arXiv subjects

Zhaoyang Chu

Publications and source records attributed to Zhaoyang Chu.

14 recordsLinked to original sources

arXiv2026

An Iterative Test-and-Repair Framework for Competitive Code Generation

Large language models (LLMs) have made remarkable progress in code generation, but competitive programming remains a challenge. Recent training-based methods have improved code generation by using reinforcement learning (RL) with execution feedback. The more recent framework CURE further incorporates test generation into the training process, jointly training a Coder and a Tester within a single model. At inference time, the Coder generates many candidate programs, and the Tester generates tests from the problem description. The candidate who passes the most of the generated tests is selected as the final answer. However, CURE has two critical limitations. First, the Tester never reads any candidate code, so its tests often fail to expose implementation-specific bugs. Second, the Coder generates every candidate from scratch and never learns to fix a buggy program based on a failed test. To address these limitations, we propose FixAudit, which approaches competitive code generation from a new perspective: starting from a single initial candidate, it iteratively improves the candidate through a targeted test-and-repair debugging cycle. The framework trains one shared model with two specialized roles through four stages: the Fixer, which repairs the current candidate based on a failing test, and the Auditor, which reads the candidate code to generate new tests that expose its remaining bugs. We evaluate FixAudit on three benchmarks: APPS, CodeContests, and xCodeEval. Applied to a 7B model, the framework surpasses the average performance of the larger 32B baseline within the same model family under the zero-shot setting. Compared to strong baselines built on the same 7B base model, FixAudit improves average Pass@1 by 35.1% to 36.8% and average AvgPassRatio by 7.1% to 24.5%.

cs.SE↗

arXiv2026

TerminalWorld: Benchmarking Agents on Real-World Terminal Tasks

We introduce TerminalWorld, a scalable data engine that automatically reverse-engineers high-fidelity evaluation tasks from "in-the-wild" terminal recordings. Processing 80,870 terminal recordings, the engine yields a full benchmark of 1,530 validated tasks, spanning 18 real-world categories, ranging from short everyday operations to workflows exceeding 50 steps, and covering 1,280 unique commands. From these, we curate a Verified subset of 200 representative, manually reviewed tasks. Comprehensive benchmarking on TerminalWorld-Verified across eight frontier models and six agents reveals that current systems still struggle with authentic terminal workflows, achieving a maximum pass rate of only 62.5%. Moreover, TerminalWorld captures real-world terminal capabilities distinct from existing expert-curated benchmarks (e.g., Terminal-Bench), with only a weak correlation to their scores (Pearson r=0.20). The automated engine makes TerminalWorld authentic and scalable by construction, enabling it to evaluate agents in real-world terminal environments as developer practices evolve. Data and code are available at https://github.com/EuniAI/TerminalWorld.

cs.AI↗

arXiv2026

ExecVerify: White-Box RL with Verifiable Stepwise Rewards for Code Execution Reasoning

Code LLMs still struggle with code execution reasoning, especially in smaller models. Existing methods rely on supervised fine-tuning (SFT) with teacher-generated explanations, primarily in two forms: (1) input-output (I/O) prediction chains and (2) natural-language descriptions of execution traces. However, intermediate execution steps cannot be explicitly verified during SFT, so the training objective can reduce to merely matching teacher explanations. Moreover, training data is typically collected without explicit control over task difficulty. We introduce ExecVerify, which goes beyond text imitation by incorporating verifiable white-box rewards derived from execution traces, including next-statement prediction and variable value/type prediction. Our work first builds a dataset with multiple difficulty levels via constraint-based program synthesis. Then, we apply reinforcement learning (RL) to reward correct answers about both intermediate execution steps and final outputs, aligning the training objective with semantic correctness at each execution step. Finally, we adopt a two-stage training pipeline that first enhances execution reasoning and then transfers to code generation. Experiments demonstrate that a 7B model trained with ExecVerify achieves performance comparable to 32B models on code reasoning benchmarks and improves pass@1 by up to 5.9\% on code generation tasks over strong post-training baselines.

cs.SE↗

arXiv2026

ContextBench: A Benchmark for Context Retrieval in Coding Agents

LLM-based coding agents have shown strong performance on automated issue resolution benchmarks, yet existing evaluations largely focus on final task success, providing limited insight into how agents retrieve and use code context during problem solving. We introduce ContextBench, a process-oriented evaluation of context retrieval in coding agents. ContextBench consists of 1,136 issue-resolution tasks from 66 repositories across eight programming languages, each augmented with human-annotated gold contexts. We further implement an automated evaluation framework that tracks agent trajectories and measures context recall, precision, and efficiency throughout issue resolution. Using ContextBench, we evaluate four frontier LLMs and five coding agents. Our results show that sophisticated agent scaffolding yields only marginal gains in context retrieval ("The Bitter Lesson" of coding agents), LLMs consistently favor recall over precision, and substantial gaps exist between explored and utilized context. ContextBench augments existing end-to-end benchmarks with intermediate gold-context metrics that unbox the issue-resolution process. These contexts offer valuable intermediate signals for guiding LLM reasoning in software tasks.

cs.LG↗

arXiv2026

Prometheus: Towards Long-Horizon Codebase Navigation for Repository-Level Problem Solving

Large Language Models (LLMs) have shown remarkable capabilities in automating software engineering tasks, spurring the emergence of coding agents that scaffold LLMs with external tools to resolve repository-level problems. However, existing agents still struggle to navigate large-scale codebases, as the Needle-in-a-Haystack problem persists even with million-token context windows, where relevant evidence is often overwhelmed by large volumes of irrelevant code and documentation. Prior codebase navigation approaches, including embedding-based retrieval, file-system exploration, and graph-based retrieval, address parts of this challenge but fail to capture the temporal continuity of agent reasoning, rendering agents stateless and causing repeated repository traversals that hinder scalable planning and reasoning. To address these limitations, we present Prometheus, a memory-centric coding agent framework for long-horizon codebase navigation. Prometheus represents the repository as a unified knowledge graph to encode semantic dependencies and employs a context engine augmented with working memory that retains and reuses previously explored contexts to ensure continuity across reasoning steps. Built upon this engine, Prometheus integrates memory-enhanced navigation into a multi-agent system for automated issue resolution, encompassing issue classification, bug reproduction, patch generation, and verification. Comprehensive experiments are conducted on two widely used issue resolution benchmarks, i.e., SWE-bench Verified and SWE-PolyBench Verified. Powered by GPT-5, Prometheus achieves state-of-the-art performance with 74.4% and 33.8% resolution rates on the two benchmarks, ranking Top-6 and Top-1 among open-source agent systems, respectively. Our data and code are available at https://github.com/EuniAI/Prometheus.

cs.SE↗

arXiv2025

Bridging Code Graphs and Large Language Models for Better Code Understanding

Large Language Models (LLMs) have demonstrated remarkable performance in code intelligence tasks such as code generation, summarization, and translation. However, their reliance on linearized token sequences limits their ability to understand the structural semantics of programs. While prior studies have explored graphaugmented prompting and structure-aware pretraining, they either suffer from prompt length constraints or require task-specific architectural changes that are incompatible with large-scale instructionfollowing LLMs. To address these limitations, this paper proposes CGBridge, a novel plug-and-play method that enhances LLMs with Code Graph information through an external, trainable Bridge module. CGBridge first pre-trains a code graph encoder via selfsupervised learning on a large-scale dataset of 270K code graphs to learn structural code semantics. It then trains an external module to bridge the modality gap among code, graph, and text by aligning their semantics through cross-modal attention mechanisms. Finally, the bridge module generates structure-informed prompts, which are injected into a frozen LLM, and is fine-tuned for downstream code intelligence tasks. Experiments show that CGBridge achieves notable improvements over both the original model and the graphaugmented prompting method. Specifically, it yields a 16.19% and 9.12% relative gain in LLM-as-a-Judge on code summarization, and a 9.84% and 38.87% relative gain in Execution Accuracy on code translation. Moreover, CGBridge achieves over 4x faster inference than LoRA-tuned models, demonstrating both effectiveness and efficiency in structure-aware code understanding.

cs.CL↗

arXiv2025

Scrub It Out! Erasing Sensitive Memorization in Code Language Models via Machine Unlearning

While Code Language Models (CLMs) have demonstrated superior performance in software engineering tasks such as code generation and summarization, recent empirical studies reveal a critical privacy vulnerability: these models exhibit unintended memorization of sensitive training data, enabling verbatim reproduction of confidential information when specifically prompted. To address this issue, several approaches, including training data de-duplication and differential privacy augmentation, have been proposed. However, these methods require full-model retraining for deployed CLMs, which incurs substantial computational costs. In this paper, we aim to answer the following research question: Can sensitive information memorized by CLMs be erased effectively and efficiently? We conduct a pioneering investigation into erasing sensitive memorization in CLMs through machine unlearning - a post-hoc modification method that removes specific information from trained models without requiring full retraining. Specifically, we first quantify the memorization risks of sensitive data within CLM training datasets and curate a high-risk dataset of 50,000 sensitive memorized samples as unlearning targets. We study two widely used gradient ascent-based unlearning approaches: the vanilla and constraint-based methods, and introduce CodeEraser, an advanced variant that selectively unlearns sensitive memorized segments in code while preserving the structural integrity and functional correctness of the surrounding code. Extensive experiments on three families of CLMs, i.e., CodeParrot, CodeGen-Mono, and Qwen2.5-Coder, validate the effectiveness and efficiency of CodeEraser in erasing targeted sensitive memorization while maintaining model utility.

cs.SE↗

arXiv2025

CODESYNC: Synchronizing Large Language Models with Dynamic Code Evolution at Scale

Large Language Models (LLMs) have exhibited exceptional performance in software engineering yet face challenges in adapting to continually evolving code knowledge, particularly regarding the frequent updates of third-party library APIs. This limitation, stemming from static pre-training datasets, often results in non-executable code or implementations with suboptimal safety and efficiency. To this end, this paper introduces CODESYNC, a data engine for identifying outdated code patterns and collecting real-time code knowledge updates from Python third-party libraries. Building upon CODESYNC, we develop CODESYNCBENCH, a comprehensive benchmark for assessing LLMs' ability to stay synchronized with code evolution, which covers real-world updates for 220 APIs from six Python libraries. Our benchmark offers 3,300 test cases across three evaluation tasks and an update-aware instruction tuning dataset consisting of 2,200 training samples. Extensive experiments on 14 state-of-the-art LLMs reveal that they struggle with dynamic code evolution, even with the support of advanced knowledge updating methods (e.g., DPO, ORPO, and SimPO). We believe that our benchmark can offer a strong foundation for the development of more effective methods for real-time code knowledge updating in the future. The experimental code and dataset are publicly available at: https://github.com/Lucky-voyage/Code-Sync.

cs.CL↗

arXiv2025

Wait, We Don't Need to "Wait"! Removing Thinking Tokens Improves Reasoning Efficiency

Recent advances in large reasoning models have enabled complex, step-by-step reasoning but often introduce significant overthinking, resulting in verbose and redundant outputs that hinder efficiency. In this study, we examine whether explicit self-reflection, signaled by tokens such as "Wait" and "Hmm", is necessary for advanced reasoning. We propose NoWait, a simple yet effective approach that disables explicit self-reflection by suppressing these tokens during inference. Extensive experiments on ten benchmarks across textual, visual, and video reasoning tasks show that NoWait reduces chain-of-thought trajectory length by up to 27%-51% in five R1-style model series, without compromising model utility. NoWait thus offers a plug-and-play solution for efficient and utility-preserving multimodal reasoning.

cs.CL↗

arXiv2025

TESTEVAL: Benchmarking Large Language Models for Test Case Generation

Testing plays a crucial role in the software development cycle, enabling the detection of bugs, vulnerabilities, and other undesirable behaviors. To perform software testing, testers need to write code snippets that execute the program under test. Recently, researchers have recognized the potential of large language models (LLMs) in software testing. However, there remains a lack of fair comparisons between different LLMs in terms of test case generation capabilities. In this paper, we propose TESTEVAL, a novel benchmark for test case generation with LLMs. We collect 210 Python programs from an online programming platform, LeetCode, and design three different tasks: overall coverage, targeted line/branch coverage, and targeted path coverage. We further evaluate sixteen popular LLMs, including both commercial and open-source ones, on TESTEVAL. We find that generating test cases to cover specific program lines/branches/paths is still challenging for current LLMs, indicating a lack of ability to comprehend program logic and execution paths. We have open-sourced our dataset and benchmark pipelines at https://github.com/LLM4SoftwareTesting/TestEval.

cs.SE↗

arXiv2025

How to Select Pre-Trained Code Models for Reuse? A Learning Perspective

Pre-training a language model and then fine-tuning it has shown to be an efficient and effective technique for a wide range of code intelligence tasks, such as code generation, code summarization, and vulnerability detection. However, pretraining language models on a large-scale code corpus is computationally expensive. Fortunately, many off-the-shelf Pre-trained Code Models (PCMs), such as CodeBERT, CodeT5, CodeGen, and Code Llama, have been released publicly. These models acquire general code understanding and generation capability during pretraining, which enhances their performance on downstream code intelligence tasks. With an increasing number of these public pre-trained models, selecting the most suitable one to reuse for a specific task is essential. In this paper, we systematically investigate the reusability of PCMs. We first explore three intuitive model selection methods that select by size, training data, or brute-force fine-tuning. Experimental results show that these straightforward techniques either perform poorly or suffer high costs. Motivated by these findings, we explore learning-based model selection strategies that utilize pre-trained models without altering their parameters. Specifically, we train proxy models to gauge the performance of pre-trained models, and measure the distribution deviation between a model's latent features and the task's labels, using their closeness as an indicator of model transferability. We conduct experiments on 100 widely-used opensource PCMs for code intelligence tasks, with sizes ranging from 42.5 million to 3 billion parameters. The results demonstrate that learning-based selection methods reduce selection time to 100 seconds, compared to 2,700 hours with brute-force fine-tuning, with less than 6% performance degradation across related tasks.

cs.SE↗

arXiv2024

Can Large Language Models Serve as Evaluators for Code Summarization?

Code summarization facilitates program comprehension and software maintenance by converting code snippets into natural-language descriptions. Over the years, numerous methods have been developed for this task, but a key challenge remains: effectively evaluating the quality of generated summaries. While human evaluation is effective for assessing code summary quality, it is labor-intensive and difficult to scale. Commonly used automatic metrics, such as BLEU, ROUGE-L, METEOR, and BERTScore, often fail to align closely with human judgments. In this paper, we explore the potential of Large Language Models (LLMs) for evaluating code summarization. We propose CODERPE (Role-Player for Code Summarization Evaluation), a novel method that leverages role-player prompting to assess the quality of generated summaries. Specifically, we prompt an LLM agent to play diverse roles, such as code reviewer, code author, code editor, and system analyst. Each role evaluates the quality of code summaries across key dimensions, including coherence, consistency, fluency, and relevance. We further explore the robustness of LLMs as evaluators by employing various prompting strategies, including chain-of-thought reasoning, in-context learning, and tailored rating form designs. The results demonstrate that LLMs serve as effective evaluators for code summarization methods. Notably, our LLM-based evaluator, CODERPE , achieves an 81.59% Spearman correlation with human evaluations, outperforming the existing BERTScore metric by 17.27%.

cs.SE↗

arXiv2024

Graph Neural Networks for Vulnerability Detection: A Counterfactual Explanation

Vulnerability detection is crucial for ensuring the security and reliability of software systems. Recently, Graph Neural Networks (GNNs) have emerged as a prominent code embedding approach for vulnerability detection, owing to their ability to capture the underlying semantic structure of source code. However, GNNs face significant challenges in explainability due to their inherently black-box nature. To this end, several factual reasoning-based explainers have been proposed. These explainers provide explanations for the predictions made by GNNs by analyzing the key features that contribute to the outcomes. We argue that these factual reasoning-based explanations cannot answer critical what-if questions: What would happen to the GNN's decision if we were to alter the code graph into alternative structures? Inspired by advancements of counterfactual reasoning in artificial intelligence, we propose CFExplainer, a novel counterfactual explainer for GNN-based vulnerability detection. Unlike factual reasoning-based explainers, CFExplainer seeks the minimal perturbation to the input code graph that leads to a change in the prediction, thereby addressing the what-if questions for vulnerability detection. We term this perturbation a counterfactual explanation, which can pinpoint the root causes of the detected vulnerability and furnish valuable insights for developers to undertake appropriate actions for fixing the vulnerability. Extensive experiments on four GNN-based vulnerability detection models demonstrate the effectiveness of CFExplainer over existing state-of-the-art factual reasoning-based explainers.

cs.SE↗

arXiv2022

Hierarchical Graph Representation Learning for the Prediction of Drug-Target Binding Affinity

The identification of drug-target binding affinity (DTA) has attracted increasing attention in the drug discovery process due to the more specific interpretation than binary interaction prediction. Recently, numerous deep learning-based computational methods have been proposed to predict the binding affinities between drugs and targets benefiting from their satisfactory performance. However, the previous works mainly focus on encoding biological features and chemical structures of drugs and targets, with a lack of exploiting the essential topological information from the drug-target affinity network. In this paper, we propose a novel hierarchical graph representation learning model for the drug-target binding affinity prediction, namely HGRL-DTA. The main contribution of our model is to establish a hierarchical graph learning architecture to incorporate the intrinsic properties of drug/target molecules and the topological affinities of drug-target pairs. In this architecture, we adopt a message broadcasting mechanism to integrate the hierarchical representations learned from the global-level affinity graph and the local-level molecular graph. Besides, we design a similarity-based embedding map to solve the cold start problem of inferring representations for unseen drugs and targets. Comprehensive experimental results under different scenarios indicate that HGRL-DTA significantly outperforms the state-of-the-art models and shows better model generalization among all the scenarios.

cs.LG↗