arXiv subjects

Zeyu Sun

Publications and source records attributed to Zeyu Sun.

At least 37 records · Page 2Linked to original sources

arXiv2025

Condor: A Code Discriminator Integrating General Semantics with Code Details

LLMs demonstrate significant potential across various software engineering tasks. However, they still face challenges in generating correct code on the first attempt when addressing complex requirements. Introducing a discriminator to select reliable outputs from multiple generated results is an effective way to enhance their reliability and stability. Currently, these discriminators fall into two categories: execution-based discriminators and non-execution-based discriminators. Execution-based discriminators face flexibility challenges due to difficulties in obtaining test cases and security concerns, while non-execution-based discriminators, although more flexible, struggle to capture subtle differences in code details. To maintain flexibility while improving the model's ability to capture fine-grained code details, this paper proposes Condor. We first design contrastive learning to optimize the code representations of the base model, enabling it to reflect differences in code details. Then, we leverage intermediate data from the code modification process to further enrich the discriminator's training data, enhancing its ability to discern code details. Experimental results indicate that on the subtle code difference dataset (i.e., CodeNanoFix), Condor significantly outperforms other discriminators in discriminative performance: Condor (1.3B) improves the discriminative F1 score of DeepSeek-Coder (1.3B) from 67% to 73%. In discriminating LLM-generated outputs, Condor (1.3B) and Condor (110M) raise the Pass@1 score of Meta-Llama-3.1-Instruct (70B) on the CodeNanoFix dataset from 52.64% to 62.63% and 59.64%, respectively. Moreover, Condor demonstrates strong generalization capabilities on the APPS, MBPP, and LiveCodeBench datasets. For example, Condor (1.3B) improves the Pass@1 of Meta-Llama-3.1-Instruct (70B) on the APPS dataset by 147.05%.

cs.SE↗

arXiv2025

Directional Diffusion-Style Code Editing Pre-training

Code pre-trained models have shown promising effectiveness in various software engineering tasks. Among these tasks, many tasks are related to software evolution and/or code editing. However, existing code pre-trained models often overlook the real-world code editing data and the evolutionary nature of the editing process. In this paper, to simulate the step-by-step code editing process of human developers, we propose DivoT5, a pre-trained model based on directional diffusion at the data level. In DivoT5, we adopt two categories of pre-training tasks. The first category is mask and denoising tasks augmented with a diffusion direction representing code evolution. That is, we first apply a noising process to the code snippets before evolution, and then ask the pre-training process to restore the snippets with noise into the code snippets after evolution. The second category is tasks aiming to reinforce the evolutionary direction. That is, we first generate various intermediate versions for each pair of snippets before and after evolution, and then ask the pre-training process to transform the intermediate versions into the snippet after evolution for each pair. We evaluate DivoT5 for two code-editing scenarios and one non-editing scenario using five downstream tasks. Given each downstream task, we fine-tune the pre-trained DivoT5 to evaluate its effectiveness. Our experimental results show that DivoT5 achieves state-of-the-art (SOTA) performance on most tasks in comparison to models of the same scale (220M), large scale (770M) models in fine-tuning, and billion-scale (6.7B, 8B, ChatGPT) models in few-shot settings. For one code-editing task (i.e., automated code review), DivoT5 pre-trained on top of CodeT5-small (60M) can even outperform CodeT5-base (220M) and other pre-trained models with 220M parameters except for DivoT5 pre-trained on top of CodeT5-base (220M).

cs.SE↗

arXiv2025

Grammar-Based Code Representation: Is It a Worthy Pursuit for LLMs?

Grammar serves as a cornerstone in programming languages and software engineering, providing frameworks to define the syntactic space and program structure. Existing research demonstrates the effectiveness of grammar-based code representations in small-scale models, showing their ability to reduce syntax errors and enhance performance. However, as language models scale to the billion level or beyond, syntax-level errors become rare, making it unclear whether grammar information still provides performance benefits. To explore this, we develop a series of billion-scale GrammarCoder models, incorporating grammar rules in the code generation process. Experiments on HumanEval (+) and MBPP (+) demonstrate a notable improvement in code generation accuracy. Further analysis shows that grammar-based representations enhance LLMs' ability to discern subtle code differences, reducing semantic errors caused by minor variations. These findings suggest that grammar-based code representations remain valuable even in billion-scale models, not only by maintaining syntax correctness but also by improving semantic differentiation.

cs.PL↗

arXiv2025

DSCodeBench: A Realistic Benchmark for Data Science Code Generation

We introduce DSCodeBench, a new benchmark designed to evaluate large language models (LLMs) on complicated and realistic data science code generation tasks. DSCodeBench consists of 1,000 carefully constructed problems sourced from realistic problems from GitHub across ten widely used Python data science libraries. DSCodeBench offers a more challenging and representative testbed, more complex code solutions, more comprehensive data science libraries, clearer and better structured problem descriptions, and stronger test suites. To construct the DSCodeBench, we develop a robust pipeline that combines task scope selection, code construction, test case generation, and problem description synthesis. The process is paired with rigorous manual editing to ensure alignment and enhance the reliability of the evaluation. Experimental result shows that DSCodeBench exhibits robust scaling behavior, where larger models systematically outperform smaller ones, validating its ability to distinguish model capabilities. The best LLM we test, GPT-4o, has a pass@1 of 0.392, indicating that LLMs still have a large room to improve for realistic data science code generation tasks. We believe DSCodeBench will serve as a rigorous and trustworthy foundation for advancing LLM-based data science programming.

cs.SE↗

arXiv2025

Interleaved Learning and Exploration: A Self-Adaptive Fuzz Testing Framework for MLIR

MLIR (Multi-Level Intermediate Representation) has rapidly become a foundational technology for modern compiler frameworks, enabling extensibility across diverse domains. However, ensuring the correctness and robustness of MLIR itself remains challenging. Existing fuzzing approaches-based on manually crafted templates or rule-based mutations-struggle to generate sufficiently diverse and semantically valid test cases, making it difficult to expose subtle or deep-seated bugs within MLIR's complex and evolving code space. In this paper, we present FLEX, a novel self-adaptive fuzzing framework for MLIR. FLEX leverages neural networks for program generation, a perturbed sampling strategy to encourage diversity, and a feedback-driven augmentation loop that iteratively improves its model using both crashing and non-crashing test cases. Starting from a limited seed corpus, FLEX progressively learns valid syntax and semantics and autonomously produces high-quality test inputs. We evaluate FLEX on the upstream MLIR compiler against four state-of-the-art fuzzers. In a 30-day campaign, FLEX discovers 80 previously unknown bugs-including multiple new root causes and parser bugs-while in 24-hour fixed-revision comparisons, it detects 53 bugs (over 3.5x as many as the best baseline) and achieves 28.2% code coverage, outperforming the next-best tool by 42%. Ablation studies further confirm the critical role of both perturbed generation and diversity augmentation in FLEX's effectiveness.

cs.SE↗

arXiv2025

GramTrans: A Better Code Representation Approach in Code Generation

Code generation has shown great promise in assisting software development. A fundamental yet underexplored question is how the choice of code representation affects model performance. While existing studies employ various representations, such as treating code as plain text, grammar rule sequences, or syntax tree sequences, they lack a principled understanding of the relationship between parsing difficulty and model effectiveness. This paper proposes a conjecture: the easier a representation is to parse, the better performance the model achieves. We formalize this idea using grammar classes, where representations in simpler classes (e.g., LL(1)) are easier to parse. Through a controlled experiment on a Python-based DSL, we show that parsing difficulty strongly correlates with model performance. Motivated by this finding, we present GramTrans, a general approach that automatically transforms a context-free language into a representation within the LL(1) class. GramTrans introduces a novel hierarchical conflict elimination algorithm, enabling a flexible trade-off between syntactic simplicity and token efficiency. We evaluate GramTrans on both Python and Java using three code generation models: StarCoder 1B, DeepSeek-Coder 1.3B, and Qwen2.5 1.5B. Across multiple benchmarks, GramTrans consistently delivers significant improvements over baseline representations. Furthermore, our analysis of existing representations reconfirms the strong alignment between parsing difficulty and model performance, providing additional support for the conjecture.

cs.SE↗

arXiv2025

Feedback-Free Resource Scheduling for Flexible Multi-BS Cooperation in FD-RAN

Flexible cooperation among base stations (BSs) is critical to improve resource utilization efficiency and meet personalized user demands. However, its practical implementation is hindered by the current radio access network (RAN), which relies on the coupling of uplink and downlink transmissions and channel state information feedback with inherent issues such as overheads and delays. To overcome these limitations, we consider the fully-decoupled RAN (FD-RAN), in which uplink and downlink functionalities are independent, and feedback-free MIMO transmission is adopted at the physical layer. To further deliver flexible cooperation in FD-RAN, we study feedback-free downlink multi-BS resource scheduling under the practical scheduling process. The problem is considered based on network load conditions. In heavy-load states where it is impossible for all user demands to be met, an optimal greedy algorithm is proposed, maximizing the weighted sum of user demand satisfaction rates. In light-load states where at least one solution exists to satisfy all user demands, an optimal two-stage resource allocation algorithm is designed to further minimize network energy consumption by leveraging the flexibility of cooperation. Extensive simulations validate the superiority of proposed algorithms in performance and running time, and highlight the potential for realizing flexible cooperation in practice.

eess.SP↗

arXiv2025

LLMs are Bug Replicators: An Empirical Study on LLMs' Capability in Completing Bug-prone Code

Large Language Models (LLMs) have demonstrated remarkable performance in code completion. However, the training data used to develop these models often contain a significant amount of buggy code. Yet, it remains unclear to what extent these buggy instances influence LLMs' performance when tackling bug-prone code completion tasks. To fill this gap, this paper presents the first empirical study evaluating the performance of LLMs in completing bug-prone code. Through extensive experiments on 7 LLMs and the Defects4J dataset, we analyze LLMs' accuracy, robustness, and limitations in this challenging context. Our experimental results show that completing bug-prone code is significantly more challenging for LLMs than completing normal code. Notably, in bug-prone tasks, the likelihood of LLMs generating correct code is nearly the same as generating buggy code, and it is substantially lower than in normal code completion tasks (e.g., 12.27% vs. 29.85% for GPT-4). To our surprise, 44.44% of the bugs LLMs make are completely identical to the pre-fix version, indicating that LLMs have been seriously biased by historical bugs when completing code. Additionally, we investigate the effectiveness of existing post-processing techniques and find that while they can improve consistency, they do not significantly reduce error rates in bug-prone code scenarios. Our research highlights the limitations of current LLMs in handling bug-prone code and underscores the need for improved models and post-processing strategies to enhance code completion accuracy in real-world development environments.

cs.SE↗

arXiv2025

Prompt Alchemy: Automatic Prompt Refinement for Enhancing Code Generation

Code generation has emerged as a key task to automate software development by converting high-level descriptions into executable code. Large language models (LLMs) excel at this but depend heavily on input prompt quality.Manual prompt engineering can be time-consuming and inconsistent, limiting LLM effectiveness. This paper introduces Prochemy, an innovative method for automatically refining prompts to boost code generation. Prochemy overcomes manual prompt limitations by automating optimization, ensuring consistency during inference, and supporting multi-agent systems.It iteratively refines prompts based on model performance, using an optimized final prompt for improved consistency across tasks. We tested Prochemy on natural language-based code generation and translation tasks using three LLM series. Results indicate Prochemy enhances existing methods, improving performance by 5.0% for GPT-3.5-Turbo and 1.9% for GPT-4o over zero-shot baselines on HumanEval. In state-of-the-art LDB, Prochemy + LDB surpasses standalone methods by 1.2-1.8%. For code translation, Prochemy boosts GPT-4o's Java-to-Python (AVATAR) performance from 74.5 to 84.1 (+12.9%) and Python-to-Java from 66.8 to 78.2 (+17.1%). Moreover, Prochemy maintains strong performance when integrated with the o1-mini model, validating its efficacy in code tasks. Designed as plug-and-play, Prochemy optimizes prompts with minimal human input, bridging the gap between simple prompts and complex frameworks.

cs.SE↗

arXiv2025

Knowledge-Enhanced Program Repair for Data Science Code

This paper introduces DSrepair, a knowledge-enhanced program repair method designed to repair the buggy code generated by LLMs in the data science domain. DSrepair uses knowledge graph based RAG for API knowledge retrieval as well as bug knowledge enrichment to construct repair prompts for LLMs. Specifically, to enable knowledge graph based API retrieval, we construct DS-KG (Data Science Knowledge Graph) for widely used data science libraries. For bug knowledge enrichment, we employ an abstract syntax tree (AST) to localize errors at the AST node level. DSrepair's effectiveness is evaluated against five state-of-the-art LLM-based repair baselines using four advanced LLMs on the DS-1000 dataset. The results show that DSrepair surpasses all five baselines. Specifically, when compared to the second-best baseline, DSrepair demonstrates significant improvements, fixing 44.4%, 14.2%, 20.6%, and 32.1% more buggy code snippets for each of the four evaluated LLMs, respectively. Additionally, it achieves greater efficiency, reducing the number of tokens required per code task by 17.49%, 34.24%, 24.71%, and 17.59%, respectively.

cs.SE↗

arXiv2025

Automatically Learning a Precise Measurement for Fault Diagnosis Capability of Test Cases

Prevalent Fault Localization (FL) techniques rely on tests to localize buggy program elements. Tests could be treated as fuel to further boost FL by providing more debugging information. Therefore, it is highly valuable to measure the Fault Diagnosis Capability (FDC) of a test for diagnosing faults, so as to select or generate tests to better help FL. To this end, researchers have proposed many FDC metrics, which serve as the selection criterion in FL-oriented test selection or the fitness function in FL-oriented test generation. Existing FDC metrics can be classified into result-agnostic and result-aware metrics depending on whether they take test results (i.e., passing or failing) as input. Although result-aware metrics perform better in test selection, they have restricted applications due to the input of test results, e.g., they cannot be applied to guide test generation. Moreover, all the existing FDC metrics are designed based on some predefined heuristics and have achieved limited FL performance due to their inaccuracy. To address these issues, in this paper, we reconsider result-agnostic metrics, and propose a novel result-agnostic metric RLFDC which predicts FDC values of tests through reinforcement learning. In particular, we treat FL results as reward signals, and train an FDC prediction model with the direct FL feedback to automatically learn a more accurate measurement rather than design one based on predefined heuristics. Finally, we evaluate the proposed RLFDC on Defects4J by applying the studied metrics to test selection and generation. According to the experimental results, the proposed RLFDC outperforms all the result-agnostic metrics in both test selection and generation.

cs.SE↗

arXiv2024

Do Advanced Language Models Eliminate the Need for Prompt Engineering in Software Engineering?

Large Language Models (LLMs) have significantly advanced software engineering (SE) tasks, with prompt engineering techniques enhancing their performance in code-related areas. However, the rapid development of foundational LLMs such as the non-reasoning model GPT-4o and the reasoning model o1 raises questions about the continued effectiveness of these prompt engineering techniques. This paper presents an extensive empirical study that reevaluates various prompt engineering techniques within the context of these advanced LLMs. Focusing on three representative SE tasks, i.e., code generation, code translation, and code summarization, we assess whether prompt engineering techniques still yield improvements with advanced models, the actual effectiveness of reasoning models compared to non-reasoning models, and whether the benefits of using these advanced models justify their increased costs. Our findings reveal that prompt engineering techniques developed for earlier LLMs may provide diminished benefits or even hinder performance when applied to advanced models. In reasoning LLMs, the ability of sophisticated built-in reasoning reduces the impact of complex prompts, sometimes making simple zero-shot prompting more effective. Furthermore, while reasoning models outperform non-reasoning models in tasks requiring complex reasoning, they offer minimal advantages in tasks that do not need reasoning and may incur unnecessary costs. Based on our study, we provide practical guidance for practitioners on selecting appropriate prompt engineering techniques and foundational LLMs, considering factors such as task requirements, operational costs, and environmental impact. Our work contributes to a deeper understanding of effectively harnessing advanced LLMs in SE tasks, informing future research and application development.

cs.SE↗

arXiv2024

Yuan 2.0-M32: Mixture of Experts with Attention Router

Yuan 2.0-M32, with a similar base architecture as Yuan-2.0 2B, uses a mixture-of-experts architecture with 32 experts of which 2 experts are active. A new router network, Attention Router, is proposed and adopted for a more efficient selection of experts, which improves the accuracy compared to the model with classical router network. Yuan 2.0-M32 is trained with 2000B tokens from scratch, and the training computation consumption is only 9.25% of a dense model at the same parameter scale. Yuan 2.0-M32 demonstrates competitive capability on coding, math, and various domains of expertise, with only 3.7B active parameters of 40B in total, and 7.4 GFlops forward computation per token, both of which are only 1/19 of Llama3-70B. Yuan 2.0-M32 surpass Llama3-70B on MATH and ARC-Challenge benchmark, with accuracy of 55.89 and 95.8 respectively. The models and source codes of Yuan 2.0-M32 are released at Github1.

cs.AI↗

arXiv2024

Improving Smart Contract Security with Contrastive Learning-based Vulnerability Detection

Currently, smart contract vulnerabilities (SCVs) have emerged as a major factor threatening the transaction security of blockchain. Existing state-of-the-art methods rely on deep learning to mitigate this threat. They treat each input contract as an independent entity and feed it into a deep learning model to learn vulnerability patterns by fitting vulnerability labels. It is a pity that they disregard the correlation between contracts, failing to consider the commonalities between contracts of the same type and the differences among contracts of different types. As a result, the performance of these methods falls short of the desired level. To tackle this problem, we propose a novel Contrastive Learning Enhanced Automated Recognition Approach for Smart Contract Vulnerabilities, named Clear. In particular, Clear employs a contrastive learning (CL) model to capture the fine-grained correlation information among contracts and generates correlation labels based on the relationships between contracts to guide the training process of the CL model. Finally, it combines the correlation and the semantic information of the contract to detect SCVs. Through an empirical evaluation of a large-scale real-world dataset of over 40K smart contracts and compare 13 state-of-the-art baseline methods. We show that Clear achieves (1) optimal performance over all baseline methods; (2) 9.73%-39.99% higher F1-score than existing deep learning methods.

cs.CR↗

arXiv2024

Channel-Feedback-Free Transmission for Downlink FD-RAN: A Radio Map based Complex-valued Precoding Network Approach

As the demand for high-quality services proliferates, an innovative network architecture, the fully-decoupled RAN (FD-RAN), has emerged for more flexible spectrum resource utilization and lower network costs. However, with the decoupling of uplink base stations and downlink base stations in FD-RAN, the traditional transmission mechanism, which relies on real-time channel feedback, is not suitable as the receiver is not able to feedback accurate and timely channel state information to the transmitter. This paper proposes a novel transmission scheme without relying on physical layer channel feedback. Specifically, we design a radio map based complex-valued precoding network~(RMCPNet) model, which outputs the base station precoding based on user location. RMCPNet comprises multiple subnets, with each subnet responsible for extracting unique modal features from diverse input modalities. Furthermore, the multi-modal embeddings derived from these distinct subnets are integrated within the information fusion layer, culminating in a unified representation. We also develop a specific RMCPNet training algorithm that employs the negative spectral efficiency as the loss function. We evaluate the performance of the proposed scheme on the public DeepMIMO dataset and show that RMCPNet can achieve 16\% and 76\% performance improvements over the conventional real-valued neural network and statistical codebook approach, respectively.

cs.IT↗

arXiv2024

A Large-Scale Empirical Study on Improving the Fairness of Image Classification Models

Fairness has been a critical issue that affects the adoption of deep learning models in real practice. To improve model fairness, many existing methods have been proposed and evaluated to be effective in their own contexts. However, there is still no systematic evaluation among them for a comprehensive comparison under the same context, which makes it hard to understand the performance distinction among them, hindering the research progress and practical adoption of them. To fill this gap, this paper endeavours to conduct the first large-scale empirical study to comprehensively compare the performance of existing state-of-the-art fairness improving techniques. Specifically, we target the widely-used application scenario of image classification, and utilized three different datasets and five commonly-used performance metrics to assess in total 13 methods from diverse categories. Our findings reveal substantial variations in the performance of each method across different datasets and sensitive attributes, indicating over-fitting on specific datasets by many existing methods. Furthermore, different fairness evaluation metrics, due to their distinct focuses, yield significantly different assessment results. Overall, we observe that pre-processing methods and in-processing methods outperform post-processing methods, with pre-processing methods exhibiting the best performance. Our empirical study offers comprehensive recommendations for enhancing fairness in deep learning models. We approach the problem from multiple dimensions, aiming to provide a uniform evaluation platform and inspire researchers to explore more effective fairness solutions via a set of implications.

cs.LG↗

arXiv2023

Minimum-Risk Recalibration of Classifiers

Recalibrating probabilistic classifiers is vital for enhancing the reliability and accuracy of predictive models. Despite the development of numerous recalibration algorithms, there is still a lack of a comprehensive theory that integrates calibration and sharpness (which is essential for maintaining predictive power). In this paper, we introduce the concept of minimum-risk recalibration within the framework of mean-squared-error (MSE) decomposition, offering a principled approach for evaluating and recalibrating probabilistic classifiers. Using this framework, we analyze the uniform-mass binning (UMB) recalibration method and establish a finite-sample risk upper bound of order $\tilde{O}(B/n + 1/B^2)$ where $B$ is the number of bins and $n$ is the sample size. By balancing calibration and sharpness, we further determine that the optimal number of bins for UMB scales with $n^{1/3}$, resulting in a risk bound of approximately $O(n^{-2/3})$. Additionally, we tackle the challenge of label shift by proposing a two-stage approach that adjusts the recalibration function using limited labeled data from the target domain. Our results show that transferring a calibrated classifier requires significantly fewer target samples compared to recalibrating from scratch. We validate our theoretical findings through numerical simulations, which confirm the tightness of the proposed bounds, the optimal number of bins, and the effectiveness of label shift adaptation.

cs.LG↗

arXiv2023

Performative Federated Learning: A Solution to Model-Dependent and Heterogeneous Distribution Shifts

We consider a federated learning (FL) system consisting of multiple clients and a server, where the clients aim to collaboratively learn a common decision model from their distributed data. Unlike the conventional FL framework that assumes the client's data is static, we consider scenarios where the clients' data distributions may be reshaped by the deployed decision model. In this work, we leverage the idea of distribution shift mappings in performative prediction to formalize this model-dependent data distribution shift and propose a performative federated learning framework. We first introduce necessary and sufficient conditions for the existence of a unique performative stable solution and characterize its distance to the performative optimal solution. Then we propose the performative FedAvg algorithm and show that it converges to the performative stable solution at a rate of O(1/T) under both full and partial participation schemes. In particular, we use novel proof techniques and show how the clients' heterogeneity influences the convergence. Numerical results validate our analysis and provide valuable insights into real-world applications.

cs.LG↗