine-grained options. MultiConf generates failing configurations through a lightweight construction process and derives the corresponding passing configurations by selectively disabling bug-related fine-grained options. We then employ a Spectrum-Based Fault Localization (SBFL) formula to rank the suspiciousness of compiler source files. Each adversarial configuration pair independently produces a ranking, which is subsequently aggregated using a weighted voting scheme to derive a final suspiciousness ranking, enabling more accurate and robust fault localization. We evaluate MultiConf on a benchmark of 60 real-world GCC compiler bugs. The results demonstrate that MultiConf significantly outperforms existing compiler fault localization techniques in both effectiveness and efficiency. In particular, MultiConf successfully localizes 27 out of 60 bugs at the Top-1 file level, representing improvements of 35.0% and 28.6% over the two state-of-the-art approaches, Odfl(20) and Basic(21), respectively.

cs.SE↗

arXiv2025

Using a Sledgehammer to Crack a Nut? Revisiting Automated Compiler Fault Isolation

Background: Compilers are fundamental to software development, translating high-level source code into executable software systems. Faults in compilers can have severe consequences and thus effective localization and resolution of compiler bugs are crucial. Problem: In practice, developers often examine version history to identify and investigate bug-inducing commit (BIC) for fixing bugs. However, while numerous sophisticated Spectrum-Based Fault Localization (SBFL) techniques have been proposed for compiler fault isolation, their effectiveness has not been evaluated against the BIC-based strategies widely adopted in practice. Objective: This study aims to bridge this gap by directly comparing a BIC-based strategy, Basic, with representative SBFL techniques in the context of compiler fault localization. The BIC-based strategy closely aligns with common developer practices, as it directly identifies the BIC and treats the files modified in that commit as faulty candidates. Method: The Basic identifies the most recent good release and earliest bad release, and then employs a binary search to pinpoint the bug-inducing commit. All files modified in the identified commit are flagged as potentially faulty. We rigorously compare Basic against SBFL-based techniques using a benchmark consisting of 60 GCC bugs and 60 LLVM bugs. Result: Our analysis reveals that Basic performs comparably to, and in many cases outperforms, state-of-the-art SBFL-based techniques, particularly on the critical Top-1 and Top-5 ranking metrics. Conclusion: This study provides new insights into the practical effectiveness of SBFL-based techniques in real-world compiler debugging scenarios. We recommend that future research adopt Basic as a baseline when developing and evaluating new compiler fault isolation methods.

cs.SE↗

arXiv2025

WITNESS: A lightweight and practical approach to fine-grained predictive mutation testing

Existing fine-grained predictive mutation testing studies predominantly rely on deep learning, which faces two critical limitations in practice: (1) Exorbitant computational costs. The deep learning models adopted in these studies demand significant computational resources for training and inference acceleration. This introduces high costs and undermines the cost-reduction goal of predictive mutation testing. (2) Constrained applicability. Although modern mutation testing tools generate mutants both inside and outside methods, current fine-grained predictive mutation testing approaches handle only inside-method mutants. As a result, they cannot predict outside-method mutants, limiting their applicability in real-world scenarios. We propose WITNESS, a new fine-grained predictive mutation testing approach. WITNESS adopts a twofold design: (1) With collected features from both inside-method and outside-method mutants, WITNESS is suitable for all generated mutants. (2) Instead of using computationally expensive deep learning, WITNESS employs lightweight classical machine learning models for training and prediction. This makes it more cost-effective and enabling straightforward explanations of the decision-making processes behind the adopted models. Evaluations on Defects4J projects show that WITNESS consistently achieves state-of-the-art predictive performance across different scenarios. Additionally, WITNESS significantly enhances the efficiency of kill matrix prediction. Post-hoc analysis reveals that features incorporating information from before and after the mutation are the most important among those used in WITNESS. Test case prioritization based on the predicted kill matrix shows that WITNESS delivers results much closer to those obtained by using the actual kill matrix, outperforming baseline approaches.

cs.SE↗

arXiv2025

Once4All: Skeleton-Guided SMT Solver Fuzzing with LLM-Synthesized Generators

Satisfiability Modulo Theory (SMT) solvers are foundational to modern systems and programming languages research, providing the foundation for tasks like symbolic execution and automated verification. Because these solvers sit on the critical path, their correctness is essential, and high-quality test formulas are key to uncovering bugs. However, while prior testing techniques performed well on earlier solver versions, they struggle to keep pace with rapidly evolving features. Recent approaches based on Large Language Models (LLMs) show promise in exploring advanced solver capabilities, but two obstacles remain: nearly half of the generated formulas are syntactically invalid, and iterative interactions with LLMs introduce substantial computational overhead. In this study, we present Once4All, a novel LLM-assisted fuzzing framework that addresses both issues by shifting from direct formula generation to the synthesis of generators for reusable terms (i.e., logical expressions). Specifically, Once4All uses LLMs to (1) automatically extract context-free grammars (CFGs) for SMT theories, including solver-specific extensions, from documentation, and (2) synthesize composable Boolean term generators that adhere to these grammars. During fuzzing, Once4All populates structural skeletons derived from existing formulas with the terms iteratively produced by the LLM-synthesized generators. This design ensures syntactic validity while promoting semantic diversity. Notably, Once4All requires only one-time LLM interaction investment, dramatically reducing runtime cost. We evaluated Once4All on two leading SMT solvers: Z3 and cvc5. Our experiments show that Once4All has identified 43 confirmed bugs, 40 of which have already been fixed by developers.

cs.SE↗

arXiv2025

Hetu v2: A General and Scalable Deep Learning System with Hierarchical and Heterogeneous Single Program Multiple Data Annotations

The Single-Program Multiple-Data (SPMD) paradigm provides a unified abstraction to annotate various parallel dimensions in distributed deep learning (DL) training. With SPMD, users can write training programs from the viewpoint of a single device, and the system will automatically deduce the tensor sharding and communication patterns. However, with the recent development in large-scale DL models, distributed training exhibits spatial and temporal workload heterogeneity, arising from both device disparities (e.g., mixed hardware, failures) and data variations (e.g., uneven sequence lengths). Such heterogeneity violates SPMD's assumption of symmetric workload partitioning, which restricts its ability to express and optimize heterogeneous parallel strategies effectively. To address this, we propose HSPMD within the Hetu v2 system to achieve general and scalable DL training. HSPMD extends SPMD's declarative annotations to support asymmetric sharding and composes standard communication primitives for hierarchical communication, all while retaining the simplicity of a single-device programming model. HSPMD handles spatial heterogeneity through progressive graph specialization, enabling device-specific execution logic, and addresses temporal heterogeneity via dynamic graph switching. Evaluations on (a) heterogeneous devices, (b) unstable devices, and (c) mixed-length data scenarios show that HSPMD matches or outperforms specialized systems, providing a flexible and efficient solution for modern distributed DL training.

cs.DC↗

arXiv2025

SCOPE-DTI: Semi-Inductive Dataset Construction and Framework Optimization for Practical Usability Enhancement in Deep Learning-Based Drug Target Interaction Prediction

Deep learning-based drug-target interaction (DTI) prediction methods have demonstrated strong performance; however, real-world applicability remains constrained by limited data diversity and modeling complexity. To address these challenges, we propose SCOPE-DTI, a unified framework combining a large-scale, balanced semi-inductive human DTI dataset with advanced deep learning modeling. Constructed from 13 public repositories, the SCOPE dataset expands data volume by up to 100-fold compared to common benchmarks such as the Human dataset. The SCOPE model integrates three-dimensional protein and compound representations, graph neural networks, and bilinear attention mechanisms to effectively capture cross domain interaction patterns, significantly outperforming state-of-the-art methods across various DTI prediction tasks. Additionally, SCOPE-DTI provides a user-friendly interface and database. We further validate its effectiveness by experimentally identifying anticancer targets of Ginsenoside Rh1. By offering comprehensive data, advanced modeling, and accessible tools, SCOPE-DTI accelerates drug discovery research.

cs.LG↗

arXiv2024

Deep Learning-based Software Engineering: Progress, Challenges, and Opportunities

Researchers have recently achieved significant advances in deep learning techniques, which in turn has substantially advanced other research disciplines, such as natural language processing, image processing, speech recognition, and software engineering. Various deep learning techniques have been successfully employed to facilitate software engineering tasks, including code generation, software refactoring, and fault localization. Many papers have also been presented in top conferences and journals, demonstrating the applications of deep learning techniques in resolving various software engineering tasks. However, although several surveys have provided overall pictures of the application of deep learning techniques in software engineering, they focus more on learning techniques, that is, what kind of deep learning techniques are employed and how deep models are trained or fine-tuned for software engineering tasks. We still lack surveys explaining the advances of subareas in software engineering driven by deep learning techniques, as well as challenges and opportunities in each subarea. To this end, in this paper, we present the first task-oriented survey on deep learning-based software engineering. It covers twelve major software engineering subareas significantly impacted by deep learning techniques. Such subareas spread out the through the whole lifecycle of software development and maintenance, including requirements engineering, software development, testing, maintenance, and developer collaboration. As we believe that deep learning may provide an opportunity to revolutionize the whole discipline of software engineering, providing one survey covering as many subareas as possible in software engineering can help future research push forward the frontier of deep learning-based software engineering more systematically.

cs.SE↗

arXiv2024

A Roadmap on Modern Code Review: Challenges and Opportunities

Over the past decade, modern code review (MCR) has been established as a cornerstone of software quality assurance and a vital channel for knowledge transfer within development teams. However, the manual inspection of increasingly complex systems remains a cognitively demanding and resource-intensive activity, often leading to significant workflow bottlenecks. This paper presents a comprehensive roadmap for the evolution of MCR, consolidating over a decade of research (2013-2025) into a unified taxonomy comprising improvement techniques, which focus on the technical optimization and automation of downstream review tasks, and understanding studies, which investigate the underlying socio-technical mechanisms and empirical phenomena of the review process. By diagnosing the current landscape through a strategic SWOT analysis, we examine the transformative impact of generative AI and identify critical gaps between burgeoning AI capabilities and industrial realities. We envision a future where MCR evolves from a human-driven task into a symbiotic partnership between developers and intelligent systems. Our roadmap charts this course by proposing three pivotal paradigm shifts, Context-Aware Proactivity, Value-Driven Evaluation, and Human-Centric Symbiosis, aiming to guide researchers and practitioners in transforming MCR into an intelligent, inclusive, and strategic asset for the AI-driven future.

cs.SE↗

arXiv2024

No Need to Lift a Finger Anymore? Assessing the Quality of Code Generation by ChatGPT

Large language models (LLMs) have demonstrated impressive capabilities across various NLP tasks. Additionally, LLMs are also highly valuable in supporting software engineering tasks, particularly in the field of code generation. Automatic code generation is a process of automatically generating source code or executable code based on given specifications or requirements, improving developer productivity. In this study, we perform a systematic empirical assessment to the quality of code generation using ChatGPT. We leverage 728 algorithm problems in five languages (i.e., C, C++, Java, Python, and JavaScript) and 18 CWEs with 54 code scenarios for the code generation task. Our evaluation encompasses a comprehensive analysis of code snippets generated by ChatGPT, focusing on three critical aspects: correctness, complexity, and security. We also specifically investigate ChatGPT's ability to engage in multi-round fixing process (i.e., ChatGPT's dialog ability) of facilitating code generation. By delving into the generated code and examining the experimental results, this work provides valuable insights into the performance of ChatGPT in tackling code generation tasks over the three critical aspects. Overall, our findings uncover potential issues and limitations that arise in the ChatGPT-based code generation and lay the groundwork for improving AI and LLM-based code generation techniques.

cs.SE↗

arXiv2024

Coverage Goal Selector for Combining Multiple Criteria in Search-Based Unit Test Generation

Unit testing is critical to the software development process, ensuring the correctness of basic programming units in a program (e.g., a method). Search-based software testing (SBST) is an automated approach to generating test cases. SBST generates test cases with genetic algorithms by specifying the coverage criterion (e.g., branch coverage). However, a good test suite must have different properties, which cannot be captured using an individual coverage criterion. Therefore, the state-of-the-art approach combines multiple criteria to generate test cases. Since combining multiple coverage criteria brings multiple objectives for optimization, it hurts the test suites' coverage for certain criteria compared with using the single criterion. To cope with this problem, we propose a novel approach named \textbf{smart selection}. Based on the coverage correlations among criteria and the subsumption relationships among coverage goals, smart selection selects a subset of coverage goals to reduce the number of optimization objectives and avoid missing any properties of all criteria. We conduct experiments to evaluate smart selection on $400$ Java classes with three state-of-the-art genetic algorithms under the $2$-minute budget. On average, smart selection outperforms combining all goals on $65.1\%$ of the classes having significant differences between the two approaches. Secondly, we conduct experiments to verify our assumptions about coverage criteria relationships. Furthermore, we assess the coverage performance of smart selection under varying budgets of $5$, $8$, and $10$ minutes and explore its effect on bug detection, confirming the advantage of smart selection over combining all goals.

cs.SE↗

arXiv2023

Human-in-the-loop online just-in-time software defect prediction

Online Just-In-Time Software Defect Prediction (O-JIT-SDP) uses an online model to predict whether a new software change will introduce a bug or not. However, existing studies neglect the interaction of Software Quality Assurance (SQA) staff with the model, which may miss the opportunity to improve the prediction accuracy through the feedback from SQA staff. To tackle this problem, we propose Human-In-The-Loop (HITL) O-JIT-SDP that integrates feedback from SQA staff to enhance the prediction process. Furthermore, we introduce a performance evaluation framework that utilizes a k-fold distributed bootstrap method along with the Wilcoxon signed-rank test. This framework facilitates thorough pairwise comparisons of alternative classification algorithms using a prequential evaluation approach. Our proposal enables continuous statistical testing throughout the prequential process, empowering developers to make real-time decisions based on robust statistical evidence. Through experimentation across 10 GitHub projects, we demonstrate that our evaluation framework enhances the credibility of model evaluation, and the incorporation of HITL feedback elevates the prediction performance of online JIT-SDP models. These advancements hold the potential to significantly enhance the value of O-JIT-SDP for industrial applications.

cs.SE↗

arXiv2023

On Code Reuse from StackOverflow: An Exploratory Study on Jupyter Notebook

Jupyter Notebook is a popular tool among data analysts and scientists for working with data. It provides a way to combine code, documentation, and visualizations in a single, interactive environment, facilitating code reuse. While code reuse can improve programming efficiency, it can also decrease readability, security, and overall performance. We conduct a large-scale exploratory study of code reuse practices in the Jupyter Notebook development community on the Stack Overflow platform to understand the potential negative impacts of code reuse. Our findings identified 1,097,470 Jupyter Notebook clone pairs that reuse Stack Overflow code snippets, and the average code snippet has 7.91 code quality violations. Through our research, we gain insight into the reasons behind Jupyter Notebook developers' decision to reuse code and the potential drawbacks of this practice.

cs.SE↗