arXiv subjects

Chaozheng Wang

Publications and source records attributed to Chaozheng Wang.

At least 37 records · Page 2Linked to original sources

arXiv2025

UMoE: Unifying Attention and FFN with Shared Experts

Sparse Mixture of Experts (MoE) architectures have emerged as a promising approach for scaling Transformer models. While initial works primarily incorporated MoE into feed-forward network (FFN) layers, recent studies have explored extending the MoE paradigm to attention layers to enhance model performance. However, existing attention-based MoE layers require specialized implementations and demonstrate suboptimal performance compared to their FFN-based counterparts. In this paper, we aim to unify MoE designs in attention and FFN layers by introducing a novel reformulation of the attention mechanism, that reveals an underlying FFN-like structure within attention modules. Our proposed architecture, UMoE, achieves superior performance through attention-based MoE layers while enabling efficient parameter sharing between FFN and attention components.

cs.LG↗

arXiv2025

SEER: Enhancing Chain-of-Thought Code Generation through Self-Exploring Deep Reasoning

Code generation, the task of creating executable programs from natural language requirements, has recently seen tremendous advances through Chain-of-Thought (CoT) reasoning, which enables Large Language Models (LLMs) to develop high-level reasoning plans before writing code. Recent research has proposed various methods to enhance models' CoT reasoning for code generation such as prompt engineering and supervised fine-tuning. However, existing approaches still face three critical limitations: (1) limited exploration of diverse reasoning paths, which constrains generalization across various programming scenarios, (2) lack of quality assessment for intermediate reasoning steps, which hampers the reliability of the generated plans and code, and (3) the potential negative impact of "overthinking", potentially leading to unnecessarily complex and incorrect solutions. To address these limitations, we frame CoT code generation as a decision making problem and present SEER, a SElf-Exploring deep Reasoning framework that enables accurate and adaptive reasoning for code generation. SEER introduces three key components: (1) Diverse reasoning path exploration, which aims at exploring diverse reasoning paths and annotating intermediate steps without relying on manual experts or closed-source proprietary models; (2) Reasoning quality-aware model training, which trains a policy model for generating candidate reasoning steps and a value model for assessing their quality; and (3) Adaptive CoT reasoning, which dynamically switches between direct generation and step-by-step reasoning for different problems.

cs.SE↗

arXiv2025

TREAT: A Code LLMs Trustworthiness / Reliability Evaluation and Testing Framework

Large foundation models are fundamentally transforming the software engineering landscape, demonstrating exceptional capabilities across diverse tasks such as code generation, debugging, and testing. Despite this rapid progress, a significant gap remains in how to comprehensively evaluate these models' trustworthiness in real-world software engineering scenarios. Existing benchmarks suffer from limited task scope and fail to incorporate critical evaluation aspects such as the robustness and reliability of models. To bridge this gap, we present an evaluation framework called TREAT (Code LLMs Trustworthiness / Reliability Evaluation And Testing) that provides a holistic assessment of model performance in code intelligence tasks. Our evaluation framework addresses key limitations in existing approaches with four main improvements: (1) Multi-Task Holistic Evaluation that spans diverse software engineering activities rather than limited coding tasks; (2) Multi-Language and Multi-Modality Assessment that extends beyond traditional single-language, text-only benchmarks to include multi-modality coding tasks; (3) Robustness Assessment that evaluates model reliability under semantically-preserving code transformations; and (4) Rigorous Evaluation Methodology that enhances the trustworthiness of evaluation results through diverse evaluation prompts and adaptive solution extraction. Based on this evaluation framework, we assess 26 state-of-the-art models and uncover both their strengths and limitations, yielding several key insights:(1) Current models show substantial performance variation across programming tasks; (2) Multi-modal language models demonstrate specific performance limitations in UI code generation and edit;

cs.SE↗

arXiv2025

CodeCrash: Exposing LLM Fragility to Misleading Natural Language in Code Reasoning

Large Language Models (LLMs) have recently demonstrated strong capabilities in code-related tasks, but their robustness in code reasoning under perturbations remains underexplored. We introduce CodeCrash, a stress-testing framework with 1,279 questions from CruxEval and LiveCodeBench, designed to evaluate reasoning reliability under structural perturbations and misleading natural language (NL) contexts. Through a systematic evaluation of 17 LLMs, we find that models often shortcut reasoning by over-relying on NL cues, leading to an average performance degradation of 23.2% in output prediction tasks. Even with Chain-of-Thought reasoning, models on average still have a 13.8% drop due to distractibility and rationalization, revealing a lack of critical reasoning capability to distinguish the actual code behaviors. While Large Reasoning Models with internal reasoning mechanisms improve robustness by fostering critical thinking, plausible yet incorrect hints can trigger pathological self-reflection, causing 2-3 times token consumption and even catastrophic cognitive dissonance in extreme cases for QwQ-32B. We refer to this phenomenon as Reasoning Collapse. CodeCrash provides a rigorous benchmark for evaluating robustness in code reasoning, guiding future research and development toward more reliable and resilient models.

cs.AI↗

arXiv2025

JSProtect: A Scalable Obfuscation Framework for Mini-Games in WeChat

The WeChat mini-game ecosystem faces rampant intellectual property theft to other platforms via secondary development, yet existing JavaScript obfuscation tools are ill-equipped for large-scale applications, suffering from prohibitive processing times, severe runtime performance degradation, and unsustainable code size inflation. This paper introduces JSProtect, a high-throughput parallelized obfuscation framework designed to overcome these fundamental limitations. At the core of our framework is the Parallel-Aware Scope Analysis (PASA) algorithm, which enables two key optimizations: independent code partitioning for multi-core processing and independent namespace management that aggressively reuses short identifiers to combat code bloat. Our evaluation demonstrates that JSProtectprocesses 20MB codebases in minutes, maintaining 100\% semantic equivalence while controlling code size inflation to as low as 20\% compared to over 1,000\% with baseline tools. Furthermore, it preserves near-native runtime performance and provides superior security effectiveness against both static analysis tools and large language models. This work presents a new paradigm for industrial-scale JavaScript protection that effectively balances robust security with high performance and scalability.

cs.SE↗

arXiv2025

JSidentify-V2: Leveraging Dynamic Memory Fingerprinting for Mini-Game Plagiarism Detection

The explosive growth of mini-game platforms has led to widespread code plagiarism, where malicious users access popular games' source code and republish them with modifications. While existing static analysis tools can detect simple obfuscation techniques like variable renaming and dead code injection, they fail against sophisticated deep obfuscation methods such as encrypted code with local or cloud-based decryption keys that completely destroy code structure and render traditional Abstract Syntax Tree analysis ineffective. To address these challenges, we present JSidentify-V2, a novel dynamic analysis framework that detects mini-game plagiarism by capturing memory invariants during program execution. Our key insight is that while obfuscation can severely distort static code characteristics, runtime memory behavior patterns remain relatively stable. JSidentify-V2 employs a four-stage pipeline: (1) static pre-analysis and instrumentation to identify potential memory invariants, (2) adaptive hot object slicing to maximize execution coverage of critical code segments, (3) Memory Dependency Graph construction to represent behavioral fingerprints resilient to obfuscation, and (4) graph-based similarity analysis for plagiarism detection. We evaluate JSidentify-V2 against eight obfuscation methods on a comprehensive dataset of 1,200 mini-games ...

cs.CR↗

arXiv2025

A Deep Dive into Retrieval-Augmented Generation for Code Completion: Experience on WeChat

Code completion, a crucial task in software engineering that enhances developer productivity, has seen substantial improvements with the rapid advancement of large language models (LLMs). In recent years, retrieval-augmented generation (RAG) has emerged as a promising method to enhance the code completion capabilities of LLMs, which leverages relevant context from codebases without requiring model retraining. While existing studies have demonstrated the effectiveness of RAG on public repositories and benchmarks, the potential distribution shift between open-source and closed-source codebases presents unique challenges that remain unexplored. To mitigate the gap, we conduct an empirical study to investigate the performance of widely-used RAG methods for code completion in the industrial-scale codebase of WeChat, one of the largest proprietary software systems. Specifically, we extensively explore two main types of RAG methods, namely identifier-based RAG and similarity-based RAG, across 26 open-source LLMs ranging from 0.5B to 671B parameters. For a more comprehensive analysis, we employ different retrieval techniques for similarity-based RAG, including lexical and semantic retrieval. Based on 1,669 internal repositories, we achieve several key findings: (1) both RAG methods demonstrate effectiveness in closed-source repositories, with similarity-based RAG showing superior performance, (2) the effectiveness of similarity-based RAG improves with more advanced retrieval techniques, where BM25 (lexical retrieval) and GTE-Qwen (semantic retrieval) achieve superior performance, and (3) the combination of lexical and semantic retrieval techniques yields optimal results, demonstrating complementary strengths. Furthermore, we conduct a developer survey to validate the practical utility of RAG methods in real-world development environments.

cs.SE↗

arXiv2025

Position Paper: Programming Language Techniques for Bridging LLM Code Generation Semantic Gaps

Large Language Models have demonstrated remarkable capabilities in automated code generation, yet their statistical nature and black-box characteristics create significant semantic gaps manifested through syntax errors, semantic hallucinations, and reliability concerns. This position paper argues that principled integration of Programming Language (PL) techniques is essential for bridging these gaps. Through structured program representations, formal correctness guarantees, and robust verification mechanisms, PL techniques can elevate LLM-generated code from statistical pattern matching to truly reliable and trustworthy levels. This integration is crucial for developing systems that generate code that is not only functionally correct but also interpretable, verifiable, and ultimately trustworthy.

cs.SE↗

arXiv2025

RAG or Fine-tuning? A Comparative Study on LCMs-based Code Completion in Industry

Code completion, a crucial practice in industrial settings, helps developers improve programming efficiency by automatically suggesting code snippets during development. With the emergence of Large Code Models (LCMs), this field has witnessed significant advancements. Due to the natural differences between open-source and industrial codebases, such as coding patterns and unique internal dependencies, it is a common practice for developers to conduct domain adaptation when adopting LCMs in industry. There exist multiple adaptation approaches, among which retrieval-augmented generation (RAG) and fine-tuning are the two most popular paradigms. However, no prior research has explored the trade-off of the two approaches in industrial scenarios. To mitigate the gap, we comprehensively compare the two paradigms including Retrieval-Augmented Generation (RAG) and Fine-tuning (FT), for industrial code completion in this paper. In collaboration with Tencent's WXG department, we collect over 160,000 internal C++ files as our codebase. We then compare the two types of adaptation approaches from three dimensions that are concerned by industrial practitioners, including effectiveness, efficiency, and parameter sensitivity, using six LCMs. Our findings reveal that RAG, when implemented with appropriate embedding models that map code snippets into dense vector representations, can achieve higher accuracy than fine-tuning alone. Specifically, BM25 presents superior retrieval effectiveness and efficiency among studied RAG methods. Moreover, RAG and fine-tuning are orthogonal and their combination leads to further improvement. We also observe that RAG demonstrates better scalability than FT, showing more sustained performance gains with larger scales of codebase.

cs.SE↗

arXiv2025

Automatically Generating UI Code from Screenshot: A Divide-and-Conquer-Based Approach

Websites are critical in today's digital world, with over 1.11 billion currently active and approximately 252,000 new sites launched daily. Converting website layout design into functional UI code is a time-consuming yet indispensable step of website development. Manual methods of converting visual designs into functional code present significant challenges, especially for non-experts. To explore automatic design-to-code solutions, we first conduct a motivating study on GPT-4o and identify three types of issues in generating UI code: element omission, element distortion, and element misarrangement. We further reveal that a focus on smaller visual segments can help multimodal large language models (MLLMs) mitigate these failures in the generation process. In this paper, we propose DCGen, a divide-and-conquer-based approach to automate the translation of webpage design to UI code. DCGen starts by dividing screenshots into manageable segments, generating code for each segment, and then reassembling them into complete UI code for the entire screenshot. We conduct extensive testing with a dataset comprised of real-world websites and various MLLMs and demonstrate that DCGen achieves up to a 15% improvement in visual similarity and 8% in code similarity for large input images. Human evaluations show that DCGen can help developers implement webpages significantly faster and more similar to the UI designs. To the best of our knowledge, DCGen is the first segment-aware MLLM-based approach for generating UI code directly from screenshots.

cs.SE↗

arXiv2025

IDInit: A Universal and Stable Initialization Method for Neural Network Training

Deep neural networks have achieved remarkable accomplishments in practice. The success of these networks hinges on effective initialization methods, which are vital for ensuring stable and rapid convergence during training. Recently, initialization methods that maintain identity transition within layers have shown good efficiency in network training. These techniques (e.g., Fixup) set specific weights to zero to achieve identity control. However, settings of remaining weight (e.g., Fixup uses random values to initialize non-zero weights) will affect the inductive bias that is achieved only by a zero weight, which may be harmful to training. Addressing this concern, we introduce fully identical initialization (IDInit), a novel method that preserves identity in both the main and sub-stem layers of residual networks. IDInit employs a padded identity-like matrix to overcome rank constraints in non-square weight matrices. Furthermore, we show the convergence problem of an identity matrix can be solved by stochastic gradient descent. Additionally, we enhance the universality of IDInit by processing higher-order weights and addressing dead neuron problems. IDInit is a straightforward yet effective initialization method, with improved convergence, stability, and performance across various settings, including large-scale datasets and deep models.

cs.LG↗

arXiv2025

The Prompt Alchemist: Automated LLM-Tailored Prompt Optimization for Test Case Generation

Test cases are essential for validating the reliability and quality of software applications. Recent studies have demonstrated the capability of Large Language Models (LLMs) to generate useful test cases for given source code. However, the existing work primarily relies on human-written plain prompts, which often leads to suboptimal results since the performance of LLMs can be highly influenced by the prompts. Moreover, these approaches use the same prompt for all LLMs, overlooking the fact that different LLMs might be best suited to different prompts. Given the wide variety of possible prompt formulations, automatically discovering the optimal prompt for each LLM presents a significant challenge. Although there are methods on automated prompt optimization in the natural language processing field, they are hard to produce effective prompts for the test case generation task. First, the methods iteratively optimize prompts by simply combining and mutating existing ones without proper guidance, resulting in prompts that lack diversity and tend to repeat the same errors in the generated test cases. Second, the prompts are generally lack of domain contextual knowledge, limiting LLMs' performance in the task.

cs.SE↗

arXiv2024

Split and Merge: Aligning Position Biases in LLM-based Evaluators

Large language models (LLMs) have shown promise as automated evaluators for assessing the quality of answers generated by AI systems. However, these LLM-based evaluators exhibit position bias, or inconsistency, when used to evaluate candidate answers in pairwise comparisons, favoring either the first or second answer regardless of content. To address this limitation, we propose PORTIA, an alignment-based system designed to mimic human comparison strategies to calibrate position bias in a lightweight yet effective manner. Specifically, PORTIA splits the answers into multiple segments, aligns similar content across candidate answers, and then merges them back into a single prompt for evaluation by LLMs. We conducted extensive experiments with six diverse LLMs to evaluate 11,520 answer pairs. Our results show that PORTIA markedly enhances the consistency rates for all the models and comparison forms tested, achieving an average relative improvement of 47.46%. Remarkably, PORTIA enables less advanced GPT models to achieve 88% agreement with the state-of-the-art GPT-4 model at just 10% of the cost. Furthermore, it rectifies around 80% of the position bias instances within the GPT-4 model, elevating its consistency rate up to 98%. Subsequent human evaluations indicate that the PORTIA-enhanced GPT-3.5 model can even surpass the standalone GPT-4 in terms of alignment with human evaluators. These findings highlight PORTIA's ability to correct position bias, improve LLM consistency, and boost performance while keeping cost-efficiency. This represents a valuable step toward a more reliable and scalable use of LLMs for automated evaluations across diverse applications.

cs.CL↗

arXiv2024

A Systematic Evaluation of Large Code Models in API Suggestion: When, Which, and How

API suggestion is a critical task in modern software development, assisting programmers by predicting and recommending third-party APIs based on the current context. Recent advancements in large code models (LCMs) have shown promise in the API suggestion task. However, they mainly focus on suggesting which APIs to use, ignoring that programmers may demand more assistance while using APIs in practice including when to use the suggested APIs and how to use the APIs. To mitigate the gap, we conduct a systematic evaluation of LCMs for the API suggestion task in the paper. To facilitate our investigation, we first build a benchmark that contains a diverse collection of code snippets, covering 176 APIs used in 853 popular Java projects. Three distinct scenarios in the API suggestion task are then considered for evaluation, including (1) ``\textit{when to use}'', which aims at determining the desired position and timing for API usage; (2) ``\textit{which to use}'', which aims at identifying the appropriate API from a given library; and (3) ``\textit{how to use}'', which aims at predicting the arguments for a given API. The consideration of the three scenarios allows for a comprehensive assessment of LCMs' capabilities in suggesting APIs for developers. During the evaluation, we choose nine popular LCMs with varying model sizes for the three scenarios. We also perform an in-depth analysis of the influence of context selection on the model performance ...

cs.SE↗

arXiv2024

ComplexCodeEval: A Benchmark for Evaluating Large Code Models on More Complex Code

In recent years, the application of large language models (LLMs) to code-related tasks has gained significant attention. However, existing evaluation benchmarks often focus on limited scenarios, such as code generation or completion, which do not reflect the diverse challenges developers face in real-world contexts. To address this, we introduce ComplexCodeEval, a benchmark designed to assess LCMs in various development tasks, including code generation, completion, API recommendation, and test case generation. It includes 3,897 Java samples and 7,184 Python samples from high-star GitHub repositories, each annotated with function signatures, docstrings, and API references to simulate real development environments. Our experiments across ten LCMs reveal that context improves performance and that data leakage can lead to overestimation, highlighting the need for more accurate evaluations.

cs.SE↗

arXiv2024

XMoE: Sparse Models with Fine-grained and Adaptive Expert Selection

Sparse models, including sparse Mixture-of-Experts (MoE) models, have emerged as an effective approach for scaling Transformer models. However, they often suffer from computational inefficiency since a significant number of parameters are unnecessarily involved in computations via multiplying values by zero or low activation values. To address this issue, we present \tool, a novel MoE designed to enhance both the efficacy and efficiency of sparse MoE models. \tool leverages small experts and a threshold-based router to enable tokens to selectively engage only essential parameters. Our extensive experiments on language modeling and machine translation tasks demonstrate that \tool can enhance model performance while decreasing the computation load at MoE layers by over 50\% without sacrificing performance. Furthermore, we present the versatility of \tool by applying it to dense models, enabling sparse computation during inference. We provide a comprehensive analysis and make our code available at https://github.com/ysngki/XMoE.

cs.LG↗

arXiv2024

Exploring Multi-Lingual Bias of Large Code Models in Code Generation

Code generation aims to synthesize code and fulfill functional requirements based on natural language (NL) specifications, which can greatly improve development efficiency. In the era of large language models (LLMs), large code models (LCMs) have been recently proposed to generate source code. LCMs can generate highly feasible solutions for programming problems described in natural language. Despite the effectiveness, we observe a noticeable multilingual bias in the generation performance of LCMs. Specifically, LCMs demonstrate proficiency in generating solutions when provided with instructions in English, yet may falter when faced with semantically equivalent instructions in other NLs such as Chinese. Moreover, the ability of LCMs to generate code exhibits variety across different programming languages (PLs), such as Python and C++. The observed phenomenon indicates the presence of multi-lingual bias within the generative capabilities of LCMs, which has remained unexplored. In this paper, we aim to investigate the multi-lingual bias that exists in current LCMs. First, we initiate our investigation by constructing the first multi-lingual evaluation benchmark X-HumanEval-X, enabling us to systematically evaluate the extent of multi-lingual bias that exists in current LCMs. In our large-scale experiments on nine popular LCMs, we observe a pronounced multi-lingual bias of LCMs in code generation, including multi-NL and multi-PL bias. Specifically, when using Chinese instructions, the code generation capabilities of LCMs decrease by at least 13% in terms of the Pass@1 metric. Furthermore, LCMs perform variously across different programming languages, e.g., the performance gap between Python and C++ reaches as high as 20.9%. ...

cs.SE↗

arXiv2023

VRPTEST: Evaluating Visual Referring Prompting in Large Multimodal Models

With recent advancements in Large Multimodal Models (LMMs) across various domains, a novel prompting method called visual referring prompting has emerged, showing significant potential in enhancing human-computer interaction within multimodal systems. This method offers a more natural and flexible approach to human interaction with these systems compared to traditional text descriptions or coordinates. However, the categorization of visual referring prompting remains undefined, and its impact on the performance of LMMs has yet to be formally examined. In this study, we conduct the first comprehensive analysis of LMMs using a variety of visual referring prompting strategies. We introduce a benchmark dataset called VRPTEST, comprising 3 different visual tasks and 2,275 images, spanning diverse combinations of prompt strategies. Using VRPTEST, we conduct a comprehensive evaluation of eight versions of prominent open-source and proprietary foundation models, including two early versions of GPT-4V. We develop an automated assessment framework based on software metamorphic testing techniques to evaluate the accuracy of LMMs without the need for human intervention or manual labeling. We find that the current proprietary models generally outperform the open-source ones, showing an average accuracy improvement of 22.70%; however, there is still potential for improvement. Moreover, our quantitative analysis shows that the choice of prompt strategy significantly affects the accuracy of LMMs, with variations ranging from -17.5% to +7.3%. Further case studies indicate that an appropriate visual referring prompting strategy can improve LMMs' understanding of context and location information, while an unsuitable one might lead to answer rejection. We also provide insights on minimizing the negative impact of visual referring prompting on LMMs.

cs.CV↗