arXiv subjects

Xin Peng

Publications and source records attributed to Xin Peng.

At least 55 records · Page 3Linked to original sources

arXiv2026

Code Digital Twin: A Knowledge Infrastructure for AI-Assisted Complex Software Development

Recent advances in AI coding tools powered by large language models (LLMs) have shown strong capabilities in software engineering tasks, raising expectations of major productivity gains. Tools such as Cursor and Claude Code have popularized "vibe coding" (where developers steer development through high-level intent), commonly relying on context engineering and Retrieval-Augmented Generation (RAG) to ground generation in a codebase. However, these paradigms struggle in ultra-complex enterprise systems, where software evolves incrementally under pervasive design constraints and depends on tacit knowledge such as responsibilities, intent, and decision rationales distributed across code, configurations, discussions, and version history. In this environment, context engineering faces a fundamental barrier: the required context is scattered across artifacts and entangled across time, beyond the capacity of LLMs to reliably capture, prioritize, and fuse evidence into correct and trustworthy decisions, even as context windows grow. To bridge this gap, we propose the Code Digital Twin, a persistent and evolving knowledge infrastructure built on the codebase. It separates long-term knowledge engineering from task-time context engineering and serves as a backend "context engine" for AI coding assistants. The Code Digital Twin models both the physical and conceptual layers of software and co-evolves with the system. By integrating hybrid knowledge representations, multi-stage extraction pipelines, incremental updates, AI-empowered applications, and human-in-the-loop feedback, it transforms fragmented knowledge into explicit and actionable representations, providing a roadmap toward sustainable and resilient development and evolution of ultra-complex systems.

cs.SE↗

arXiv2026

AutoMerge: Search-Based Model Merging Framework for Effective Model Reuse

Software reuse has long been recognized as a critical and widely studied topic in software engineering, offering substantial benefits in reducing development costs, improving software quality, and enhancing operational efficiency. This paradigm extends into deep learning through model reuse. Recently, model merging has emerged in the domain of large language models (LLMs) as a training-free approach that takes multiple task-specific models with the same architecture as source models and merges them without retraining, enhancing model reuse within LLMs. However, no prior work has systematically investigated whether such an approach can be effectively applied to other deep learning models with different architectures across domains. To bridge this gap, we present the first systematic study that evaluates five model merging techniques on three distinct model architectures across three domains: LLMs, image classification, and autonomous driving. Our findings reveal that directly applying existing model merging techniques leads to highly inconsistent results and falls notably short of their success within LLMs. Moreover, a single model merging technique often fails to handle the heterogeneous structural properties within a model, limiting its applicability to different model architectures across domains. Furthermore, the effectiveness of model merging techniques is highly sensitive to hyperparameter configurations, thereby constraining their potential for broader adoption. Inspired by these insights, we propose AutoMerge, a novel search-based model merging framework that first segments complex models into multiple heterogeneous blocks and then systematically explores the merging space to identify the merging technique and its hyperparameter configuration.

cs.SE↗

arXiv2026

Flow Perturbation++: Multi-Step Unbiased Jacobian Estimation for High-Dimensional Boltzmann Sampling

The scalability of continuous normalizing flows (CNFs) for unbiased Boltzmann sampling remains limited in high-dimensional systems due to the cost of Jacobian-determinant evaluation, which requires $D$ backpropagation passes through the flow layers. Existing stochastic Jacobian estimators such as the Hutchinson trace estimator reduce computation but introduce bias, while the recently proposed Flow Perturbation method is unbiased yet suffers from high variance. We present \textbf{Flow Perturbation++}, a variance-reduced extension of Flow Perturbation that discretizes the probability-flow ODE and performs unbiased stepwise Jacobian estimation at each integration step. This multi-step construction retains the unbiasedness of Flow Perturbation while achieves substantially lower estimator variance. Integrated into a Sequential Monte Carlo framework, Flow Perturbation++ achieves significantly improved equilibrium sampling on a 1000D Gaussian Mixture Model and the all-atom Chignolin protein compared with Hutchinson-based and single-step Flow Perturbation baselines.

cs.LG↗

arXiv2026

Rethinking Refinement: Correcting Generative Bias without Noise Injection

Generative models, including diffusion and flow-based models, often exhibit systematic biases that degrade sample quality, particularly in high-dimensional settings. We revisit refinement methods and show that effective bias correction can be achieved as a post-hoc procedure, without noise injection or multi-step resampling of the sampling process. We propose a flow-matching-based \textbf{Bi-stage Flow Refinement (BFR)} framework with two refinement strategies operating at different stages: latent space alignment for approximately invertible generators and data space refinement trained with lightweight augmentations. Unlike previous refiners that perturb sampling dynamics, BFR preserves the original ODE trajectory and applies deterministic corrections to generated samples. Experiments on MNIST, CIFAR-10, and FFHQ at 256x256 resolution demonstrate consistent improvements in fidelity and coverage; notably, starting from base samples with FID 3.95, latent space refinement achieves a \textbf{state-of-the-art} FID of \textbf{1.46} on MNIST using only a single additional function evaluation (1-NFE), while maintaining sample diversity.

cs.LG↗

arXiv2026

Reducing False Positives in Static Bug Detection with LLMs: An Empirical Study in Industry

Static analysis tools (SATs) are widely adopted in both academia and industry for improving software quality, yet their practical use is often hindered by high false positive rates, especially in large-scale enterprise systems. These false alarms demand substantial manual inspection, creating severe inefficiencies in industrial code review. While recent work has demonstrated the potential of large language models (LLMs) for false alarm reduction on open-source benchmarks, their effectiveness in real-world enterprise settings remains unclear. To bridge this gap, we conduct the first comprehensive empirical study of diverse LLM-based false alarm reduction techniques in an industrial context at Tencent, one of the largest IT companies in China. Using data from Tencent's enterprise-customized SAT on its large-scale Advertising and Marketing Services software, we construct a dataset of 433 alarms (328 false positives, 105 true positives) covering three common bug types. Through interviewing developers and analyzing the data, our results highlight the prevalence of false positives, which wastes substantial manual effort (e.g., 10-20 minutes of manual inspection per alarm). Meanwhile, our results show the huge potential of LLMs for reducing false alarms in industrial settings (e.g., hybrid techniques of LLM and static analysis eliminate 94-98% of false positives with high recall). Furthermore, LLM-based techniques are cost-effective, with per-alarm costs as low as 2.1-109.5 seconds and $0.0011-$0.12, representing orders-of-magnitude savings compared to manual review. Finally, our case analysis further identifies key limitations of LLM-based false alarm reduction in industrial settings.

cs.SE↗

arXiv2026

CI4A: Semantic Component Interfaces for Agents Empowering Web Automation

While Large Language Models demonstrate remarkable proficiency in high-level semantic planning, they remain limited in handling fine-grained, low-level web component manipulations. To address this limitation, extensive research has focused on enhancing model grounding capabilities through techniques such as Reinforcement Learning. However, rather than compelling agents to adapt to human-centric interfaces, we propose constructing interaction interfaces specifically optimized for agents. This paper introduces Component Interface for Agent (CI4A), a semantic encapsulation mechanism that abstracts the complex interaction logic of UI components into a set of unified tool primitives accessible to agents. We implemented CI4A within Ant Design, an industrial-grade front-end framework, covering 23 categories of commonly used UI components. Furthermore, we developed a hybrid agent featuring an action space that dynamically updates according to the page state, enabling flexible invocation of available CI4A tools. Leveraging the CI4A-integrated Ant Design, we refactored and upgraded the WebArena benchmark to evaluate existing SoTA methods. Experimental results demonstrate that the CI4A-based agent significantly outperforms existing approaches, achieving a new SoTA task success rate of 86.3%, alongside substantial improvements in execution efficiency.

cs.AI↗

arXiv2026

A Survey of Fuzzing Open-Source Operating Systems

Vulnerabilities in open-source operating systems (OSs) pose substantial security risks to software systems, making their detection crucial. While fuzzing has been an effective vulnerability detection technique in various domains, OS fuzzing (OSF) faces unique challenges due to OS complexity and multi-layered interaction, and has not been comprehensively reviewed. Therefore, this work systematically surveys the state-of-the-art OSF techniques, categorizes them based on the general fuzzing process, and investigates challenges specific to kernel, file system, driver, and hypervisor fuzzing. Finally, future research directions for OSF are discussed.

cs.OS↗

arXiv2025

Large Language Model-Based Agents for Software Engineering: A Survey

The recent advance in Large Language Models (LLMs) has shaped a new paradigm of AI agents, i.e., LLM-based agents. Compared to standalone LLMs, LLM-based agents substantially extend the versatility and expertise of LLMs by enhancing LLMs with the capabilities of perceiving and utilizing external resources and tools. To date, LLM-based agents have been applied and shown remarkable effectiveness in Software Engineering (SE). The synergy between multiple agents and human interaction brings further promise in tackling complex real-world SE problems. In this work, we present a comprehensive and systematic survey on LLM-based agents for SE. We collect 124 papers and categorize them from two perspectives, i.e., the SE and agent perspectives. In addition, we discuss open challenges and future directions in this critical domain. The repository of this survey is at https://github.com/FudanSELab/Agent4SE-Paper-List.

cs.SE↗

arXiv2025

Robust, Observable, and Evolvable Agentic Systems Engineering: A Principled Framework Validated via the Fairy GUI Agent

The Agentic Paradigm faces a significant Software Engineering Absence, yielding Agentic systems commonly lacking robustness, observability, and evolvability. To address these deficiencies, we propose a principled engineering framework comprising Runtime Goal Refinement (RGR), Observable Cognitive Architecture (OCA), and Evolutionary Memory Architecture (EMA). In this framework, RGR ensures robustness and intent alignment via knowledge-constrained refinement and human-in-the-loop clarification; OCA builds an observable and maintainable white-box architecture using component decoupling, logic layering, and state-control separation; and EMA employs an execution-evolution dual-loop for evolvability. We implemented and empirically validated Fairy, a mobile GUI agent based on this framework. On RealMobile-Eval, our novel benchmark for ambiguous and complex tasks, Fairy outperformed the best SoTA baseline in user requirement completion by 33.7%. Subsequent controlled experiments, human-subject studies, and ablation studies further confirmed that the RGR enhances refinement accuracy and prevents intent deviation; the OCA improves maintainability; and the EMA is crucial for long-term performance. This research provides empirically validated specifications and a practical blueprint for building reliable, observable, and evolvable Agentic AI systems.

cs.AI↗

arXiv2025

Evolving Triple Knowledge-Augmented LLMs for Code Translation in Repository Context

Large language models (LLMs) have behaved well in function-level code translation without repository-level context. However, the performance of LLMs in repository-level context code translation remains suboptimal due to complex dependencies and context, hindering their adoption in industrial settings. In this work, we propose a novel LLM-based code translation technique K-Trans, which leverages triple knowledge augmentation to enhance LLM's translation quality under repository context in real-world software development. First, K-Trans constructs a evolving translation knowledge base by extracting relevant information from target-language codebases, the repository being translated, and prior translation results. Second, for each function to be translated, K-Trans retrieves relevant triple knowledge, including target-language code samples, dependency usage examples, and successful translation function pairs, serving as references to enhance LLM for translation. Third, K-Trans constructs a knowledge-augmented translation prompt using the retrieved triple knowledge and employs LLMs to generate the translated code while preserving repository context. It further leverages LLMs for self-debugging, enhancing translation correctness. Lastly, K-Trans continuously evolves the translation knowledge base. The experiments show that K-Trans substantially outperforms the baseline adapted from previous work by 19.4%/40.2% relative improvement in pass@1 and 0.138 in CodeBLEU. It is important to note that the results also demonstrate that each knowledge significantly contributes to K-Trans's effectiveness in handling repository-level context code translation, with dependency usage examples making the most notable contribution. Moreover, as the self-evolution process progresses, the knowledge base continuously enhances the LLM's performance across various aspects of the repository-level code translation.

cs.SE↗

arXiv2025

ExpertAD: Enhancing Autonomous Driving Systems with Mixture of Experts

Recent advancements in end-to-end autonomous driving systems (ADSs) underscore their potential for perception and planning capabilities. However, challenges remain. Complex driving scenarios contain rich semantic information, yet ambiguous or noisy semantics can compromise decision reliability, while interference between multiple driving tasks may hinder optimal planning. Furthermore, prolonged inference latency slows decision-making, increasing the risk of unsafe driving behaviors. To address these challenges, we propose ExpertAD, a novel framework that enhances the performance of ADS with Mixture of Experts (MoE) architecture. We introduce a Perception Adapter (PA) to amplify task-critical features, ensuring contextually relevant scene understanding, and a Mixture of Sparse Experts (MoSE) to minimize task interference during prediction, allowing for effective and efficient planning. Our experiments show that ExpertAD reduces average collision rates by up to 20% and inference latency by 25% compared to prior methods. We further evaluate its multi-skill planning capabilities in rare scenarios (e.g., accidents, yielding to emergency vehicles) and demonstrate strong generalization to unseen urban environments. Additionally, we present a case study that illustrates its decision-making process in complex driving scenarios.

cs.RO↗

arXiv2025

Argus: Resilience-Oriented Safety Assurance Framework for End-to-End ADSs

End-to-end autonomous driving systems (ADSs), with their strong capabilities in environmental perception and generalizable driving decisions, are attracting growing attention from both academia and industry. However, once deployed on public roads, ADSs are inevitably exposed to diverse driving hazards that may compromise safety and degrade system performance. This raises a strong demand for resilience of ADSs, particularly the capability to continuously monitor driving hazards and adaptively respond to potential safety violations, which is crucial for maintaining robust driving behaviors in complex driving scenarios. To bridge this gap, we propose a runtime resilience-oriented framework, Argus, to mitigate the driving hazards, thus preventing potential safety violations and improving the driving performance of an ADS. Argus continuously monitors the trajectories generated by the ADS for potential hazards and, whenever the EGO vehicle is deemed unsafe, seamlessly takes control through a hazard mitigator. We integrate Argus with three state-of-the-art end-to-end ADSs, i.e., TCP, UniAD and VAD. Our evaluation has demonstrated that Argus effectively and efficiently enhances the resilience of ADSs, improving the driving score of the ADS by up to 150.30% on average, and preventing up to 64.38% of the violations, with little additional time overhead.

cs.AI↗

arXiv2025

Code Digital Twin: Empowering LLMs with Tacit Knowledge for Complex Software Development

Recent advances in large language models (LLMs) have demonstrated strong capabilities in software engineering tasks, raising expectations of revolutionary productivity gains. However, enterprise software development is largely driven by incremental evolution, where challenges extend far beyond routine coding and depend critically on tacit knowledge, including design decisions at different levels and historical trade-offs. To achieve effective AI-powered support for complex software development, we should align emerging AI capabilities with the practical realities of enterprise development. To this end, we systematically identify challenges from both software and LLM perspectives. Alongside these challenges, we outline opportunities where AI and structured knowledge frameworks can enhance decision-making in tasks such as issue localization and impact analysis. To address these needs, we propose the Code Digital Twin, a living framework that models both the physical and conceptual layers of software, preserves tacit knowledge, and co-evolves with the codebase. By integrating hybrid knowledge representations, multi-stage extraction pipelines, incremental updates, LLM-empowered applications, and human-in-the-loop feedback, the Code Digital Twin transforms fragmented knowledge into explicit and actionable representations. Our vision positions it as a bridge between AI advancements and enterprise software realities, providing a concrete roadmap toward sustainable, intelligent, and resilient development and evolution of ultra-complex systems.

cs.SE↗

arXiv2025

RustRepoTrans: Repository-level Code Translation Benchmark Targeting Rust

Recent advancements in large language models (LLMs) have demonstrated impressive capabilities in code translation, typically evaluated using benchmarks like CodeTransOcean and RepoTransBench. However, dependency-free benchmarks fail to capture real-world complexities by focusing primarily on simple function-level translations and overlooking repository-level context (e.g., dependencies). Full-repository translation benchmarks significantly exceed the current capabilities of existing models, resulting in performance bottlenecks that fail to provide actionable insights for guiding model development. Furthermore, existing benchmarks do not account for the scenario of incrementally translating new or modified modules from the source to the target language, which demands careful handling of repository-level contexts such as dependencies, cross-module references, and architectural divergence. Moreover, LLMs' effectiveness in translating to newer, low-resource languages like Rust remains largely underexplored. To address these gaps, we introduce RustRepoTrans, the first repository-level context code translation benchmark targeting incremental translation, comprising 375 tasks translating into Rust from C, Java, and Python. Using this benchmark, we evaluate seven representative LLMs, analyzing their errors to assess limitations in complex translation scenarios. Among them, DeepSeek-R1 performs best with 51.5% Pass@1, excelling in both basic functionality and additional translation abilities, such as noise robustness and syntactical difference identification. However, even DeepSeek-R1 experiences a 22.2% performance drop (Pass@1 from 73.7% to 51.5%) when handling repository-level context compared to previous benchmarks without such context.

cs.SE↗

arXiv2025

PARL-MT: Learning to Call Functions in Multi-Turn Conversation with Progress Awareness

Large language models (LLMs) have achieved impressive success in single-turn function calling, yet real-world applications such as travel planning or multi-stage data analysis typically unfold across multi-turn conversations. In these settings, LLMs must not only issue accurate function calls at each step but also maintain progress awareness, the ability to summarize past interactions and plan future actions to ensure coherent, long-horizon task execution. Existing approaches, however, either reduce multi-turn training to isolated single-turn samples, which neglects task-level planning, or employ end-to-end reinforcement learning (RL) that struggles with redundancy and lacks explicit integration of progress awareness. To overcome these limitations, we introduce PARL-MT, a framework that explicitly incorporates progress awareness into LLM training for multi-turn function calling. PARL-MT combines (i) a Progress Awareness Generation (PAG) pipeline, which automatically constructs datasets coupling conversation summaries with future task planning, and (ii) a Progress Awareness-Guided Reinforcement Learning (PAG-RL) algorithm, which integrates progress awareness into RL training to reduce contextual redundancy and improve alignment between local actions and global task completion. Empirical results on two public benchmarks demonstrate that PARL-MT significantly outperforms existing methods, highlighting the effectiveness of progress awareness in enabling robust and efficient multi-turn function calling.

cs.CL↗

arXiv2025

Generating High-Quality Datasets for Code Editing via Open-Source Language Models

Code editing plays a vital role in software engineering, requiring developers to adjust existing code according to natural language instructions while keeping functionality intact and avoiding unnecessary modifications. However, commit-based datasets commonly used for this task are often noisy, lack diversity, and fail to reflect the style of real-world edit instructions. To address this, we introduce OpenCodeEdit, an open-source pipeline that leverages multiple LLMs to synthesize realistic code-edit triplets. The pipeline produces both concise "lazy" instructions and more detailed "descriptive" ones, and applies filtering based on diffs and topics to guarantee data quality and variety. Using this process, we construct OCEDataFT, a curated dataset of 20K samples. Fine-tuning three advanced base models on OCEDataFT leads to significant performance boosts on the CanItEdit benchmark, with relative pass@1 improvements ranging from 4.50% to 20.79%. Notably, the resulting models achieve performance close to closed-source systems, narrowing the gap to GPT-4 to just 3.54%, without relying on proprietary resources or manual annotation.

cs.SE↗

arXiv2025

EvolMathEval: Towards Evolvable Benchmarks for Mathematical Reasoning via Evolutionary Testing

The rapid advancement of Large Language Models (LLMs) poses a significant challenge to existing mathematical reasoning benchmarks. However, these benchmarks tend to become easier over time as LLMs can learn from the published benchmarks. This limitation hinder the precise evaluation of the true capabilities of SOTA models. To address this challenge, this paper introduces EvolMathEval, an automated mathematical benchmark generation and evolution framework based on evolutionary testing. Experimental results demonstrate that EvolMathEval can not only generate a large volume of high-difficulty problems through continuous self-iteration, but it can also significantly enhance the complexity of public datasets like GSM8K through evolution, reducing model accuracy by an average of 48\%. Deeper investigation reveals that when solving these evolved problems, LLMs tend to bypass complex multi-step logical reasoning by relying on simplistic and fuzzy conditions, consequently leading to incorrect solutions. We define this phenomenon as the ``Pseudo Aha Moment", which we find accounts for 77\% to 100\% of errors on targeted problems. Code and resources are available at: https://anonymous.4open.science/r/EvolMathEval

cs.AI↗

arXiv2025

Extracting Conceptual Knowledge to Locate Software Issues

Issue localization, which identifies faulty code elements such as files or functions, is critical for effective bug fixing. While recent LLM-based and LLM-agent-based approaches improve accuracy, they struggle in large-scale repositories due to concern tangling, where relevant logic is buried in large functions, and concern scattering, where related logic is dispersed across files. To address these challenges, we propose RepoLens, a novel approach that abstracts and leverages conceptual knowledge from code repositories. RepoLens decomposes fine-grained functionalities and recomposes them into high-level concerns, semantically coherent clusters of functionalities that guide LLMs. It operates in two stages: an offline stage that extracts and enriches conceptual knowledge into a repository-wide knowledge base, and an online stage that retrieves issue-specific terms, clusters and ranks concerns by relevance, and integrates them into localization workflows via minimally intrusive prompt enhancements. We evaluate RepoLens on SWE-Lancer-Loc, a benchmark of 216 tasks derived from SWE-Lancer. RepoLens consistently improves three state-of-the-art tools, namely AgentLess, OpenHands, and mini-SWE-agent, achieving average gains of over 22% in Hit@k and 46% in Recall@k for file- and function-level localization. It generalizes across models (GPT-4o, GPT-4o-mini, GPT-4.1) with Hit@1 and Recall@10 gains up to 504% and 376%, respectively. Ablation studies and manual evaluation confirm the effectiveness and reliability of the constructed concerns.

cs.SE↗