arXiv subjects

Jiajun Jiang

Publications and source records attributed to Jiajun Jiang.

At least 19 recordsLinked to original sources

arXiv2026

PolyCodeEval: Benchmarking Multilingual Code Generation from Functions to Repositories

As large language models increasingly move toward repository-level software engineering, existing code-generation benchmarks remain fragmented across language coverage, task granularity, and evaluation protocols, impeding systematic comparison. To address this gap, we present PolyCodeEval, a unified multilingual and multi-granularity benchmark for code generation. It comprises 2,590 code generation tasks spanning functions to repositories, derived from 58 real, executable open-source repositories in five programming languages. All tasks are evaluated under a unified execution-based protocol with integration procedures tailored to their generation targets. Building on this benchmark, we evaluate frontier large language models, state-of-the-art specialized methods, and general coding agents. Our results show that existing approaches still struggle to correctly generate complete code fragments across granularities and languages. Specifically, the studied methods generate at most 71.7%, 76.7%, and 31.0% correct functions, files, and repositories, respectively, with performance varying widely across languages. Paired experiments further show that implementation context from related functions in the same file improves the executable correctness of function generation. Method rankings also vary across task granularities and programming languages, highlighting the importance of multilingual, multi-granularity evaluation for comprehensively assessing code generation capabilities.

cs.SE↗

arXiv2026

LLM-based Vulnerability Detection at Project Scale: An Empirical Study

As software complexity grows, automated vulnerability detection becomes increasingly important. Recent LLM-based detectors combine semantic reasoning with static analysis for project-scale scanning, but their practical effectiveness and failure causes remain unclear. We present the first comprehensive empirical study of specialized LLM-based detectors at project scale, evaluating five specialized methods, two general-purpose agents, and four traditional static analyzers on 265 known C/C++ and Java vulnerabilities and 24 active open-source projects/modules. Using Codex-assisted labeling with stratified manual validation, we analyze 6,442 sampled warnings and classify 5,896 false positives using a taxonomy derived from 355 manually inspected reports. Our study yields three findings. First, general-purpose agents achieve the highest recall but struggle with complex code, while both LLM-based and traditional methods suffer from incomplete API modeling. Second, many tools exhibit high false discovery rates on real-world projects/modules. Truncated interprocedural context and incorrect program-point classification are the dominant false-positive causes, with LLM-based tools exhibiting additional reasoning and prompt-compliance failures. Third, LLM-based methods incur substantial costs: hundreds of thousands to hundreds of millions of tokens, API charges exceeding $1,000 per project/module, and runtimes ranging from hours to days. These findings reveal limitations in the robustness, reliability, and scalability of current detectors and inform directions for more effective and practical project-scale vulnerability detection.

cs.SE↗

arXiv2026

SatNav: A Scalable Benchmark for Long-Horizon UAV Vision-Language Navigation from Satellite Imagery

Urban uncrewed aerial vehicle (UAV) vision-language navigation (VLN) requires agents to follow instructions across extended urban spaces, inherently demanding long-term memory and geospatial grounding. However, scaling existing benchmarks remains difficult because of their reliance on costly reconstructed 3D assets, limiting geographic diversity and episode scale. To address this, we introduce SatNav, a scalable, long-horizon UAV VLN benchmark built from high-resolution satellite imagery. SatNav targets city-level navigation missions and uses satellite crops as approximations of UAV nadir views for visual observations. Through an automated cue-to-episode pipeline, SatNav constructs 118K episodes from 59 scenes across 18 cities, with an average trajectory length of 379 m. To stress-test long-horizon memory and geospatial reasoning, SatNav defines three task families: Boundary, Landmark, and Route, targeting loop progress tracking, landmark-based spatial grounding, and route following with counting cues. Benchmarking classical VLN agents and recent agents based on large vision-language models (LVLMs) on SatNav shows that city-scale navigation remains challenging. We further introduce SwiftVLN, a modular framework with switchable memory components, and conduct systematic memory-design ablations. Finally, satellite-to-UAV transfer experiments show that satellite-trained navigation models can operate on real-flight UAV observations, showing the practical relevance of SatNav. Our project page: https://eku127.github.io/SatNav/

cs.CV↗

arXiv2026

Accurate Measurement of 3D and 2D Circular Centers With Application to LiDAR-Camera Extrinsic Calibration

Accurate measurement of circular centers is a fun-damental geometric sensing problem in instrumentation and measurement tasks involving cameras, LiDARs, and other spa-tial sensors. In circular-target-based LiDAR-camera extrinsic calibration, a 3D circular center measured from LiDAR and its corresponding 2D projected center measured in the image serve as cross-modal geometric observations for estimating the rigid transformation between the two sensor frames. Conven-tional pipelines can bias both measurements: the 3D center is often obtained by sequential plane fitting, point projection, and 2D circle fitting, while the image ellipse center is frequently treated as the projected circle center even though the two generally differ under perspective projection. This paper focuses on accurate 3D and 2D circular-center measurement and uses LiDAR-camera extrinsic calibration as a representative applica-tion. A conformal-geometric-algebra estimator is integrated with RANSAC to jointly recover the 3D center, normal, and radius from noisy or partially observed LiDAR points. A chord-length-variance criterion then estimates the 2D projected center, with its twofold ambiguity resolved by homography validation or a quasi-RANSAC fallback. Synthetic and real-sensor experiments show that the proposed method improves circular-center measurement accuracy and reduces extrinsic calibration error across different target and sensor configurations.

cs.CV↗

arXiv2026

Toward Understanding Operating System Defects

Operating systems (OS) serve as the foundation for all other software systems, and thus defects in OSes can lead to severe conquences, such as system crashes and data corruption, affecting billions of users. This broad impact underscores the necessity and importance of ensuring OS quality. Understanding the characteristics of OS defects is a fundamental step in this quality assurance task, as it facilitates the design of effective defect detection and debugging approaches. In this work, we conduct a large-scale study of 1,500 defects from three distinct and representative operating systems (Android, Linux, and HarmonyOS) spanning both mobile and desktop environments. To the best of our knowledge, this is the largest study of its kind in this domain. By analyzing the distribution of OS defects across multiple classification dimensions, including the OS layer where defects occur, the functions they affect, how they are triggered, their severity, and the code elements involved in their repair, as well as performing joint analysis between dimensions and cross-OS similarity comparisons, we summarize several major findings that contribute to a comprehensive understanding of OS defects across systems. Based on these findings, we provide a series of actionable implications for better OS defect detection and debugging, offering guidelines for future research aimed at improving OS quality assurance.

cs.SE↗

arXiv2026

An Extensive Empirical Study on Code Translation Technique

Automated code translation is increasingly important for software evolution, yet the relative strengths and limitations of learning-based and large language model (LLM)-based techniques remain insufficiently understood. To address this gap, we conduct a large-scale empirical study comparing representative code translation techniques across methodological paradigms and translation granularities. We evaluate learning-based methods, LLM-based methods, and general-purpose LLMs on multilingual method-level and class-level benchmarks involving multiple programming languages. Our analysis considers executable correctness, code similarity, translation direction, translation granularity, and failure patterns. The results show that LLMs and LLM-based methods generally outperform learning-based methods in method-level correctness, although similarity metrics alone do not reliably reflect functional correctness. Translation direction substantially affects performance, particularly when translating between languages with different type-system characteristics. Class-level translation remains considerably more difficult than method-level translation because it requires preserving global semantics, interfaces, member relationships, and cross-method dependencies. Our error analysis further shows that static semantic errors and logical errors are the primary challenges in existing code translation systems. These findings provide empirical evidence and practical guidance for developing more robust, type-aware, structure-aware, and context-aware code translation techniques.

cs.SE↗

arXiv2026

Beyond Pass@k: Measuring Reliability and Security of Agentic Code Generation

AI coding agent benchmarks rank agents with the Chen et al. (2021) pass@k estimator, but current implementations misapply it: they set n to the number of unit tests in a single submission rather than the number of independent rollout attempts, conflating test-suite size with attempt independence. We diagnose this operationalization error, prove it by counterexample, and propose reliability@k, the same estimator applied correctly, with n = independent rollouts and c = fully-passing rollouts per (task, agent) pair. In a synthetic multi-rollout benchmark, the misapplied metric inflates reported scores by 0.85-0.97 in absolute terms (0.96-0.98 reported vs. 0.00-0.12 corrected), and a cheap single-rollout proxy fails to substitute for repeated runs (Spearman $ρ= 0.417$). Motivated by evidence that functional correctness does not imply security safety, we additionally propose security-adjusted reliability@k, which counts only rollouts that are both functionally correct and free of high-severity insecure patterns. In an initial live-API test with three agents, the adjustment did not change any ranking under our current scanner and threshold, so we present it as a proposed complementary lens whose decisive evaluation requires better-powered future runs. Finally, a preliminary 5-task SWE-bench Verified pilot observes the same core concern in a real repository setting: macro-averaged hidden-test pass rate was 0.80 while strict task resolution was 0.20.

cs.AI↗

arXiv2026

Towards Better Linux Kernel Fault Localization: Leveraging Contrastive Reasoning and Hierarchical Context Analysis

Debugging the Linux kernel remains a formidable challenge due to its vast codebase, complex architecture, and low-level programming intricacies. Effective fault localization (FL) is thus essential for efficient kernel debugging and maintenance. While existing FL techniques (both traditional and LLM-based) have shown promise in general-purpose software, they are ill-suited for the kernel context. In particular, recent LLM-based techniques often treat bug reports and source code as plain text, lacking deep integration of kernel-specific knowledge, which limits their ability to identify root causes and achieve fine-grained localization. We present CoHiKer, a novel LLM-based FL technique tailored to the Linux kernel. CoHiKer introduces two key innovations: (1) contrastive reasoning, which identifies root causes by analyzing the behavioral divergence between carefully mutated passing and failing test cases, and (2) hierarchical context analysis, which systematically narrows the localization scope from files to methods by integrating crash reports, syscall semantics, inter-file dependencies, and kernel-specific features. Unlike prior techniques that rely on static understanding and full-code input, CoHiKer decomposes the localization task and enables structured LLM prompting to reason semantically over meaningful contexts. We evaluate CoHiKer on an extended Linux kernel bug dataset against five state-of-the-art baselines. CoHiKer consistently outperforms all competitors, improving Top-1 localization accuracy by up to 26.07% at the file level and 56.85% at the method level over state-of-the-art LLM-based baselines, while achieving up to 8.84% and 28.9% reductions in token consumption, respectively. Furthermore, CoHiKer demonstrates strong generalizability on the non-kernel dataset, with comparable gains (15.5% and 5.3% in Top-1 at file and method levels).

cs.SE↗

arXiv2026

ERNIE 5.0 Technical Report

In this report, we introduce ERNIE 5.0, a natively autoregressive foundation model desinged for unified multimodal understanding and generation across text, image, video, and audio. All modalities are trained from scratch under a unified next-group-of-tokens prediction objective, based on an ultra-sparse mixture-of-experts (MoE) architecture with modality-agnostic expert routing. To address practical challenges in large-scale deployment under diverse resource constraints, ERNIE 5.0 adopts a novel elastic training paradigm. Within a single pre-training run, the model learns a family of sub-models with varying depths, expert capacities, and routing sparsity, enabling flexible trade-offs among performance, model size, and inference latency in memory- or time-constrained scenarios. Moreover, we systematically address the challenges of scaling reinforcement learning to unified foundation models, thereby guaranteeing efficient and stable post-training under ultra-sparse MoE architectures and diverse multimodal settings. Extensive experiments demonstrate that ERNIE 5.0 achieves strong and balanced performance across multiple modalities. To the best of our knowledge, among publicly disclosed models, ERNIE 5.0 represents the first production-scale realization of a trillion-parameter unified autoregressive model that supports both multimodal understanding and generation. To facilitate further research, we present detailed visualizations of modality-agnostic expert routing in the unified model, alongside comprehensive empirical analysis of elastic training, aiming to offer profound insights to the community.

cs.CL↗

arXiv2025

DualMap: Online Open-Vocabulary Semantic Mapping for Natural Language Navigation in Dynamic Changing Scenes

We introduce DualMap, an online open-vocabulary mapping system that enables robots to understand and navigate dynamically changing environments through natural language queries. Designed for efficient semantic mapping and adaptability to changing environments, DualMap meets the essential requirements for real-world robot navigation applications. Our proposed hybrid segmentation frontend and object-level status check eliminate the costly 3D object merging required by prior methods, enabling efficient online scene mapping. The dual-map representation combines a global abstract map for high-level candidate selection with a local concrete map for precise goal-reaching, effectively managing and updating dynamic changes in the environment. Through extensive experiments in both simulation and real-world scenarios, we demonstrate state-of-the-art performance in 3D open-vocabulary segmentation, efficient scene mapping, and online language-guided navigation. Project page: https://eku127.github.io/DualMap/

cs.RO↗

arXiv2025

SwarmSys: Decentralized Swarm-Inspired Agents for Scalable and Adaptive Reasoning

Large language model (LLM) agents have shown remarkable reasoning abilities. However, existing multi-agent frameworks often rely on fixed roles or centralized control, limiting scalability and adaptability in long-horizon reasoning. We introduce SwarmSys, a closed-loop framework for distributed multi-agent reasoning inspired by swarm intelligence. Coordination in SwarmSys emerges through iterative interactions among three specialized roles, Explorers, Workers, and Validators, that continuously cycle through exploration, exploitation, and validation. To enable scalable and adaptive collaboration, we integrate adaptive agent and event profiles, embedding-based probabilistic matching, and a pheromone-inspired reinforcement mechanism, supporting dynamic task allocation and self-organizing convergence without global supervision. Across symbolic reasoning, research synthesis, and scientific programming tasks, SwarmSys consistently outperforms baselines, improving both accuracy and reasoning stability. These findings highlight swarm-inspired coordination as a promising paradigm for scalable, robust, and adaptive multi-agent reasoning, suggesting that coordination scaling may rival model scaling in advancing LLM intelligence.

cs.AI↗

arXiv2025

Boosting Redundancy-based Automated Program Repair by Fine-grained Pattern Mining

Redundancy-based automated program repair (APR), which generates patches by referencing existing source code, has gained much attention since they are effective in repairing real-world bugs with good interpretability. However, since existing approaches either demand the existence of multi-line similar code or randomly reference existing code, they can only repair a small number of bugs with many incorrect patches, hindering their wide application in practice. In this work, we aim to improve the effectiveness of redundancy-based APRs by exploring more effective source code reuse methods for improving the number of correct patches and reducing incorrect patches. Specifically, we have proposed a new repair technique named Repatt, which incorporates a two-level pattern mining process for guiding effective patch generation (i.e., token and expression levels). We have conducted an extensive experiment on the widely-used Defects4J benchmark and compared Repatt with ten state-of-the-art APR approaches. The results show that it complements existing approaches by repairing 9 unique bugs compared with the latest Large Language Model (LLM)-based and deep learning-based methods and 19 unique bugs compared with traditional repair methods when providing the perfect fault localization. In addition, when the perfect fault localization is unknown in real practice, Repatt significantly outperforms the baseline approaches by achieving much higher patch precision, i.e., 83.8\%, although it repairs fewer bugs. Moreover, we further proposed an effective patch ranking strategy for combining the strength of Repatt and the baseline methods. The result shows that it repairs 124 bugs when only considering the Top-1 patches and improves the best-performing repair method by repairing 39 more bugs. The results demonstrate the effectiveness of our approach for practical use.

cs.SE↗

arXiv2025

Improving Compiler Bug Isolation by Leveraging Large Language Models

Compilers play a foundational role in building reliable software systems, and bugs within them can lead to catastrophic consequences. The compilation process typically involves hundreds of files, making traditional automated bug isolation techniques inapplicable due to scalability or effectiveness issues. Current mainstream compiler bug localization techniques have limitations in test program mutation and resource consumption. Inspired by the recent advances of pre-trained Large Language Models (LLMs), we propose an innovative approach named AutoCBI, which (1) uses LLMs to summarize compiler file functions and (2) employs specialized prompts to guide LLM in reordering suspicious file rankings. This approach leverages four types of information: the failing test program, source file function summaries, lists of suspicious files identified through analyzing test coverage, as well as compilation configurations with related output messages, resulting in a refined ranking of suspicious files. Our evaluation of AutoCBI against state-of-the-art approaches (DiWi, RecBi and FuseFL) on 120 real-world bugs from the widely-used GCC and LLVM compilers demonstrates its effectiveness. Specifically, AutoCBI isolates 66.67%/69.23%, 300%/340%, and 100%/57.14% more bugs than RecBi, DiWi, and FuseFL, respectively, in the Top-1 ranked results for GCC/LLVM. Additionally, the ablation study underscores the significance of each component in our approach.

cs.SE↗

arXiv2025

Empirical Evaluation of Large Language Models in Automated Program Repair

The increasing prevalence of software bugs has made automated program repair (APR) a key research focus. Large language models (LLMs) offer new opportunities for APR, but existing studies mostly rely on smaller, earlier-generation models and Java benchmarks. The repair capabilities of modern, large-scale LLMs across diverse languages and scenarios remain underexplored. To address this, we conduct a comprehensive empirical study of four open-source LLMs, CodeLlama, LLaMA, StarCoder, and DeepSeek-Coder, spanning 7B to 33B parameters, diverse architectures, and purposes. We evaluate them across two bug scenarios (enterprise-grades and algorithmic), three languages (Java, C/C++, Python), and four prompting strategies, analyzing over 600K generated patches on six benchmarks. Key findings include: (1) model specialization (e.g., CodeLlama) can outperform larger general-purpose models (e.g., LLaMA); (2) repair performance does not scale linearly with model size; (3) correct patches often appear early in generation; and (4) prompts significantly affect results. These insights offer practical guidance for designing effective and efficient LLM-based APR systems.

cs.SE↗

arXiv2025

DESIL: Detecting Silent Bugs in MLIR Compiler Infrastructure

MLIR (Multi-Level Intermediate Representation) compiler infrastructure provides an efficient framework for introducing a new abstraction level for programming languages and domain-specific languages. It has attracted widespread attention in recent years and has been applied in various domains, such as deep learning compiler construction. Recently, several MLIR compiler fuzzing techniques, such as MLIRSmith and MLIRod, have been proposed. However, none of them can detect silent bugs, i.e., bugs that incorrectly optimize code silently. The difficulty in detecting silent bugs arises from two main aspects: (1) UB-Free Program Generation: Ensures the generated programs are free from undefined behaviors to suit the non-UB assumptions required by compiler optimizations. (2) Lowering Support: Converts the given MLIR program into an executable form, enabling execution result comparisons, and selects a suitable lowering path for the program to reduce redundant lowering pass and improve the efficiency of fuzzing. To address the above issues, we propose DESIL. DESIL enables silent bug detection by defining a set of UB-elimination rules based on the MLIR documentation and applying them to input programs to produce UB-free MLIR programs. To convert dialects in MLIR program into the executable form, DESIL designs a lowering path optimization strategy to convert the dialects in given MLIR program into executable form. Furthermore, DESIL incorporates the differential testing for silent bug detection. To achieve this, it introduces an operation-aware optimization recommendation strategy into the compilation process to generate diverse executable files. We applied DESIL to the latest revisions of the MLIR compiler infrastructure. It detected 23 silent bugs and 19 crash bugs, of which 12/14 have been confirmed or fixed

cs.SE↗

arXiv2025

Evaluating the Generalizability of LLMs in Automated Program Repair

LLM-based automated program repair methods have attracted significant attention for their state-of-the-art performance. However, they were primarily evaluated on a few well known datasets like Defects4J, raising questions about their effectiveness on new datasets. In this study, we evaluate 11 top-performing LLMs on DEFECTS4J-TRANS, a new dataset derived from transforming Defects4J while maintaining the original semantics. Results from experiments on both Defects4J and DEFECTS4J-TRANS show that all studied LLMs have limited generalizability in APR tasks, with the average number of correct and plausible patches decreasing by 49.48% and 42.90%, respectively, on DEFECTS4J-TRANS. Further investigation into incorporating additional repair-relevant information in repair prompts reveals that, although this information significantly enhances the LLMs' capabilities (increasing the number of correct and plausible patches by up to 136.67% and 121.82%, respectively), performance still falls short of their original results. This indicates that prompt engineering alone is insufficient to substantially enhance LLMs' repair capabilities. Based on our study, we also offer several recommendations for future research.

cs.SE↗

arXiv2024

Dependency-Aware Code Naturalness

Code naturalness, which captures repetitiveness and predictability in programming languages, has proven valuable for various code-related tasks in software engineering. However, precisely measuring code naturalness remains a fundamental challenge. Existing methods measure code naturalness over individual lines of code while ignoring the deep semantic relations among different lines, e.g., program dependency, which may negatively affect the precision of the measure. In this study, we aim to perform the first empirical study to investigate whether incorporating code dependency, instead of analyzing individual lines, can enhance the precision of measuring code naturalness. To achieve that, we first propose a new method named DAN for measuring code naturalness by incorporating the rich dependency information in the code. Specifically, DAN extracts multiple sequences of code lines by traversing the program dependency graph, where different code lines are connected by dependencies in each sequence, and then the code naturalness will be measured by taking each sequence as a whole. In this way, the dependency information can be well captured. Finally, we have conducted an extensive study to evaluate the influence of code dependency for measuring code naturalness with DAN, and compared it with the state-of-the-art methods under three emerging application scenarios of code naturalness. The results demonstrate that DAN can not only better distinguish natural and unnatural code, but also substantially boost two important downstream applications of code naturalness, i.e., distinguishing buggy and non-buggy code lines and data cleansing for training better code models, reflecting the significance of code dependency in measuring code naturalness.

cs.SE↗

arXiv2024

Toward Understanding Deep Learning Framework Bugs

DL frameworks are the basis of constructing all DL programs and models, and thus their bugs could lead to the unexpected behaviors of any DL program or model relying on them. Such a wide effect demonstrates the necessity and importance of guaranteeing DL frameworks' quality. Understanding the characteristics of DL framework bugs is a fundamental step for this quality assurance task, facilitating designing effective bug detection and debugging approaches. Hence, in this work we conduct the most large-scale study on 1,000 bugs from four popular and diverse DL frameworks (i.e., TensorFlow, PyTorch, MXNet, and DL4J). By analyzing the root causes and symptoms of DL framework bugs associated with 5 components decomposed from DL frameworks, as well as measuring test coverage achieved by three state-of-the-art testing techniques, we obtain 12 major findings for the comprehensive understanding of DL framework bugs and the current status of existing DL framework testing practice, and then provide a series of actionable guidelines for better DL framework bug detection and debugging. Finally, based on the guidelines, we design and implement a prototype DL-framework testing tool, called TenFuzz, which is evaluated to be effective and finds 3 unknown bugs on the latest TensorFlow framework in a preliminary study, indicating the significance of our guidelines.

cs.SE↗