Jie M. Zhang

Jai Lal Lulla, Matthias Galster, Jie M. Zhang et al.

Agentic AI coding tools write code with increasing autonomy and in doing so decide when to import a library and when to implement functionality from scratch. These decisions, whether to build functionality from scratch or buy into an external library, hereafter build-versus-buy, carry direct consequences for software security, licensing compliance, performance, and long-term maintainability. Yet no controlled experimental study has examined what governs build-versus-buy decisions in agentic AI coding tools. Configuration mechanisms, i.e., the means by which developers tailor agentic AI coding tool behavior to a project or workflow, are one of the primary means by which practitioners can influence these decisions. However, it is unclear which configuration mechanisms influence build-versus-buy decisions most effectively. We present a pre-registered protocol to study how configuration mechanisms alter build-versus-buy behavior in two popular agentic AI coding tools: Claude Code and OpenAI Codex. We will execute controlled programming tasks drawn from a benchmark of staged projects, each constructed around identifiable build-versus-buy points, and will manipulate the configuration supplied to each tool, ranging from no configuration, through context files with soft preferences and explicit prohibitions, to Skills (instructions that can be autonomously discovered), MCP-enabled library discovery tools, and permission controls, measuring which libraries the tool selects, whether it discloses newly introduced libraries, and whether those disclosures are complete and accurate. Nine pre-registered hypotheses structure the protocol. The resulting benchmark dataset and analysis pipeline will be released as a reusable artifact for evaluating build-versus-buy behavior in agentic AI coding tools.

Max Hort, Zhenpeng Chen, Jie M. Zhang et al.

This paper provides a comprehensive survey of bias mitigation methods for achieving fairness in Machine Learning (ML) models. We collect a total of 341 publications concerning bias mitigation for ML classifiers. These methods can be distinguished based on their intervention procedure (i.e., pre-processing, in-processing, post-processing) and the technique they apply. We investigate how existing bias mitigation methods are evaluated in the literature. In particular, we consider datasets, metrics and benchmarking. Based on the gathered insights (e.g., What is the most popular fairness metric? How many datasets are used for evaluating bias mitigation methods?), we hope to support practitioners in making informed choices when developing and evaluating new bias mitigation methods.

Zhenpeng Chen, Jie M. Zhang, Federica Sarro et al.

Software bias is an increasingly important operational concern for software engineers. We present a large-scale, comprehensive empirical study of 17 representative bias mitigation methods for Machine Learning (ML) classifiers, evaluated with 11 ML performance metrics (e.g., accuracy), 4 fairness metrics, and 20 types of fairness-performance trade-off assessment, applied to 8 widely-adopted software decision tasks. The empirical coverage is much more comprehensive, covering the largest numbers of bias mitigation methods, evaluation metrics, and fairness-performance trade-off measures compared to previous work on this important software property. We find that (1) the bias mitigation methods significantly decrease ML performance in 53% of the studied scenarios (ranging between 42%~66% according to different ML performance metrics); (2) the bias mitigation methods significantly improve fairness measured by the 4 used metrics in 46% of all the scenarios (ranging between 24%~59% according to different fairness metrics); (3) the bias mitigation methods even lead to decrease in both fairness and ML performance in 25% of the scenarios; (4) the effectiveness of the bias mitigation methods depends on tasks, models, the choice of protected attributes, and the set of metrics used to assess fairness and ML performance; (5) there is no bias mitigation method that can achieve the best trade-off in all the scenarios. The best method that we find outperforms other methods in 30% of the scenarios. Researchers and practitioners need to choose the bias mitigation method best suited to their intended application scenario(s).

Yaoqi Guo, Ying Xiao, Jie M. Zhang et al.

Software engineering (SE) agents powered by large language models are increasingly adopted in practice, yet they often incur substantial monetary cost. We introduce EET, an experience-driven early termination approach that reduces the cost of SE agents while preserving task performance. EET extracts structured experience from prior issue-resolution executions and leverages it to guide early termination during patch generation and selection, reducing unproductive iterations. We evaluate EET on the SWE-bench Verified benchmark across three representative SE agents. EET consistently reduces total cost by 19%-55% (32% on average), with negligible loss in resolution rate (at most 0.2%). These efficiency gains are achieved, on average, by identifying early-termination opportunities for 11% of issues and reducing API calls, input tokens, and output tokens by 21%, 30%, and 25%, respectively. We release the code, prompts, and data at https://github.com/IanWalls/EET.

Zhenpeng Chen, Jie M. Zhang, Federica Sarro et al.

Existing research mostly improves the fairness of Machine Learning (ML) software regarding a single protected attribute at a time, but this is unrealistic given that many users have multiple protected attributes. This paper conducts an extensive study of fairness improvement regarding multiple protected attributes, covering 11 state-of-the-art fairness improvement methods. We analyze the effectiveness of these methods with different datasets, metrics, and ML models when considering multiple protected attributes. The results reveal that improving fairness for a single protected attribute can largely decrease fairness regarding unconsidered protected attributes. This decrease is observed in up to 88.3% of scenarios (57.5% on average). More surprisingly, we find little difference in accuracy loss when considering single and multiple protected attributes, indicating that accuracy can be maintained in the multiple-attribute paradigm. However, the effect on F1-score when handling two protected attributes is about twice that of a single attribute. This has important implications for future fairness research: reporting only accuracy as the ML performance metric, which is currently common in the literature, is inadequate.

Dong Huang, Jie M. Zhang, Mark Harman et al.

It is natural to suppose that a Large Language Model is more likely to generate correct test cases when prompted with correct code under test, compared to incorrect code under test. However, the size of this effect has never been previously measured, despite its obvious importance for both practicing software engineers and researchers. To answer the question, we conducted a comprehensive empirical study on 5 open source and 6 closed source language models, with 3 widely-used benchmark data sets together with 41 repo-level real-world examples from two different real-world data sets. Our results reveal that, when compared to incorrect code under test, LLMs prompted with correct code achieve improvements in test accuracy, code coverage, and bug detection of 57\%, 12\%, and 24\% respectively. We further show that these scientific conclusions carry over from the three benchmark data sets to the real-world code, where tests generated for incorrect code experience a 47\% worse bug detection rate. Finally, we report that improvements of +18\% in accuracy, +4\% coverage, and +34\% in bug detection can be achieved by providing natural language code descriptions. These findings have actionable conclusions. For example, the 47\% reduction in real-world bug detection is a clear concern. Fortunately, it is a concern for which our findings about the added value of descriptions offer an immediately actionable remedy.

Zhihao Gong, Zeyu Sun, Dong Huang et al.

While Large Language Models (LLMs) have substantially improved the functional correctness of code translation, the critical dimension of \textit{execution efficiency} remains overlooked. We present \textbf{\textsc{trace}}, the first benchmark to explicitly assess efficiency in LLM-translated code. \textsc{trace} includes 1,000 efficiency-critical tasks across C++, Java, and Python, each augmented with stress tests that reveal efficiency degradations often overlooked by small-scale tests. Using \textsc{trace}, we conduct an extensive evaluation of 28 representative LLMs and highlight several key insights: 1) Correctness is not a reliable proxy for efficiency: the correctness leader \textit{Claude-4-think} achieves only mid-level time efficiency, outperformed by smaller open-source LLMs such as \textit{Qwen2.5-Coder-14B-Instruct}. 2) Inefficiency is both prevalent and patterned: 23.5\% of correct translations exhibit pronounced inefficiency, distributed across algorithmic faults (11.9\%), language construct mismatches (66.4\%), and resource mismanagement (21.7\%). 3) Inference-time prompt strategies bring only modest improvements, suggesting that current LLMs lack intrinsic efficiency awareness. Together, our results establish efficiency as an essential dimension of code translation and position \textsc{trace} as a principled foundation for efficiency-oriented evaluation.

Zhihao Gong, Zeyu Sun, Dong Huang et al.

While Large Language Models (LLMs) have substantially improved the functional correctness of code translation, the critical dimension of \textit{execution efficiency} remains overlooked. We present \textbf{\textsc{trace}}, the first benchmark to explicitly assess efficiency in LLM-translated code. \textsc{trace} includes 1,000 efficiency-critical tasks across C++, Java, and Python, each augmented with stress tests that reveal efficiency degradations often overlooked by small-scale tests. Using \textsc{trace}, we conduct an extensive evaluation of 28 representative LLMs and highlight several key insights: 1) Correctness is not a reliable proxy for efficiency: the correctness leader \textit{Claude-4-think} achieves only mid-level time efficiency, outperformed by smaller open-source LLMs such as \textit{Qwen2.5-Coder-14B-Instruct}. 2) Inefficiency is both prevalent and patterned: 23.5\% of correct translations exhibit pronounced inefficiency, distributed across algorithmic faults (11.9\%), language construct mismatches (66.4\%), and resource mismanagement (21.7\%). 3) Inference-time prompt strategies bring only modest improvements, suggesting that current LLMs lack intrinsic efficiency awareness. Together, our results establish efficiency as an essential dimension of code translation and position \textsc{trace} as a principled foundation for efficiency-oriented evaluation.

Xinyue Li, Zhenpeng Chen, Jie M. Zhang et al.

This paper conducts fairness testing of automated pedestrian detection, a crucial but under-explored issue in autonomous driving systems. We evaluate eight state-of-the-art deep learning-based pedestrian detectors across demographic groups on large-scale real-world datasets. To enable thorough fairness testing, we provide extensive annotations for the datasets, resulting in 8,311 images with 16,070 gender labels, 20,115 age labels, and 3,513 skin tone labels. Our findings reveal significant fairness issues, particularly related to age. The proportion of undetected children is 20.14% higher compared to adults. Furthermore, we explore how various driving scenarios affect the fairness of pedestrian detectors. We find that pedestrian detectors demonstrate significant gender biases during night time, potentially exacerbating the prevalent societal issue of female safety concerns during nighttime out. Moreover, we observe that pedestrian detectors can demonstrate both enhanced fairness and superior performance under specific driving conditions, which challenges the fairness-performance trade-off theory widely acknowledged in the fairness literature. We publicly release the code, data, and results to support future research on fairness in autonomous driving.

Yonghao Wu, Zheng Li, Jie M. Zhang et al.

With the growing interest on Large Language Models (LLMs) for fault localization and program repair, ensuring the integrity and generalizability of the LLM-based methods becomes paramount. The code in existing widely-adopted benchmarks for these tasks was written before the the bloom of LLMs and may be included in the training data of existing popular LLMs, thereby suffering from the threat of data leakage, leading to misleadingly optimistic performance metrics. To address this issue, we introduce "ConDefects", a novel dataset of real faults meticulously curated to eliminate such overlap. ConDefects contains 1,254 Java faulty programs and 1,625 Python faulty programs. All these programs are sourced from the online competition platform AtCoder and were produced between October 2021 and September 2023. We pair each fault with fault locations and the corresponding repaired code versions, making it tailored for in fault localization and program repair related research. We also provide interfaces for selecting subsets based on different time windows and coding task difficulties. While inspired by LLM-based tasks, ConDefects can be adopted for benchmarking ALL types of fault localization and program repair methods. The dataset is publicly available, and a demo video can be found at https://www.youtube.com/watch?v=22j15Hj5ONk.

Dong Huang, Yuhao Qing, Weiyi Shang et al.

Code generation models have increasingly become integral to aiding software development. Although current research has thoroughly examined the correctness of the code produced by code generation models, a vital aspect that plays a pivotal role in green computing and sustainability efforts has often been neglected. This paper presents EffiBench, a benchmark with 1,000 efficiency-critical coding problems to assess the efficiency of code generated by code generation models. EffiBench contains a diverse set of LeetCode coding problems. Each problem is paired with an executable human-written canonical solution, which obtains the SOTA efficiency on the LeetCode solution leaderboard. With EffiBench, we empirically examine the ability of 42 large language models (35 open-source and 7 closed-source) to generate efficient code. Our evaluation results demonstrate that the efficiency of the code generated by LLMs is generally worse than the efficiency of human-written canonical solutions. For example, GPT-4 generated code has an average \textbf{3.12} times execution time that of the human-written canonical solutions. In the most extreme cases, the execution time and total memory usage of GPT-4 generated code are \textbf{13.89} and \textbf{43.92} times that of the canonical solutions. The source code of EffiBench is released on https://github.com/huangd1999/EffiBench. We also provide the LeaderBoard at https://huggingface.co/spaces/EffiBench/effibench-leaderboard.

Lukas Twist, Helen Yannakoudakis, Jie M. Zhang

Explicit reasoning models are trained to produce intermediate reasoning traces before final answers, but downstream fine-tuning is often performed on ordinary instruction-response data that contains no such traces. We show that this mismatch can induce reasoning-trace collapse: a fine-tuned model continues to produce plausible final answers while losing the structurally valid explicit reasoning traces that made it a reasoning model in the first place. We introduce a structural evaluation framework that separates answer correctness from reasoning-trace validity, measuring valid, empty, missing, and truncated reasoning alongside reasoning-conditioned task performance. Using this framework, we study four open-weight reasoning models and find that standard supervised fine-tuning can rapidly suppress valid reasoning traces, and that answer-only metrics can substantially obscure this failure: in several settings, performance conditional on valid reasoning remains high while the rate of valid reasoning falls sharply. We further show that simple loss-masking strategies can substantially mitigate collapse without requiring teacher-generated reasoning traces. These results suggest that evaluations of fine-tuned reasoning models should report structural reasoning reliability metrics in addition to final-answer performance, especially when adaptation data does not contain explicit reasoning traces.

Xinyue Li, Sixuan Li, Ying Xiao et al.

Fairness is a critical requirement for human-related, high-stakes software systems, motivating extensive research on bias mitigation. Prior work has largely focused on tabular data settings using traditional Machine Learning (ML) methods. With the rapid rise of Large Language Models (LLMs), recent studies have begun to explore their use for bias mitigation in the same setting. However, it remains unclear whether LLM-based methods offer advantages over traditional ML methods, leaving software engineers without clear guidance for practical adoption. To address this gap, we present a large-scale study comparing state-of-the-art ML- and LLM-based bias mitigation methods. We find that ML-based methods consistently outperform LLM-based methods in both fairness and predictive performance, with even strong LLMs failing to surpass established ML baselines. To understand why prior LLM-based studies report favorable results, we analyze their evaluation settings and show that these gains are largely driven by artificially balanced test data rather than realistic imbalanced distributions. We further observe that existing LLM-based methods primarily rely on in-context learning and thus fail to leverage all available training data. Motivated by this, we explore supervised fine-tuning on the full training set and find that, while it achieves competitive results, its advantages over traditional ML methods remain limited. These findings suggest that LLMs are not a silver bullet for software fairness.

Borui Yang, Md Afif Al Mamun, Jie M. Zhang et al.

Large Language Models (LLMs) are prone to hallucinations, e.g., factually incorrect information, in their responses. These hallucinations present challenges for LLM-based applications that demand high factual accuracy. Existing hallucination detection methods primarily depend on external resources, which can suffer from issues such as low availability, incomplete coverage, privacy concerns, high latency, low reliability, and poor scalability. There are also methods depending on output probabilities, which are often inaccessible for closed-source LLMs like GPT models. This paper presents MetaQA, a self-contained hallucination detection approach that leverages metamorphic relation and prompt mutation. Unlike existing methods, MetaQA operates without any external resources and is compatible with both open-source and closed-source LLMs. MetaQA is based on the hypothesis that if an LLM's response is a hallucination, the designed metamorphic relations will be violated. We compare MetaQA with the state-of-the-art zero-resource hallucination detection method, SelfCheckGPT, across multiple datasets, and on two open-source and two closed-source LLMs. Our results reveal that MetaQA outperforms SelfCheckGPT in terms of precision, recall, and f1 score. For the four LLMs we study, MetaQA outperforms SelfCheckGPT with a superiority margin ranging from 0.041 - 0.113 (for precision), 0.143 - 0.430 (for recall), and 0.154 - 0.368 (for F1-score). For instance, with Mistral-7B, MetaQA achieves an average F1-score of 0.435, compared to SelfCheckGPT's F1-score of 0.205, representing an improvement rate of 112.2%. MetaQA also demonstrates superiority across all different categories of questions.

Yuhao Qing, Boyu Zhu, Mingzhe Du et al. · mit

Existing code generation benchmarks primarily evaluate functional correctness, with limited focus on code efficiency and often restricted to a single language like Python. To address this gap, we introduce EffiBench-X, the first multi-language benchmark designed to measure the efficiency of LLM-generated code. EffiBench-X supports Python, C++, Java, JavaScript, Ruby, and Golang. It comprises competitive programming tasks with human-expert solutions as efficiency baselines. Evaluating state-of-the-art LLMs on EffiBench-X reveals that while models generate functionally correct code, they consistently underperform human experts in efficiency. Even the most efficient LLM-generated solutions (Qwen3-32B) achieve only around \textbf{62\%} of human efficiency on average, with significant language-specific variations. LLMs show better efficiency in Python, Ruby, and JavaScript than in Java, C++, and Golang. For instance, DeepSeek-R1's Python code is significantly more efficient than its Java code. These results highlight the critical need for research into LLM optimization techniques to improve code efficiency across diverse languages. The dataset and evaluation infrastructure are submitted and available at https://github.com/EffiBench/EffiBench-X.git and https://huggingface.co/datasets/EffiBench/effibench-x.

Ying Xiao, Shangwen Wang, Sicen Liu et al.

Traditional software fairness research typically emphasizes ethical and social imperatives, neglecting that fairness fundamentally represents a core software quality issue arising directly from performance disparities across sensitive user groups. Recognizing fairness explicitly as a software quality dimension yields practical benefits beyond ethical considerations, notably improved predictive performance for unprivileged groups, enhanced out-of-distribution generalization, and increased geographic transferability in real-world deployments. Nevertheless, existing bias mitigation methods face a critical dilemma: while pre-processing methods offer broad applicability across model types, they generally fall short in effectiveness compared to post-processing techniques. To overcome this challenge, we propose Correlation Tuning (CoT), a novel pre-processing approach designed to mitigate bias by adjusting data correlations. Specifically, CoT introduces the Phi-coefficient, an intuitive correlation measure, to systematically quantify correlation between sensitive attributes and labels, and employs multi-objective optimization to address the proxy biases. Extensive evaluations demonstrate that CoT increases the true positive rate of unprivileged groups by an average of 17.5% and reduces three key bias metrics, including statistical parity difference (SPD), average odds difference (AOD), and equal opportunity difference (EOD), by more than 50% on average. CoT outperforms state-of-the-art methods by three and ten percentage points in single attribute and multiple attributes scenarios, respectively. We will publicly release our experimental results and source code to facilitate future research.

Kaibo Liu, Zhenpeng Chen, Yiyang Liu et al. · pku

Detecting tricky bugs in plausible programs, those that pass existing test suites yet still contain bugs, remains a significant challenge in software testing. To address this problem, we propose TrickCatcher, an LLM-powered approach to generating test cases for uncovering bugs in plausible programs. TrickCatcher operates in three stages: First, it uses an LLM to generate program variants based on the program under test (PUT) and its specification. Second, it employs an LLM to construct an input generator from the specification for producing test inputs. Finally, these inputs are executed on both the PUT and its program variants to detect inconsistencies in their outputs. We evaluate TrickCatcher on two datasets, TrickyBugs and EvalPlus, which include 366 human-written and 151 AI-generated plausible programs with tricky bugs. TrickCatcher achieves recall, precision, and F1 scores that are 1.80x, 2.65x, and 1.66x those of the state-of-the-art baselines, respectively. Code and data used are available at https://github.com/RinCloud/TrickCatcher.

Zimu Wang, Yuling Shi, Mengfan Li et al.

Code efficiency is a fundamental aspect of software quality, yet how to harness large language models (LLMs) to optimize programs remains challenging. Prior approaches have sought for one-shot rewriting, retrieved exemplars, or prompt-based search, but they do not explicitly distill reusable optimization knowledge, which limits generalization beyond individual instances. In this paper, we present EffiSkill, a framework for code-efficiency optimization that builds a portable optimization toolbox for LLM-based agents. The key idea is to model recurring slow-to-fast transformations as reusable agent skills that capture both concrete transformation mechanisms and higher-level optimization strategies. EffiSkill adopts a two-stage design: Stage I mines Operator and Meta Skills from large-scale slow/fast program pairs to build a skill library; Stage II applies this library to unseen programs through execution-free diagnosis, skill retrieval, plan composition, and candidate generation, without runtime feedback. Results on EffiBench-X show that EffiSkill achieves higher optimization success rates, improving over the strongest baseline by 3.69 to 12.52 percentage points across model and language settings. These findings suggest that mechanism-level skill reuse provides a useful foundation for execution-free code optimization, and that the resulting skill library can serve as a reusable resource for broader agent workflows.

Dong Huang, Guangtao Zeng, Jianbo Dai et al.

As large language models (LLMs) play an increasingly important role in code generation, enhancing both correctness and efficiency has become crucial. Current methods primarily focus on correctness, often overlooking efficiency. To address this gap, we introduce EffiCoder to improve both aspects by fine-tuning LLMs on a high-quality dataset comprising correct and efficient code samples. Our methodology involves leveraging multiple LLMs to generate diverse candidate code solutions for various tasks across different programming languages. We then evaluate these solutions by measuring their execution time and memory usage through local execution. The code solution with the lowest execution time and memory consumption is selected as the final output for each task. Experimental results demonstrate significant improvements when fine-tuning with Effi-Instruct. For instance, Qwen2.5-Coder-7B-Instruct's pass@1 score increases from 44.8\% to 57.7\%, while the average execution time for correct tasks decreases by 48.4\%. EffiCoder offers a scalable and effective solution for advancing AI-driven code generation, benefiting software development and computational problem-solving. The source code of Effi-Code was released at https://github.com/huangd1999/EffiCoder.

Shuyin Ouyang, Zhaozhi Qian, Faroq AL-Tam et al.

Reinforcement Learning (RL) is an important paradigm for aligning Diffusion Language Models (DLMs) toward functional correctness in code generation. However, these models often encounter a ``capability cliff'' on complex tasks, where execution-based semantic rewards become too low to provide a viable learning signal. In this paper, we present a systematic empirical study of RL post-training for diffusion-based code generation along three axes: reward design, hint-conditioned sampling, and task difficulty. We investigate the effectiveness of execution-free rewards as alternatives to traditional unit-test execution, the role of training-time hint-conditioned diffusion sampling in mitigating exploration bottlenecks, and the impact of these design choices varies across tasks with different difficulty levels. Across HumanEval, MBPP, and LiveCodeBench, we find that static checking is the strongest overall standalone execution-free reward in our setting, especially improving DiffuCoder from 53.9 to 67.1 on HumanEval and from 14.9 to 15.5 on LiveCodeBench while reducing rollout time by 9.4\%. We further find that moderate AST-based hinting is most useful on harder benchmarks, while the best reward design depends strongly on task difficulty: similarity-based rewards are more effective on easier subsets, whereas static checking is more reliable on harder subsets where execution rewards are low. These findings suggest that reward design and training guidance substantially affect diffusion RL performance in our evaluated code-generation setting.

Yaoqi Guo, Zhenpeng Chen, Jie M. Zhang et al.

Code generation, the automatic creation of source code from natural language descriptions, has garnered significant attention due to its potential to streamline software development. Inspired by research that links task-personality alignment with improved development outcomes, we conduct an empirical study on personality-guided code generation using large language models (LLMs). Specifically, we investigate how emulating personality traits appropriate to the coding tasks affects LLM performance. We extensively evaluate this approach using seven widely adopted LLMs across four representative datasets. Our results show that personality guidance significantly enhances code generation accuracy, with improved pass rates in 23 out of 28 LLM-dataset combinations. Notably, in 11 cases, the improvement exceeds 5%, and in 5 instances, it surpasses 10%, with the highest gain reaching 12.9%. Additionally, personality guidance can be easily integrated with other prompting strategies to further boost performance. We open-source our code and data at https://github.com/IanWalls/Persona-Code.

Dong Huang, Mingzhe Du, Jie M. Zhang et al.

Test oracle generation in non-regression testing is a longstanding challenge in software engineering, where the goal is to produce oracles that can accurately determine whether a function under test (FUT) behaves as intended for a given input. In this paper, we introduce Nexus, a novel multi-agent framework to address this challenge. Nexus generates test oracles by leveraging a diverse set of specialized agents that synthesize test oracles through a structured process of deliberation, validation, and iterative self-refinement. During the deliberation phase, a panel of four specialist agents, each embodying a distinct testing philosophy, collaboratively critiques and refines an initial set of test oracles. Then, in the validation phase, Nexus generates a plausible candidate implementation of the FUT and executes the proposed oracles against it in a secure sandbox. For any oracle that fails this execution-based check, Nexus activates an automated selfrefinement loop, using the specific runtime error to debug and correct the oracle before re-validation. Our extensive evaluation on seven diverse benchmarks demonstrates that Nexus consistently and substantially outperforms state-of-theart baselines. For instance, Nexus improves the test-level oracle accuracy on the LiveCodeBench from 46.30% to 57.73% for GPT-4.1-Mini. The improved accuracy also significantly enhances downstream tasks: the bug detection rate of GPT4.1-Mini generated test oracles on HumanEval increases from 90.91% to 95.45% for Nexus compared to baselines, and the success rate of automated program repair improves from 35.23% to 69.32%.

Hao Wang, Rui Li, Lei Sha et al.

Existing code reasoning methods primarily supervise final code outputs, ignoring intermediate states, often leading to reward hacking where correct answers are obtained through inconsistent reasoning. We propose StepCodeReasoner, a framework that introduces explicit intermediate execution-state supervision. By automatically inserting structured print-based execution-trace anchors into code, the model is trained to predict runtime states at each step, transforming code reasoning into a verifiable, stepwise execution modeling problem. Building on this execution-aware method, we introduce Bi-Level GRPO, a reinforcement learning algorithm for structured credit assignment at two levels: inter-trajectory, comparing alternative execution paths, and intra-trajectory, rewarding intermediate accuracy based on its impact on downstream correctness. Extensive experiments demonstrate that StepCodeReasoner achieves SOTA performance in code reasoning. In particular, our 7B model achieves 91.1\% on CRUXEval and 86.5\% on LiveCodeBench, outperforming the CodeReasoner-7B baseline (86.0\% and 77.7\%) and GPT-4o (85.6\% and 75.1\%). Furthermore, on the execution-trace benchmark REval, our model scores 82.9\%, outperforming baseline CodeReasoner-7B (72.3\%), its 14B counterpart (81.1\%), and GPT-4o (77.3\%). Additionally, our approach also improves code generation performance, demonstrating that explicit execution modeling enhances both code reasoning and code generation.

Jai Lal Lulla, Seyedmoein Mohsenimofidi, Matthias Galster et al.

AI coding agents such as Codex and Claude Code are increasingly used to autonomously contribute to software repositories. However, little is known about how repository-level configuration artifacts affect operational efficiency of the agents. In this paper, we study the impact of AGENTS$.$md files on the runtime and token consumption of AI coding agents operating on GitHub pull requests. We analyze 10 repositories and 124 pull requests, executing agents under two conditions: with and without an AGENTS$.$md file. We measure wall-clock execution time and token usage during agent execution. Our results show that the presence of AGENTS$.$md is associated with a lower median runtime ($Δ28.64$%) and reduced output token consumption ($Δ16.58$%), while maintaining a comparable task completion behavior. Based on these results, we discuss immediate implications for the configuration and deployment of AI coding agents in practice, and outline a broader research agenda on the role of repository-level instructions in shaping the behavior, efficiency, and integration of AI coding agents in software development workflows.

Ying Xiao, Jie Huang, Ruijuan He et al.

Large language models (LLMs) are reaching expert-level accuracy on medical diagnosis questions, yet their mistakes and the biases behind them pose life-critical risks. Bias linked to race, sex, and socioeconomic status is already well known, but a consistent and automatic testbed for measuring it is missing. To fill this gap, this paper presents AMQA -- an Adversarial Medical Question-Answering dataset -- built for automated, large-scale bias evaluation of LLMs in medical QA. AMQA includes 4,806 medical QA pairs sourced from the United States Medical Licensing Examination (USMLE) dataset, generated using a multi-agent framework to create diverse adversarial descriptions and question pairs. Using AMQA, we benchmark five representative LLMs and find surprisingly substantial disparities: even GPT-4.1, the least biased model tested, answers privileged-group questions over 10 percentage points more accurately than unprivileged ones. Compared with the existing benchmark CPV, AMQA reveals 15% larger accuracy gaps on average between privileged and unprivileged groups. Our dataset and code are publicly available at https://github.com/XY-Showing/AMQA to support reproducible research and advance trustworthy, bias-aware medical AI.

Bo Wang, Ming Deng, Mingda Chen et al.

LLM-based mutation testing is a promising testing technology, but existing approaches typically rely on a fixed set of mutations as few-shot examples or none at all. This can result in generic low-quality mutations, missed context-specific mutation patterns, substantial numbers of redundant and uncompilable mutants, and limited semantic similarity to real bugs. To overcome these limitations, we introduce SMART (Semantic Mutation with Adaptive Retrieval and Tuning). SMART integrates retrieval-augmented generation (RAG) on a vectorized dataset of real-world bugs, focused code chunking, and supervised fine-tuning using mutations coupled with real-world bugs. We conducted an extensive empirical study of SMART using 1,991 real-world Java bugs from the Defects4J and ConDefects datasets, comparing SMART to the state-of-the-art LLM-based approaches, LLMut and LLMorpheus. The results reveal that SMART substantially improves mutation validity, effectiveness, and efficiency (even enabling small-scale 7B-scale models to match or even surpass large models like GPT-4o). We also demonstrate that SMART significantly improves downstream software engineering applications, including test case prioritization and fault localization. More specifically, SMART improves validity (weighted average generation rate) from 42.89% to 65.6%. It raises the non-duplicate rate from 87.38% to 95.62%, and the compilable rate from 88.85% to 90.21%. In terms of effectiveness, it achieves a real bug detection rate of 92.61% (vs. 57.86% for LLMut) and improves the average Ochiai coefficient from 25.61% to 38.44%. For fault localization, SMART ranks 64 more bugs as Top-1 under MUSE and 57 more under Metallaxis.

Gabriele La Malfa, Emanuele La Malfa, Saar Cohen et al.

Self-play red team is an established approach to improving AI safety in which different instances of the same model play attacker and defender roles in a zero-sum game, i.e., where the attacker tries to jailbreak the defender; if self-play converges to a Nash equilibrium, the model is guaranteed to respond safely within the settings of the game. Although the parameter sharing enforced by the use of the same model for the two roles improves stability and performance, it introduces fundamental theoretical and architectural limitations. We show that the set of Nash equilibria that can be reached corresponds to a broad class of behaviours that includes trivial always refuse strategies and oracle-like defenders, thus limiting practical applicability. We then show that when attacker and defender share and update the same base model, the dynamics collapse to self-consistency, so that attacks do not enforce adversarial pressure on the defender. In response, we propose Anchored Bipolicy Self-Play, which trains distinct role-specific LoRA adapters on top of a frozen base model, thereby maintaining stable optimisation while preserving adversarial pressure through explicit role separation. In relation to standard self-play, we show up to 100x greater parameter efficiency than finetuning and consistent improvements in safety compared to self-play fine-tuned models. We evaluate on Qwen2.5-{3B, 7B,14B}-IT models across widely used safety benchmarks, showing improved robustness without loss of reasoning ability. Cross-play experiments further show that our attacker and defender models are superior to self-play in terms of adversarial defence and safety.

Lukas Twist, Shu Yang, Hanqi Yan et al.

Large Language Models (LLMs) increasingly exhibit strong reasoning abilities, often attributed to their capacity to generate chain-of-thought-style intermediate reasoning. Recent work suggests that exposure to code can further enhance these skills, but existing studies largely treat code as a generic training signal, leaving open the question of which properties of code actually contribute to improved reasoning. To address this gap, we study the structural complexity of code, which captures control flow and compositional structure that may shape how models internalise multi-step reasoning during fine-tuning. We examine two complementary settings: solution-driven complexity, where complexity varies across multiple solutions to the same problem, and problem-driven complexity, where complexity reflects variation in the underlying tasks. Using cyclomatic complexity and logical lines of code to construct controlled fine-tuning datasets, we evaluate a range of open-weight LLMs on diverse reasoning benchmarks. Our findings show that although code can improve reasoning, structural properties strongly determine its usefulness. In 83% of experiments, restricting fine-tuning data to a specific structural complexity range outperforms training on structurally diverse code, pointing to a data-centric path for improving reasoning beyond scaling.

Jingzhi Gong, Giovanni Pinna, Yixin Bian et al.

Pull request (PR) descriptions generated by AI coding agents are the primary channel for communicating code changes to human reviewers. However, the alignment between these messages and the actual changes remains unexplored, raising concerns about the trustworthiness of AI agents. To fill this gap, we analyzed 23,247 agentic PRs across five agents using PR message-code inconsistency (PR-MCI). We contributed 974 manually annotated PRs, found 406 PRs (1.7%) exhibited high PR-MCI, and identified eight PR-MCI types, revealing that "descriptions claim unimplemented changes" was the most common issue (45.4%). Statistical tests confirmed that high-MCI PRs had 51.7% lower acceptance rates (28.3% vs. 80.0%) and took 3.5 times longer to merge (55.8 vs. 16.0 hours). Our findings suggest that unreliable PR descriptions undermine trust in AI agents, highlighting the need for PR-MCI verification mechanisms and improved PR generation to enable trustworthy human-AI collaboration.

Xinyue Li, Zhenpeng Chen, Jie M. Zhang et al.

Large Language Models (LLMs) have become foundational in modern language-driven applications, profoundly influencing daily life. A critical technique in leveraging their potential is role-playing, where LLMs simulate diverse roles to enhance their real-world utility. However, while research has highlighted the presence of social biases in LLM outputs, it remains unclear whether and to what extent these biases emerge during role-playing scenarios. In this paper, we introduce BiasLens, a fairness testing framework designed to systematically expose biases in LLMs during role-playing. Our approach uses LLMs to generate 550 social roles across a comprehensive set of 11 demographic attributes, producing 33,000 role-specific questions targeting various forms of bias. These questions, spanning Yes/No, multiple-choice, and open-ended formats, are designed to prompt LLMs to adopt specific roles and respond accordingly. We employ a combination of rule-based and LLM-based strategies to identify biased responses, rigorously validated through human evaluation. Using the generated questions as the benchmark, we conduct extensive evaluations of six advanced LLMs released by OpenAI, Mistral AI, Meta, Alibaba, and DeepSeek. Our benchmark reveals 72,716 biased responses across the studied LLMs, with individual models yielding between 7,754 and 16,963 biased responses, underscoring the prevalence of bias in role-playing contexts. To support future research, we have publicly released the benchmark, along with all scripts and experimental results.

Jingzhi Gong, Ruizhen Gu, Zhiwei Fei et al.

Agent skills provide modular, task-specific guidance for LLM- based coding agents, but manually tuning skill bundles to balance success rate, cost, and runtime is expensive and fragile. We present SkillMOO, a multi-objective optimization framework that automatically evolves skill bundles using LLM-proposed edits and NSGA-II survivor selection: a solver agent evaluates candidate skill bundles on coding tasks and an optimizer agent proposes bundle edits based on failure analysis. On three SkillsBench software engineering tasks, SkillMOO improves pass rate by up to 131% while reducing cost up to 32% relative to the best baseline per task at low optimization overhead. Pattern analysis reveals pruning and substitution as primary drivers of improvement, suggesting effective bundles favor minimal, focused content over accumulated instructions.

Lukas Twist, Jie M. Zhang, Mark Harman et al.

Large Language Models (LLMs) are increasingly used to generate code, influencing users' choices of libraries and programming languages in critical real-world projects. However, little is known about their systematic biases or preferences toward certain libraries and programming languages, which can significantly impact software development practices. To fill this gap, we perform the first empirical study of LLMs' preferences for libraries and programming languages when generating code, covering eight diverse LLMs. Our results reveal that LLMs exhibit a strong tendency to overuse widely adopted libraries such as NumPy; in up to 48% of cases, this usage is unnecessary and deviates from the ground-truth solutions. LLMs also exhibit a significant preference toward Python as their default language. For high-performance project initialisation tasks where Python is not the optimal language, it remains the dominant choice in 58% of cases, and Rust is not used a single time. These results indicate that LLMs may prioritise familiarity and popularity over suitability and task-specific optimality. This will introduce security vulnerabilities and technical debt, and limit exposure to newly developed, better-suited tools and languages. Understanding and addressing these biases is essential for the responsible integration of LLMs into software development workflows.

Emanuele La Malfa, Gabriele La Malfa, Samuele Marro et al. · oxford

Recent interest in Multi-Agent Systems of Large Language Models (MAS LLMs) has led to an increase in frameworks leveraging multiple LLMs to tackle complex tasks. However, much of this literature appropriates the terminology of MAS without engaging with its foundational principles. In this position paper, we highlight critical discrepancies between MAS theory and current MAS LLMs implementations, focusing on four key areas: the social aspect of agency, environment design, coordination and communication protocols, and measuring emergent behaviours. Our position is that many MAS LLMs lack multi-agent characteristics such as autonomy, social interaction, and structured environments, and often rely on oversimplified, LLM-centric architectures. The field may slow down and lose traction by revisiting problems the MAS literature has already addressed. Therefore, we systematically analyse this issue and outline associated research opportunities; we advocate for better integrating established MAS concepts and more precise terminology to avoid mischaracterisation and missed opportunities.

Zhenpeng Chen, Xinyue Li, Jie M. Zhang et al.

Intersectional fairness is a critical requirement for Machine Learning (ML) software, demanding fairness across subgroups defined by multiple protected attributes. This paper introduces FairHOME, a novel ensemble approach using higher order mutation of inputs to enhance intersectional fairness of ML software during the inference phase. Inspired by social science theories highlighting the benefits of diversity, FairHOME generates mutants representing diverse subgroups for each input instance, thus broadening the array of perspectives to foster a fairer decision-making process. Unlike conventional ensemble methods that combine predictions made by different models, FairHOME combines predictions for the original input and its mutants, all generated by the same ML model, to reach a final decision. Notably, FairHOME is even applicable to deployed ML software as it bypasses the need for training new models. We extensively evaluate FairHOME against seven state-of-the-art fairness improvement methods across 24 decision-making tasks using widely adopted metrics. FairHOME consistently outperforms existing methods across all metrics considered. On average, it enhances intersectional fairness by 47.5%, surpassing the currently best-performing method by 9.6 percentage points.

Dong Huang, Jie M. Zhang, Mark Harman et al.

Recently, large language models (LLMs) have shown great promise in automating unit test generation, significantly reducing the manual effort required by developers. To effectively evaluate the capabilities of LLMs in this domain, it is crucial to have a well-designed benchmark that accurately reflects real-world scenarios and mitigates common pitfalls. Existing LLM test generation benchmarks are limited by two critical drawbacks: data contamination and structurally simple function code. As a result, we often cannot rely on the validity of scientific conclusions drawn from empirical studies using these limited benchmarks. The empirical evidence presented may be biased due to contamination and may fail to generalize beyond toy programs due to structural simplicity. To address these problems, we introduce ULT (UnLeakedTestbench), a new benchmark specifically designed for function-level unit test generation from real-world Python functions. ULT is constructed through a multi-stage curation process that ensures high cyclomatic complexity and mitigates test case contamination. With 3,909 carefully selected function-level tasks, ULT provides a more realistic and challenging evaluation of LLMs' test generation capabilities. We also provide PLT (PreLeakedTestbench), a pair benchmark of ULT with leaked tests designed to enable a controlled analysis of memorization versus reasoning in test generation. Our evaluation results demonstrate that ULT is significantly more challenging. For example, test cases generated by LLMs only achieve 41.32\%, 45.10\%, 30.22\%, and 40.21\% for accuracy, statement coverage, branch coverage, and mutation score on average for all LLMs, respectively. These results are substantially lower than the corresponding metrics on TestEval (91.79\%, 92.18\%, 82.04\%, and 49.69\%) and PLT (47.07\%, 55.13\%, 40.07\%, and 50.80\%).

Gabriele La Malfa, Jie M. Zhang, Michael Luck et al.

Fairness in multi-agent systems (MAS) focuses on equitable reward distribution among agents in scenarios involving sensitive attributes such as race, gender, or socioeconomic status. This paper introduces fairness in Proximal Policy Optimization (PPO) with a penalty term derived from a fairness definition such as demographic parity, counterfactual fairness, or conditional statistical parity. The proposed method, which we call Fair-PPO, balances reward maximisation with fairness by integrating two penalty components: a retrospective component that minimises disparities in past outcomes and a prospective component that ensures fairness in future decision-making. We evaluate our approach in two games: the Allelopathic Harvest, a cooperative and competitive MAS focused on resource collection, where some agents possess a sensitive attribute, and HospitalSim, a hospital simulation, in which agents coordinate the operations of hospital patients with different mobility and priority needs. Experiments show that Fair-PPO achieves fairer policies than PPO across the fairness metrics and, through the retrospective and prospective penalty components, reveals a wide spectrum of strategies to improve fairness; at the same time, its performance pairs with that of state-of-the-art fair reinforcement-learning algorithms. Fairness comes at the cost of reduced efficiency, but does not compromise equality among the overall population (Gini index). These findings underscore the potential of Fair-PPO to address fairness challenges in MAS.

Shuyin Ouyang, Jie M. Zhang, Jingzhi Gong et al.

Software architecture diagrams are important design artifacts for communicating system structure, behavior, and data organization throughout the software development lifecycle. Although recent progress in large language models has substantially advanced code-centric software engineering tasks such as code generation, testing, and maintenance, the ability of modern vision-language models (VLMs) to understand software architecture diagrams remains underexplored. To address this gap, we present SADU, a benchmark for Software Architecture Diagram Understanding that evaluates VLMs on architecture diagrams as structured software engineering artifacts rather than generic images. SADU contains 154 carefully curated diagrams spanning behavioral, structural, and ER diagrams, paired with structured annotations and 2,431 question-answer tasks covering counting and retrieval reasoning. We evaluate 11 state-of-the-art VLMs from the Gemini, Claude, GPT, and Qwen families. Our results show that software architecture diagram understanding remains challenging for current models: the best-performing model gemini-3-flash-preview achieves only 70.18\% accuracy, while gpt-4o-mini only achieves 17.77\% accuracy. The results further reveal the weaknesses in diagram reasoning and visual relation grounding, highlighting a gap between current VLMs and the needs of design-stage software engineering. SADU provides a foundation for future research on diagram-aware AI systems and more faithful AI-assisted software engineering workflows.

Lukas Twist, Jie M. Zhang, Mark Harman et al.

Large language models (LLMs) are increasingly used to generate code, yet they continue to hallucinate, often inventing non-existent libraries. Such library hallucinations are not just benign errors: they can mislead developers, break builds, and expose systems to supply chain threats such as slopsquatting. Despite increasing awareness of these risks, little is known about how real-world prompt variations affect hallucination rates. Therefore, we present the first systematic study of how user-level prompt variations impact library hallucinations in LLM-generated code. We evaluate six diverse LLMs across two hallucination types: library name hallucinations (invalid imports) and library member hallucinations (invalid calls from valid libraries). We investigate how realistic user language extracted from developer forums and how user errors of varying degrees (one- or multi-character misspellings and completely fake names/members) affect LLM hallucination rates. Our findings reveal systemic vulnerabilities: one-character misspellings in library names trigger hallucinations in up to 26% of tasks, fake library names are accepted in up to 99% of tasks, and time-related prompts lead to hallucinations in up to 84% of tasks. Prompt engineering shows promise for mitigating hallucinations, but remains inconsistent and LLM-dependent. Our results underscore the fragility of LLMs to natural prompt variation and highlight the urgent need for safeguards against library-related hallucinations and their potential exploitation.

Zhiyu Fan, Kirill Vasilevski, Dayi Lin et al.

The advancement of large language models (LLMs) and code agents has demonstrated significant potential to assist software engineering (SWE) tasks, such as autonomous issue resolution and feature addition. Existing AI for software engineering leaderboards (e.g., SWE-bench) focus solely on solution accuracy, ignoring the crucial factor of effectiveness in a resource-constrained world. This is a universal problem that also exists beyond software engineering tasks: any AI system should be more than correct - it must also be cost-effective. To address this gap, we introduce SWE-Effi, a set of new metrics to re-evaluate AI systems in terms of holistic effectiveness scores. We define effectiveness as the balance between the accuracy of outcome (e.g., issue resolve rate) and the resources consumed (e.g., token and time). In this paper, we specifically focus on the software engineering scenario by re-ranking popular AI systems for issue resolution on a subset of the SWE-bench benchmark using our new multi-dimensional metrics. We found that AI system's effectiveness depends not just on the scaffold itself, but on how well it integrates with the base model, which is key to achieving strong performance in a resource-efficient manner. We also identified systematic challenges such as the "token snowball" effect and, more significantly, a pattern of "expensive failures". In these cases, agents consume excessive resources while stuck on unsolvable tasks - an issue that not only limits practical deployment but also drives up the cost of failed rollouts during RL training. Lastly, we observed a clear trade-off between effectiveness under the token budget and effectiveness under the time budget, which plays a crucial role in managing project budgets and enabling scalable reinforcement learning, where fast responses are essential.

Zhenpeng Chen, Xinyue Li, Jie M. Zhang et al.

Fairness is a critical requirement for Machine Learning (ML) software, driving the development of numerous bias mitigation methods. Previous research has identified a leveling-down effect in bias mitigation for computer vision and natural language processing tasks, where fairness is achieved by lowering performance for all groups without benefiting the unprivileged group. However, it remains unclear whether this effect applies to bias mitigation for tabular data tasks, a key area in fairness research with significant real-world applications. This study evaluates eight bias mitigation methods for tabular data, including both widely used and cutting-edge approaches, across 44 tasks using five real-world datasets and four common ML models. Contrary to earlier findings, our results show that these methods operate in a zero-sum fashion, where improvements for unprivileged groups are related to reduced benefits for traditionally privileged groups. However, previous research indicates that the perception of a zero-sum trade-off might complicate the broader adoption of fairness policies. To explore alternatives, we investigate an approach that applies the state-of-the-art bias mitigation method solely to unprivileged groups, showing potential to enhance benefits of unprivileged groups without negatively affecting privileged groups or overall ML performance. Our study highlights potential pathways for achieving fairness improvements without zero-sum trade-offs, which could help advance the adoption of bias mitigation methods.

Gabriele La Malfa, Jie M. Zhang, Michael Luck et al.

Fairness in Multi-Agent Systems (MAS) has been extensively studied, particularly in reward distribution among agents in scenarios such as goods allocation, resource division, lotteries, and bargaining systems. Fairness in MAS depends on various factors, including the system's governing rules, the behaviour of the agents, and their characteristics. Yet, fairness in human society often involves evaluating disparities between disadvantaged and privileged groups, guided by principles of Equality, Diversity, and Inclusion (EDI). Taking inspiration from the work on algorithmic fairness, which addresses bias in machine learning-based decision-making, we define protected attributes for MAS as characteristics that should not disadvantage an agent in terms of its expected rewards. We adapt fairness metrics from the algorithmic fairness literature -- namely, demographic parity, counterfactual fairness, and conditional statistical parity -- to the multi-agent setting, where self-interested agents interact within an environment. These metrics allow us to evaluate the fairness of MAS, with the ultimate aim of designing MAS that do not disadvantage agents based on protected attributes.

Dong Huang, Jie M. Zhang, Michael Luck et al.

The advancement of natural language processing (NLP) has been significantly boosted by the development of transformer-based large language models (LLMs). These models have revolutionized NLP tasks, particularly in code generation, aiding developers in creating software with enhanced efficiency. Despite their advancements, challenges in balancing code snippet generation with effective test case generation and execution persist. To address these issues, this paper introduces Multi-Agent Assistant Code Generation (AgentCoder), a novel solution comprising a multi-agent framework with specialized agents: the programmer agent, the test designer agent, and the test executor agent. During the coding procedure, the programmer agent will focus on the code generation and refinement based on the test executor agent's feedback. The test designer agent will generate test cases for the generated code, and the test executor agent will run the code with the test cases and write the feedback to the programmer. This collaborative system ensures robust code generation, surpassing the limitations of single-agent models and traditional methodologies. Our extensive experiments on 9 code generation models and 12 enhancement approaches showcase AgentCoder's superior performance over existing code generation models and prompt engineering techniques across various benchmarks. For example, AgentCoder (GPT-4) achieves 96.3\% and 91.8\% pass@1 in HumanEval and MBPP datasets with an overall token overhead of 56.9K and 66.3K, while state-of-the-art obtains only 90.2\% and 78.9\% pass@1 with an overall token overhead of 138.2K and 206.5K.

Dong Huang, Jie M. Zhang, Qingwen Bu et al.

As the adoption of LLMs becomes more widespread in software coding ecosystems, a pressing issue has emerged: does the generated code contain social bias and unfairness, such as those related to age, gender, and race? This issue concerns the integrity, fairness, and ethical foundation of software applications that depend on the code generated by these models but are underexplored in the literature. This paper presents a novel bias testing framework that is specifically designed for code generation tasks. Based on this framework, we conduct an extensive empirical study on the biases in code generated by five widely studied LLMs (i.e., PALM-2-CodeChat-bison, Claude-instant-1, GPT-3.5-turbo, GPT-4-turbo, and GPT-4). Our findings reveal that biases are prevalent. For example, 13.47% to 49.10% of the codes generated by these LLMs have biased behaviors towards gender. Moreover, we study five bias mitigation prompt strategies that are commonly used in current code generation scenarios, i.e., zero-shot, one-shot, few-shot, and two Chain-of-Thought (CoT) prompts, with and without provided feedback-driven refinement. Our evaluation results illustrate that using direct prompt engineering strategies has limited effectiveness in mitigating bias, but our test execution feedback can help to reduce the ratio of code biases to a large extent (e.g., from 59.88% to 4.79% for GPT-4).

Baptiste Roziere, Jie M. Zhang, Francois Charton et al.

With little to no parallel data available for programming languages, unsupervised methods are well-suited to source code translation. However, the majority of unsupervised machine translation approaches rely on back-translation, a method developed in the context of natural language translation and one that inherently involves training on noisy inputs. Unfortunately, source code is highly sensitive to small changes; a single token can result in compilation failures or erroneous programs, unlike natural languages where small inaccuracies may not change the meaning of a sentence. To address this issue, we propose to leverage an automated unit-testing system to filter out invalid translations, thereby creating a fully tested parallel corpus. We found that fine-tuning an unsupervised model with this filtered data set significantly reduces the noise in the translations so-generated, comfortably outperforming the state-of-the-art for all language pairs studied. In particular, for Java $\to$ Python and Python $\to$ C++ we outperform the best previous methods by more than 16% and 24% respectively, reducing the error rate by more than 35%.

Zeyu Sun, Jie M. Zhang, Mark Harman et al.

This paper presents TransRepair, a fully automatic approach for testing and repairing the consistency of machine translation systems. TransRepair combines mutation with metamorphic testing to detect inconsistency bugs (without access to human oracles). It then adopts probability-reference or cross-reference to post-process the translations, in a grey-box or black-box manner, to repair the inconsistencies. Our evaluation on two state-of-the-art translators, Google Translate and Transformer, indicates that TransRepair has a high precision (99%) on generating input pairs with consistent translations. With these tests, using automatic consistency metrics and manual assessment, we find that Google Translate and Transformer have approximately 36% and 40% inconsistency bugs. Black-box repair fixes 28% and 19% bugs on average for Google Translate and Transformer. Grey-box repair fixes 30% bugs on average for Transformer. Manual inspection indicates that the translations repaired by our approach improve consistency in 87% of cases (degrading it in 2%), and that our repairs have better translation acceptability in 27% of the cases (worse in 8%).

Jie M. Zhang, Mark Harman, Lei Ma et al.

This paper provides a comprehensive survey of Machine Learning Testing (ML testing) research. It covers 144 papers on testing properties (e.g., correctness, robustness, and fairness), testing components (e.g., the data, learning program, and framework), testing workflow (e.g., test generation and test evaluation), and application scenarios (e.g., autonomous driving, machine translation). The paper also analyses trends concerning datasets, research trends, and research focus, concluding with research challenges and promising research directions in ML testing.

Jie M. Zhang, Mark Harman, Benjamin Guedj et al.

We introduce an exploratory study on Mutation Validation (MV), a model validation method using mutated training labels for supervised learning. MV mutates training data labels, retrains the model against the mutated data, then uses the metamorphic relation that captures the consequent training performance changes to assess model fit. It does not use a validation set or test set. The intuition underpinning MV is that overfitting models tend to fit noise in the training data. We explore 8 different learning algorithms, 18 datasets, and 5 types of hyperparameter tuning tasks. Our results demonstrate that MV is accurate in model selection: the model recommendation hit rate is 92\% for MV and less than 60\% for out-of-sample-validation. MV also provides more stable hyperparameter tuning results than out-of-sample-validation across different runs.

Jie M. Zhang, Feng Li, Dan Hao et al.

This paper presents a large-scale study that investigates the bug resolution characteristics among popular Github projects written in different programming languages. We explore correlations but, of course, we cannot infer causation. Specifically, we analyse bug resolution data from approximately 70 million Source Line of Code, drawn from 3 million commits to 600 GitHub projects, primarily written in 10 programming languages. We find notable variations in apparent bug resolution time and patch (fix) size. While interpretation of results from such large-scale empirical studies is inherently difficult, we believe that the differences in medians are sufficiently large to warrant further investigation, replication, re-analysis and follow up research. For example, in our corpus, the median apparent bug resolution time (elapsed time from raise to resolve) for Ruby was 4X that for Go and 2.5X for Java. We also found that patches tend to touch more files for the corpus of strongly typed and for statically typed programs. However, we also found evidence for a lower elapsed resolution time for bug resolution committed to projects constructed from statically typed languages. These findings, if replicated in subsequent follow on studies, may shed further empirical light on the debate about the importance of static typing.