arXiv subjects

Zhongxin Liu

Publications and source records attributed to Zhongxin Liu.

At least 55 records · Page 3Linked to original sources

arXiv2025

PEACE: Towards Efficient Project-Level Efficiency Optimization via Hybrid Code Editing

Large Language Models (LLMs) have demonstrated significant capability in code generation, but their potential in code efficiency optimization remains underexplored. Previous LLM-based code efficiency optimization approaches exclusively focus on function-level optimization and overlook interaction between functions, failing to generalize to real-world development scenarios. Code editing techniques show great potential for conducting project-level optimization, yet they face challenges associated with invalid edits and suboptimal internal functions. To address these gaps, we propose Peace, a novel hybrid framework for Project-level code Efficiency optimization through Automatic Code Editing, which also ensures the overall correctness and integrity of the project. Peace integrates three key phases: dependency-aware optimizing function sequence construction, valid associated edits identification, and efficiency optimization editing iteration. To rigorously evaluate the effectiveness of Peace, we construct PeacExec, the first benchmark comprising 146 real-world optimization tasks from 47 high-impact GitHub Python projects, along with highly qualified test cases and executable environments. Extensive experiments demonstrate Peace's superiority over the state-of-the-art baselines, achieving a 69.2% correctness rate (pass@1), +46.9% opt rate, and 0.840 speedup in execution efficiency. Notably, our Peace outperforms all baselines by significant margins, particularly in complex optimization tasks with multiple functions. Moreover, extensive experiments are also conducted to validate the contributions of each component in Peace, as well as the rationale and effectiveness of our hybrid framework design.

cs.SE↗

arXiv2025

Issue Localization via LLM-Driven Iterative Code Graph Searching

Issue solving aims to generate patches to fix reported issues in real-world code repositories according to issue descriptions. Issue localization forms the basis for accurate issue solving. Recently, LLM-based issue localization methods have demonstrated state-of-the-art performance. However, these methods either search from files mentioned in issue descriptions or in the whole repository and struggle to balance the breadth and depth of the search space to converge on the target efficiently. Moreover, they allow LLM to explore whole repositories freely, making it challenging to control the search direction to prevent the LLM from searching for incorrect targets. This paper introduces CoSIL, an LLM-driven, powerful function-level issue localization method without training or indexing. CoSIL employs a two-phase code graph search strategy. It first conducts broad exploration at the file level using dynamically constructed module call graphs, and then performs in-depth analysis at the function level by expanding the module call graph into a function call graph and executing iterative searches. To precisely control the search direction, CoSIL designs a pruner to filter unrelated directions and irrelevant contexts. To avoid incorrect interaction formats in long contexts, CoSIL introduces a reflection mechanism that uses additional independent queries in short contexts to enhance formatted abilities. Experiment results demonstrate that CoSIL achieves a Top-1 localization accuracy of 43.3\% and 44.6\% on SWE-bench Lite and SWE-bench Verified, respectively, with Qwen2.5-Coder-32B, average outperforming the state-of-the-art methods by 96.04\%. When CoSIL is integrated into an issue-solving method, Agentless, the issue resolution rate improves by 2.98\%--30.5\%.

cs.SE↗

arXiv2025

Unit Test Update through LLM-Driven Context Collection and Error-Type-Aware Refinement

Unit testing is critical for ensuring software quality and software system stability. The current practice of manually maintaining unit tests suffers from low efficiency and the risk of delayed or overlooked fixes. Therefore, an automated approach is required to instantly update unit tests, with the capability to both repair and enhance unit tests. However, existing automated test maintenance methods primarily focus on repairing broken tests, neglecting the scenario of enhancing existing tests to verify new functionality. Meanwhile, due to their reliance on rule-based context collection and the lack of verification mechanisms, existing approaches struggle to handle complex code changes and often produce test cases with low correctness. To address these challenges, we propose TESTUPDATER, a novel LLM based approach that enables automated just-in-time test updates in response to production code changes. TESTUPDATER first leverages the LLM to analyze code changes and identify relevant context, which it then extracts and filters. Then, through carefully designed prompts, TESTUPDATER guides the LLM step by step to handle various types of code changes and introduce new dependencies, enabling both test repair and enhancement. Finally, we introduce an error-type-aware iterative refinement mechanism that executes the LLM-updated tests and repairs failures, which significantly improves the overall correctness of test updates. Since existing test repair datasets lack scenarios of test enhancement, we further construct a new benchmark, UPDATES4J, with 195 real-world samples from 7 projects. Experimental results show that TESTUPDATER achieves a compilation pass rate of 94.4% and a test pass rate of 86.7%, outperforming the state-of-the-art method SYNTER by 15.9% and 20.0%, respectively. Furthermore, TESTUPDATER exhibits 12.9% higher branch coverage and 15.2% greater line coverage than SYNTER.

cs.SE↗

arXiv2025

Are "Solved Issues" in SWE-bench Really Solved Correctly? An Empirical Study

Automated issue solving aims to resolve real-world issues in software repositories. The most popular benchmarks for automated issue solving are SWE-bench and its human-filtered subset SWE-bench Verified. These benchmarks leverage testing to validate generated patches. However, because testing is rarely exhaustive, a patch may pass the tests but nevertheless fail to match the developers' expectations. Unfortunately, it is currently unclear to what extent evaluations performed with SWE-bench suffer from such plausible but incorrect patches. This paper presents an in-depth empirical study of the correctness of plausible patches generated by three state-of-the-art issue-solving tools evaluated on SWE-bench Verified. We extensively test and inspect generated patches, and compare them against human-written ground truth patches. The core of our methodology is a novel technique PatchDiff for differential patch testing, which automatically exposes behavioral discrepancies between two patches. Our findings reveal critical weaknesses in SWE-bench's patch validation mechanism, which causes 7.8% of all patches to count as correct while failing the developer-written test suite. Moreover, our novel automated technique reveals that even more (29.6%) plausible patches induce different behavior than the ground truth patches. These behavioral differences are often due to similar, but divergent implementations (46.8%) and due to generated patches that adapt more behavior than the ground truth patches (27.3%). Our manual inspection shows that 28.6% of behaviorally divergent patches are certainly incorrect. Combined, the different weaknesses lead to an inflation of reported resolution rates by 6.2 absolute percent points. Our findings are a call to arms for more robust and reliable evaluation of issue-solving tools. We envision our automated differential patch testing technique to be useful for this purpose.

cs.SE↗

arXiv2025

Enhancing Project-Specific Code Completion by Inferring Internal API Information

Project-specific code completion is a critical task that leverages context from a project to generate accurate code. State-of-the-art methods use retrieval-augmented generation (RAG) with large language models (LLMs) and project information for code completion. However, they often struggle to incorporate internal API information, which is crucial for accuracy, especially when APIs are not explicitly imported in the file. To address this, we propose a method to infer internal API information without relying on imports. Our method extends the representation of APIs by constructing usage examples and semantic descriptions, building a knowledge base for LLMs to generate relevant completions. We also introduce ProjBench, a benchmark that avoids leaked imports and consists of large-scale real-world projects. Experiments on ProjBench and CrossCodeEval show that our approach significantly outperforms existing methods, improving code exact match by 22.72% and identifier exact match by 18.31%. Additionally, integrating our method with existing baselines boosts code match by 47.80% and identifier match by 35.55%.

cs.SE↗

arXiv2025

MultiKernelBench: A Multi-Platform Benchmark for Kernel Generation

The automatic generation of deep learning (DL) kernels using large language models (LLMs) has emerged as a promising approach to reduce the manual effort and hardware-specific expertise required for writing high-performance operator implementations. However, existing benchmarks for evaluating LLMs in this domain suffer from limited hardware support, coarse-grained kernel categorization, and imbalanced task coverage. To address these limitations, we introduce MultiKernelBench, the first comprehensive, multi-platform benchmark for LLM-based DL kernel generation. MultiKernelBench spans 285 tasks across 14 well-defined kernel categories and supports three major hardware platforms: Nvidia GPUs, Huawei NPUs, and Google TPUs. To enable future extensibility, we design a modular backend abstraction layer that decouples platform-specific logic from the core benchmarking infrastructure, allowing easy integration of new hardware platforms. We further propose a simple yet effective category-aware one-shot prompting method that improves generation quality by providing in-category exemplars. Through systematic evaluations of seven state-of-the-art LLMs, we reveal significant variation in task difficulty, poor generalization to platforms with less training exposure, and the effectiveness of targeted prompting strategies. MultiKernelBench is publicly available at https://github.com/wzzll123/MultiKernelBench.

cs.DC↗

arXiv2025

CodeReasoner: Enhancing the Code Reasoning Ability with Reinforcement Learning

Code reasoning is a fundamental capability for large language models (LLMs) in the code domain. It involves understanding and predicting a program's execution behavior, such as determining the output for a given input or whether a specific statement will be executed. This capability is essential for downstream tasks like debugging, code generation, and program repair. Prior approaches mainly rely on supervised fine-tuning to improve performance in code reasoning tasks. However, they often show limited gains and fail to generalize across diverse scenarios. We argue this is due to two core issues: the low quality of training data and the limitations of supervised fine-tuning, which struggles to teach general reasoning skills. To address these challenges, we propose CodeReasoner, a framework that spans both dataset construction and a two-stage training process. First, we introduce a method to construct datasets that focus on the core execution logic of Python programs. Next, we apply instruction tuning to inject execution-specific knowledge distilled from a powerful teacher model. We then enhance reasoning and generalization through GRPO reinforcement learning on top of the fine-tuned model. Experiments on three widely-used code reasoning benchmarks show that CodeReasoner improves performance by 27.1% to 40.2% over prior methods using a 7B model. Notably, the 7B model matches GPT-4o on key tasks like input/output and coverage prediction. When scaled to 14B, CodeReasoner outperforms GPT-4o across all benchmarks. Ablation studies confirm the effectiveness of each training stage and highlight the importance of reasoning chains.

cs.SE↗

arXiv2025

Parallel Scaling Law for Language Models

It is commonly believed that scaling language models should commit a significant space or time cost, by increasing the parameters (parameter scaling) or output tokens (inference-time scaling). We introduce the third and more inference-efficient scaling paradigm: increasing the model's parallel computation during both training and inference time. We apply $P$ diverse and learnable transformations to the input, execute forward passes of the model in parallel, and dynamically aggregate the $P$ outputs. This method, namely parallel scaling (ParScale), scales parallel computation by reusing existing parameters and can be applied to any model structure, optimization procedure, data, or task. We theoretically propose a new scaling law and validate it through large-scale pre-training, which shows that a model with $P$ parallel streams is similar to scaling the parameters by $O(\log P)$ while showing superior inference efficiency. For example, ParScale can use up to 22$\times$ less memory increase and 6$\times$ less latency increase compared to parameter scaling that achieves the same performance improvement. It can also recycle an off-the-shelf pre-trained model into a parallelly scaled one by post-training on a small amount of tokens, further reducing the training budget. The new scaling law we discovered potentially facilitates the deployment of more powerful models in low-resource scenarios, and provides an alternative perspective for the role of computation in machine learning.

cs.LG↗

arXiv2025

Zero-Shot Cross-Domain Code Search without Fine-Tuning

Code search aims to retrieve semantically relevant code snippets for natural language queries. While pre-trained language models (PLMs) have shown remarkable performance in this task, they struggle in cross-domain scenarios, often requiring costly fine-tuning or facing performance drops in zero-shot settings. RAPID, which generates synthetic data for model fine-tuning, is currently the only effective method for zero-shot cross-domain code search. Despite its effectiveness, RAPID demands substantial computational resources for fine-tuning and needs to maintain specialized models for each domain, underscoring the need for a zero-shot, fine-tuning-free approach for cross-domain code search. The key to tackling zero-shot cross-domain code search lies in bridging the gaps among domains. In this work, we propose to break the query-code matching process of code search into two simpler tasks: query-comment matching and code-code matching. Our empirical study reveals the strong complementarity among the three matching schemas in zero-shot cross-domain settings, i.e., query-code, query-comment, and code-code matching. Based on the findings, we propose CodeBridge, a zero-shot, fine-tuning-free approach for cross-domain code search. Specifically, CodeBridge uses Large Language Models (LLMs) to generate comments and pseudo-code, then combines query-code, query-comment, and code-code matching via PLM-based similarity scoring and sampling-based fusion. Experimental results show that our approach outperforms the state-of-the-art PLM-based code search approaches, i.e., CoCoSoDa and UniXcoder, by an average of 21.4% and 24.9% in MRR, respectively, across three datasets. Our approach also yields results that are better than or comparable to those of the zero-shot cross-domain code search approach RAPID, which requires costly fine-tuning.

cs.SE↗

arXiv2025

AdaCoder: An Adaptive Planning and Multi-Agent Framework for Function-Level Code Generation

Recently, researchers have proposed many multi-agent frameworks for function-level code generation, which aim to improve software development productivity by automatically generating function-level source code based on task descriptions. A typical multi-agent framework consists of Large Language Model (LLM)-based agents that are responsible for task planning, code generation, testing, debugging, etc. Studies have shown that existing multi-agent code generation frameworks perform well on ChatGPT. However, their generalizability across other foundation LLMs remains unexplored systematically. In this paper, we report an empirical study on the generalizability of four state-of-the-art multi-agent code generation frameworks across six open-source LLMs with varying parameter sizes, architectures, and performance levels. Our study reveals the unstable generalizability of existing frameworks on diverse foundation LLMs. Based on the findings obtained from the empirical study, we propose AdaCoder, a novel adaptive planning, multi-agent framework for function-level code generation. AdaCoder has two phases. Phase-1 is an initial code generation step without planning, which uses an LLM-based coding agent and a script-based testing agent to unleash LLM's native power, identify cases beyond LLM's power, and determine the errors hindering execution. Phase-2 adds a rule-based debugging agent and an LLM-based planning agent for iterative code generation with planning. Our evaluation shows that AdaCoder achieves higher generalizability on diverse LLMs. Compared to the best baseline MapCoder, AdaCoder is on average 27.69% higher in Pass@1, 16 times faster in inference, and 12 times lower in token consumption.

cs.SE↗

arXiv2025

LLM4SZZ: Enhancing SZZ Algorithm with Context-Enhanced Assessment on Large Language Models

The SZZ algorithm is the dominant technique for identifying bug-inducing commits and serves as a foundation for many software engineering studies, such as bug prediction and static code analysis. Researchers have proposed many variants to enhance the SZZ algorithm's performance since its introduction. The majority of them rely on static techniques or heuristic assumptions, making them easy to implement, but their performance improvements are often limited. Recently, a deep learning-based SZZ algorithm has been introduced to enhance the original SZZ algorithm. However, it requires complex preprocessing and is restricted to a single programming language. Additionally, while it enhances precision, it sacrifices recall. Furthermore, most of variants overlook crucial information, such as commit messages and patch context, and are limited to bug-fixing commits involving deleted lines. The emergence of large language models (LLMs) offers an opportunity to address these drawbacks. In this study, we investigate the strengths and limitations of LLMs and propose LLM4SZZ, which employs two approaches (i.e., rank-based identification and context-enhanced identification) to handle different types of bug-fixing commits. We determine which approach to adopt based on the LLM's ability to comprehend the bug and identify whether the bug is present in a commit. The context-enhanced identification provides the LLM with more context and requires it to find the bug-inducing commit among a set of candidate commits. In rank-based identification, we ask the LLM to select buggy statements from the bug-fixing commit and rank them based on their relevance to the root cause. Experimental results show that LLM4SZZ outperforms all baselines across three datasets, improving F1-score by 6.9% to 16.0% without significantly sacrificing recall.

cs.SE↗

arXiv2025

Propensity Formation-Containment Control of Fully Heterogeneous Multi-Agent Systems via Online Data-Driven Learning

This paper introduces an online data-driven learning scheme designed to address a novel problem in propensity formation and containment control for fully heterogeneous multi-agent systems. Unlike traditional approaches that rely on the eigenvalues of the Laplacian matrix, this problem considers the determination of follower positions based on propensity factors released by leaders. To address the challenge of incomplete utilization of leader information in existing multi-leader control methods, the concept of an influential transit formation leader (ITFL) is introduced. An adaptive observer is developed for the agents, including the ITFL, to estimate the state of the tracking leader or the leader's formation. Building on these observations, a model-based control protocol is proposed, elucidating the relationship between the regulation equations and control gains, ensuring the asymptotic convergence of the agent's state. To eliminate the necessity for model information throughout the control process, a new online data-driven learning algorithm is devised for the control protocol. Finally, numerical simulation results are given to verify the effectiveness of the proposed method.

eess.SY↗

arXiv2025

FlexFL: Flexible and Effective Fault Localization with Open-Source Large Language Models

Due to the impressive code comprehension ability of Large Language Models (LLMs), a few studies have proposed to leverage LLMs to locate bugs, i.e., LLM-based FL, and demonstrated promising performance. However, first, these methods are limited in flexibility. They rely on bug-triggering test cases to perform FL and cannot make use of other available bug-related information, e.g., bug reports. Second, they are built upon proprietary LLMs, which are, although powerful, confronted with risks in data privacy. To address these limitations, we propose a novel LLM-based FL framework named FlexFL, which can flexibly leverage different types of bug-related information and effectively work with open-source LLMs. FlexFL is composed of two stages. In the first stage, FlexFL reduces the search space of buggy code using state-of-the-art FL techniques of different families and provides a candidate list of bug-related methods. In the second stage, FlexFL leverages LLMs to delve deeper to double-check the code snippets of methods suggested by the first stage and refine fault localization results. In each stage, FlexFL constructs agents based on open-source LLMs, which share the same pipeline that does not postulate any type of bug-related information and can interact with function calls without the out-of-the-box capability. Extensive experimental results on Defects4J demonstrate that FlexFL outperforms the baselines and can work with different open-source LLMs. Specifically, FlexFL with a lightweight open-source LLM Llama3-8B can locate 42 and 63 more bugs than two state-of-the-art LLM-based FL approaches AutoFL and AgentFL that both use GPT-3.5.

cs.SE↗

arXiv2024

Self-Explained Keywords Empower Large Language Models for Code Generation

Large language models (LLMs) have achieved impressive performance in code generation. However, due to the long-tail distribution of LLMs' training data, low-frequency terms are typically underrepresented in the training process. Consequently, LLMs often misunderstand or overlook problem-specific, low-frequency keywords during code generation, compromising the accuracy of the generated code. To address this, we propose a novel technique named SEK(\textbf{S}elf-\textbf{E}xplained \textbf{K}eywords), which empowers an LLM for better code generation by extracting and explaining the key terms in the problem description with the LLM itself and ranking them based on frequency. Comprehensive experiments across three benchmarks, i.e., HumanEval(+), MBPP(+), and APPS, with five representative LLMs, show that SEK can significantly improve LLMs in code generation, yielding substantial and consistent gains. For instance, SEK improves the Pass@1 of DeepSeek-Coder-V2-Instruct from 85.4\% to 93.3\% on the Humaneval benchmark. Further analysis confirms that SEK enables the LLMs to shift their attention from low-frequency keywords to their corresponding high-frequency counterparts.

cs.CL↗

arXiv2024

Instructive Code Retriever: Learn from Large Language Model's Feedback for Code Intelligence Tasks

Recent studies proposed to leverage large language models (LLMs) with In-Context Learning (ICL) to handle code intelligence tasks without fine-tuning. ICL employs task instructions and a set of examples as demonstrations to guide the model in generating accurate answers without updating its parameters. While ICL has proven effective for code intelligence tasks, its performance heavily relies on the selected examples. Previous work has achieved some success in using BM25 to retrieve examples for code intelligence tasks. However, existing approaches lack the ability to understand the semantic and structural information of queries, resulting in less helpful demonstrations. Moreover, they do not adapt well to the complex and dynamic nature of user queries in diverse domains. In this paper, we introduce a novel approach named Instructive Code Retriever (ICR), which is designed to retrieve examples that enhance model inference across various code intelligence tasks and datasets. We enable ICR to learn the semantic and structural information of the corpus by a tree-based loss function. To better understand the correlation between queries and examples, we incorporate the feedback from LLMs to guide the training of the retriever. Experimental results demonstrate that our retriever significantly outperforms state-of-the-art approaches. We evaluate our model's effectiveness on various tasks, i.e., code summarization, program synthesis, and bug fixing. Compared to previous state-of-the-art algorithms, our method achieved improvements of 50.0% and 90.0% in terms of BLEU-4 for two code summarization datasets, 74.6% CodeBLEU on program synthesis dataset, and increases of 3.6 and 3.2 BLEU-4 on two bug fixing datasets.

cs.SE↗

arXiv2024

B4: Towards Optimal Assessment of Plausible Code Solutions with Plausible Tests

Selecting the best code solution from multiple generated ones is an essential task in code generation, which can be achieved by using some reliable validators (e.g., developer-written test cases) for assistance. Since reliable test cases are not always available and can be expensive to build in practice, researchers propose to automatically generate test cases to assess code solutions. However, when both code solutions and test cases are plausible and not reliable, selecting the best solution becomes challenging. Although some heuristic strategies have been proposed to tackle this problem, they lack a strong theoretical guarantee and it is still an open question whether an optimal selection strategy exists. Our work contributes in two ways. First, we show that within a Bayesian framework, the optimal selection strategy can be defined based on the posterior probability of the observed passing states between solutions and tests. The problem of identifying the best solution is then framed as an integer programming problem. Second, we propose an efficient approach for approximating this optimal (yet uncomputable) strategy, where the approximation error is bounded by the correctness of prior knowledge. We then incorporate effective prior knowledge to tailor code generation tasks. Both theoretical and empirical studies confirm that existing heuristics are limited in selecting the best solutions with plausible test cases. Our proposed approximated optimal strategy B4 significantly surpasses existing heuristics in selecting code solutions generated by large language models (LLMs) with LLM-generated tests, achieving a relative performance improvement by up to 50% over the strongest heuristic and 246% over the random selection in the most challenging scenarios. Our code is publicly available at https://github.com/ZJU-CTAG/B4.

cs.SE↗

arXiv2024

Improving Retrieval-Augmented Code Comment Generation by Retrieving for Generation

Code comment generation aims to generate high-quality comments from source code automatically and has been studied for years. Recent studies proposed to integrate information retrieval techniques with neural generation models to tackle this problem, i.e., Retrieval-Augmented Comment Generation (RACG) approaches, and achieved state-of-the-art results. However, the retrievers in previous work are built independently of their generators. This results in that the retrieved exemplars are not necessarily the most useful ones for generating comments, limiting the performance of existing approaches. To address this limitation, we propose a novel training strategy to enable the retriever to learn from the feedback of the generator and retrieve exemplars for generation. Specifically, during training, we use the retriever to retrieve the top-k exemplars and calculate their retrieval scores, and use the generator to calculate a generation loss for the sample based on each exemplar. By aligning high-score exemplars retrieved by the retriever with low-loss exemplars observed by the generator, the retriever can learn to retrieve exemplars that can best improve the quality of the generated comments. Based on this strategy, we propose a novel RACG approach named JOINTCOM and evaluate it on two real-world datasets, JCSD and PCSD. The experimental results demonstrate that our approach surpasses the state-of-the-art baselines by 7.3% to 30.0% in terms of five metrics on the two datasets. We also conduct a human evaluation to compare JOINTCOM with the best-performing baselines. The results indicate that JOINTCOM outperforms the baselines, producing comments that are more natural, informative, and useful.

cs.SE↗

arXiv2024

Exploring the Capabilities of LLMs for Code Change Related Tasks

Developers deal with code-change-related tasks daily, e.g., reviewing code. Pre-trained code and code-change-oriented models have been adapted to help developers with such tasks. Recently, large language models (LLMs) have shown their effectiveness in code-related tasks. However, existing LLMs for code focus on general code syntax and semantics rather than the differences between two code versions. Thus, it is an open question how LLMs perform on code-change-related tasks. To answer this question, we conduct an empirical study using \textgreater 1B parameters LLMs on three code-change-related tasks, i.e., code review generation, commit message generation, and just-in-time comment update, with in-context learning (ICL) and parameter-efficient fine-tuning (PEFT, including LoRA and prefix-tuning). We observe that the performance of LLMs is poor without examples and generally improves with examples, but more examples do not always lead to better performance. LLMs tuned with LoRA have comparable performance to the state-of-the-art small pre-trained models. Larger models are not always better, but \textsc{Llama~2} and \textsc{Code~Llama} families are always the best. The best LLMs outperform small pre-trained models on the code changes that only modify comments and perform comparably on other code changes. We suggest future work should focus more on guiding LLMs to learn the knowledge specific to the changes related to code rather than comments for code-change-related tasks.

cs.SE↗