Institution profile

Nanjing University of Science and Technology

Academic institutionasia · cn
Official website
Research library752linked papers
Opportunities0open roles
Selected work

Representative Papers

A Systematic Literature Review on Large Language Models for Automated Program Repair

May 02, 2024arXiv.org

Research on large language models (LLMs) for automated program repair (APR) remains fragmented and lacks a systematic, unified understanding. Method: We conduct a systematic literature review (SLR) covering 127 papers published between 2020 and 2024, establishing the first comprehensive conceptual framework for LLM-based APR. We categorize model utilization strategies into three types—fine-tuning, prompt engineering, and hybrid ensemble—and perform multidimensional thematic analysis across input representation, semantic/security-specific repair scenarios, and open-science practices. Contribution/Results: We identify core challenges including model robustness, evaluation bias, and real-world deployment adaptability. The study yields a reusable taxonomy, benchmark insights, and methodological guidelines—delivering the APR community’s first holistic landscape map to precisely identify research gaps and inform future innovation pathways.

39 citations1 influentialRead paper

Are the confidence scores of reviewers consistent with the review content? Evidence from top conference proceedings in AI

Jun 20, 2024Scientometrics

This study investigates the alignment between reviewers’ confidence scores and textual content in AI top-conference peer reviews, addressing a gap in fine-grained alignment analysis. We propose the first multi-granular (word-, sentence-, and aspect-level) quantitative alignment framework, integrating deep learning–based aspect extraction and fuzzy sentence detection with NLP features—including sentiment polarity, hedge word/sentence frequency, and comment length—and multi-level statistical tests (correlation, regression, significance). Results show strong alignment between confidence scores and review text across all granularities. Counterintuitively, higher confidence scores significantly predict paper rejection (p < 0.01), supporting reviewer expertise and the credibility of the review process. Our core contribution is the first cross-granular confidence–text alignment paradigm, demonstrating that confidence scores serve not merely as ordinal indicators but as interpretable, high-fidelity proxies for review quality.

5 citationsRead paper

Test smells in LLM-Generated Unit Tests

Oct 14, 2024arXiv.org

This study addresses the understudied issue of test smells in large language model (LLM)-generated unit tests. Method: We conduct the first large-scale empirical analysis of test smells across 20,505 Java class-level test suites drawn from five sources: human-written tests, EvoSuite-generated tests, and LLM-generated tests from GPT-3.5, GPT-4, Mistral, and Mixtral. Our multi-benchmark, cross-model analysis framework encompasses over 770,000 test cases, leveraging dual smell-detection tools—TsDetect and JNose—across 34,635 open-source projects and the TestBench benchmark. Contribution/Results: We identify prevalent smells—including Assertion Roulette and Magic Number Test—in LLM-generated tests; their occurrence patterns are significantly influenced by prompting strategies, context length, and model scale. Notably, LLM-generated tests exhibit smell profiles closer to human-written tests than to search-based software testing (SBST) outputs, suggesting potential training data contamination. These findings provide critical empirical grounding for developing smell-aware test generation frameworks.

4 citationsRead paper

A Comprehensive Evaluation of Parameter-Efficient Fine-Tuning on Method-Level Code Smell Detection

Dec 18, 2024arXiv.org

Existing code smell detection approaches—ranging from heuristic rules to machine learning (ML) and deep learning (DL)—exhibit limited performance, while full fine-tuning of large language models (LLMs) incurs prohibitive computational costs. This work presents the first systematic evaluation of parameter-efficient fine-tuning (PEFT) techniques—including Prompt Tuning, Prefix Tuning, LoRA, and (IA)³—for method-level code bad smells (e.g., Complex Conditional, Long Method) detection. We conduct experiments across four small language models and six LLMs using a high-quality GitHub dataset. Key findings: (1) Small models combined with PEFT significantly outperform both large models with PEFT and even fully fine-tuned LLMs; (2) Training data scale exerts a far greater impact on detection accuracy than the number of tunable parameters; (3) PEFT achieves comparable or superior accuracy to full fine-tuning while drastically reducing GPU memory consumption and consistently surpassing traditional detectors. Our study establishes a novel, lightweight, efficient, and deployable paradigm for code quality analysis.

2 citationsRead paper

Small but Mighty: Dynamic Wavelet Expert-Guided Fine-Tuning of Large-Scale Models for Optical Remote Sensing Object Segmentation

Jan 14, 2026

This work addresses the challenge of deploying large-scale foundation models for object segmentation in optical remote sensing imagery, where full-parameter fine-tuning incurs prohibitive memory and computational costs. To this end, we propose WEFT, an efficient fine-tuning approach guided by dynamic wavelet experts. WEFT introduces, for the first time in remote sensing segmentation, a learnable wavelet expert extractor coupled with a conditional adapter, which enhances the fine-grained perceptual capabilities of a frozen foundation model while tuning only a minimal number of parameters. By integrating wavelet transformation, dynamic expert mechanisms, and parameter-efficient fine-tuning, WEFT outperforms 21 state-of-the-art methods across three remote sensing benchmarks and achieves superior performance on camouflaged, natural, and medical image segmentation tasks, significantly reducing training resource consumption.

1 citationsRead paper
Recent publications

Latest Papers