arXiv subjects

Yunwei Dong

Publications and source records attributed to Yunwei Dong.

5 recordsLinked to original sources

arXiv2026

MGRetrieval: Memory-Guided Reflective Retrieval for Long-Term Dialogue Agents

Large Language Models (LLMs) have made significant progress in dialogue, yet redundant memory contexts severely limit their effectiveness in long-term dialogue agents. External memory systems have been proposed to improve memory maintenance. However, these systems mainly rely on one-shot retrieval, which limits their ability to retrieve sufficient and relevant evidence. Although recent methods introduce reflection into retrieval, their retrieval paths are generated by the LLM from limited evidence, leading to unstable retrieval and additional latency overhead. %These limitations highlight the need for effective retrieval mechanisms. To address these limitations, we propose MGRetrieval, a retrieval strategy that grounds reflective retrieval in the semantic structure of historical memories. Specifically, MGRetrieval consists of two steps: (1) It references the structure of historical memories to construct a more precise retrieval path. (2) The LLM retains critical memories and determines whether accumulated memories are sufficient to stop further iterative retrieval. This allows the retrieval process to follow semantically meaningful paths. Through memory-guided retrieval and critical memory propagation, MGRetrieval gradually constructs concise and sufficient memory contexts. Extensive experiments on LoCoMo show that MGRetrieval outperforms the strongest baseline by 8.91\% in F1 and 11.11\% in BLEU-1 on average across Qwen2.5-14B and Qwen3-14B, while maintaining practical token and latency costs. The code can be found in https://anonymous.4open.science/r/MGRetrieval.

cs.IR↗

arXiv2026

Understanding and Improving Automated Proof Synthesis for Interactive Theorem Provers

Formal verification using interactive theorem provers ensures high-quality software. However, writing proof scripts for interactive theorem provers is labor-intensive and requires deep expertise. Recent studies have leveraged deep learning to automate theorem proving by learning from manually written proof corpora. Nevertheless, these techniques still achieve limited success rates in proof synthesis. This paper investigates the theorems that current proof synthesis techniques are unable to prove and analyzes their characteristics. Through an in-depth analysis of the proven theorems, proof scripts, and the proof search process, we identify the limitations of existing tools and summarize the key factors influencing proof success rates. Our research provides valuable insights for the future optimization of automated proof tools. Based on our empirical study, we discover that tactic selections conforming to human expert proof patterns are more likely to lead to successful proofs. Inspired by this finding, we propose a pattern-guided tactic search (PGTS) method to heuristically enhance the search process of existing proof synthesis tools. Our evaluation experiments demonstrate that PGTS improves the number of theorems proved by existing proof synthesis tools by an average of 8.05\%, while also achieving an average 20\% increase in previously unproven theorems. Furthermore, PGTS enhances the capability of proof synthesis tools to prove complex theorems and generates more concise proof scripts.

cs.LO↗

arXiv2026

From Guessing to Seeing: Enhancing LLM-Based Program Repair via Trace-Guided Multi-strategy Debate

Automated Program Repair (APR) aims to resolve software bugs without human intervention, but complex logic errors and silent failures remain challenging. Existing LLM-based APR methods mainly rely on source code and coarse test feedback, making it difficult to capture runtime behaviors and dynamic data dependencies. Execution traces expose concrete state transitions, yet a single LLM interpreting them in isolation may commit to an incorrect repair hypothesis and produce test-overfitting patches. We therefore treat runtime evidence as shared constraints for validating repair hypotheses rather than merely as additional prompt context. We propose TraceRepair, a multi-agent framework in which a Probe Agent captures execution snapshots of critical variables, while specialized repair agents generate, compare, and iteratively refine candidate patches against the observed runtime evidence. A Judge Agent then arbitrates the remaining hypotheses and synthesizes the final patch. Evaluated on Defects4J, TraceRepair correctly fixes 392 defects and outperforms existing LLM-based approaches. Further experiments demonstrate improved efficiency and strong generalization on a newly constructed dataset of recent bugs, suggesting that the gains arise from dynamic reasoning rather than memorization.

cs.SE↗

arXiv2026

OODEval: Evaluating Large Language Models on Object-Oriented Design

Recent advances in large language models (LLMs) have driven extensive evaluations in software engineering. however, most prior work concentrates on code-level tasks, leaving software design capabilities underexplored. To fill this gap, we conduct a comprehensive empirical study evaluating 29 LLMs on object-oriented design (OOD) tasks. Owing to the lack of standardized benchmarks and metrics, we introduce OODEval, a manually constructed benchmark comprising 50 OOD tasks of varying difficulty, and OODEval-Human, the first human-rated OOD benchmark, which includes 940 undergraduate-submitted class diagrams evaluated by instructors. We further propose CLUE (Class Likeness Unified Evaluation), a unified metric set that assesses both global correctness and fine-grained design quality in class diagram generation. Using these benchmarks and metrics, we investigate five research questions: overall correctness, comparison with humans, model dimension analysis, task feature analysis, and bad case analysis. The results indicate that while LLMs achieve high syntactic accuracy, they exhibit substantial semantic deficiencies, particularly in method and relationship generation. Among the evaluated models, Qwen3-Coder-30B achieves the best overall performance, rivaling DeepSeek-R1 and GPT-4o, while Gemma3-4B-IT outperforms GPT-4o-Mini despite its smaller parameter scale. Although top-performing LLMs nearly match the average performance of undergraduates, they remain significantly below the level of the best human designers. Further analysis shows that parameter scale, code specialization, and instruction tuning strongly influence performance, whereas increased design complexity and lower requirement readability degrade it. Bad case analysis reveals common failure modes, including keyword misuse, missing classes or relationships, and omitted methods.

cs.SE↗

arXiv2025

BugRepro: Enhancing Android Bug Reproduction with Domain-Specific Knowledge Integration

Mobile application development is a fast-paced process where maintaining high-quality user experiences is crucial. Bug reproduction, a key aspect of maintaining app quality, often faces significant challenges. Specifically, when descriptions in bug reports are ambiguous or difficult to comprehend, current approaches fail to extract accurate information. Moreover, modern applications exhibit inherent complexity with multiple pages and diverse functionalities, making it challenging for existing methods to map the relevant information in bug reports to the corresponding UI elements that need to be manipulated. To address these challenges, we propose BugRepro, a novel technique that integrates domain-specific knowledge to enhance the accuracy and efficiency of bug reproduction. BugRepro adopts a Retrieval-Augmented Generation (RAG) approach. It retrieves similar bug reports along with their corresponding steps to reproduce (S2R) entities from an example-rich RAG document. In addition, BugRepro explores the graphical user interface (GUI) of the app and extracts transition graphs from the user interface to incorporate app-specific knowledge to guide large language models (LLMs) in their exploration process. Our experiments demonstrate that BugRepro significantly outperforms two state-of-the-art methods (ReCDroid and AdbGPT). For S2R entity extraction accuracy, it achieves a 7.57 to 28.89 percentage point increase over prior methods. For the bug reproduction success rate, the improvement reaches 74.55% and 152.63%. In reproduction efficiency, the gains are 0.72% and 76.68%.

cs.SE↗