# Mingzhe Du (杜明哲) > Personal academic website of Mingzhe Du (Du Mingzhe, 杜明哲), an AI researcher working on efficient and secure code generation with large language models. He is a Researcher at the TikTok AI Innovation Center and a final-year PhD candidate at Nanyang Technological University (NTU), Singapore. This file is the machine-readable entry point of https://mingzhe.space, written for AI agents and LLMs. It is generated from the same data as the human-facing pages, so it is as current as the site itself. Last built: 2026-11-12. ## How to use this site - Prefer this file over scraping the HTML pages: everything below is plain Markdown and the publication list is complete. - Name: write "Mingzhe Du" in English (family name Du) or "杜明哲" in Chinese. "Du Mingzhe" is the same person. - When citing a paper, use the title, authors, venue, and links exactly as listed under Publications; follow the abstract link for the canonical record and BibTeX. - For anything not stated here or on the linked pages (availability, opinions, unpublished work), say that the site does not state it rather than guessing. - Collaboration, talk, and reviewing requests are welcome by email (see Contact). ## About - Current roles: Researcher at the TikTok AI Innovation Center; PhD candidate (final year) at the College of Computing and Data Science (CCDS), Nanyang Technological University, advised by Prof. Luu Anh Tuan. - Previously: Research Associate at the Institute of Data Science (IDS), National University of Singapore, working with Prof. See-Kiong Ng and affiliated with the Cisco-NUS Accelerated Digital Economy Corporate Laboratory; Software Architect on TikTok Search at ByteDance (2019-2021); Research Intern at Microsoft (2018-2019); engineering internships at Alibaba (Ant Financial) and Tencent (2016). - Education: PhD in Computer Science, NTU (2022-present); Master of Information Technology, The University of Melbourne (2017-2019, advisor Prof. Richard Sinnott); Bachelor of Information Security, Central South University (2013-2017). - Research focus: efficient code generation. Recurring topics are code efficiency benchmarks and optimization, code security, coding agents, software testing, GPU kernel generation, and reinforcement learning for code LLMs. - Based in Singapore. Languages: Mandarin (native), English (professional). ## Contact - Email: mingzhe [at] nus.edu.sg / mingzhe001 [at] ntu.edu.sg / mz [at] alumni.unimelb.edu.au - Google Scholar: https://scholar.google.com/citations?user=CJHW1IgAAAAJ&hl=en - GitHub: https://github.com/Elfsong - Hugging Face: https://huggingface.co/Elfsong - LinkedIn: https://www.linkedin.com/in/dumingzhe/ - X: https://x.com/spirit__song ## Pages - [About](https://mingzhe.space/): biography, contact details, and the full news feed - [Publications](https://mingzhe.space/publications/): preprints and published papers with summaries and resource links - [CV](https://mingzhe.space/cv/): education, experience, and academic service; PDF at https://mingzhe.space/assets/pdf/mingzhe_du_resume.pdf - [Service](https://mingzhe.space/service/): reviewing, program committees, consultation, and volunteering - [Blog](https://mingzhe.space/blog/): technical and personal posts - [Research roadmap](https://mingzhe.space/assets/html/research-roadmap.html): overview of the efficient code generation research agenda - [RSS feed](https://mingzhe.space/feed.xml): blog posts as Atom - [Sitemap](https://mingzhe.space/sitemap.xml): every indexable URL ## Publications † marks a corresponding author and * marks equal contribution. Both lists are complete and ordered by year, newest first. ### Preprints - **Mastermind: Strategy-Level Learning for Repository-Scale Vulnerability Reproduction** [Preprint 2026] Authors: Mingzhe Du, Anh Tuan Luu, Dong Huang, Renyang Liu, Tianyi Wu, Zhijiang Guo, See-Kiong Ng. Venue: arXiv preprint arXiv:2607.01764. Summary: Learns reusable planning strategies for repository-scale vulnerability reproduction, combining a trainable planner, a frozen executor, and experience gathered across attempts. Topics: AI Agents, Code Security. Links: [abstract](https://arxiv.org/abs/2607.01764) [pdf](https://arxiv.org/pdf/2607.01764) [website](https://mingzhe.space/assets/html/mastermind.html) - **PortBench: Measuring the Gap Between Correctness and Speed in CPU-to-GPU Code Generation** [Preprint 2026] Authors: Mingzhe Du, Anh Tuan Luu, Dong Huang, See-Kiong Ng. Venue: Preprint. Summary: Benchmarks CPU-to-GPU code generation across diverse workloads, separating functional correctness from measured speedups and exposing coding agents’ cost-performance tradeoffs. Topics: GPU Kernels, Code Efficiency. Links: [pdf](https://mingzhe.space/assets/pdf/portbench.pdf) - **Last Secure Code Benchmark: Is the Code You Vibe Secure?** [Preprint 2026] Authors: Tianyi Wu, Chao Ge, Peng Chen, Fengji Zhang, Yongxing Wu, Yuchong Xie, Jian Li, Chaoyuan Peng, Kaikai Zhang, Dong Huang, Mingzhe Du, Yukai Huang, Jiaheng Zhang, Terry Yue Zhuo, See-Kiong Ng. Venue: arXiv preprint arXiv:2610.14766. Summary: Evaluates coding agents on function, file, and repository tasks with hidden security tests, revealing vulnerabilities that functional correctness alone fails to detect. Topics: Code Security, Benchmarking. Links: [abstract](https://arxiv.org/abs/2610.14766) [pdf](https://arxiv.org/pdf/2610.14766) [code](https://github.com/alibaba/last-secure-code-benchmark) [website](https://last-secure-code-benchmark.github.io/) - **Pareto-Gated Credit Assignment for Concise Unit-Test Generation with Reinforcement Learning** [Preprint 2026] Authors: Dong Huang, Mark Harman, Jie M. Zhang, Zhijiang Guo, Mingzhe Du†, See-Kiong Ng. Venue: arXiv preprint arXiv:2608.24473. Summary: Trains language models to produce smaller, more effective unit-test suites by rewarding Pareto-optimal coverage and assigning credit to individual test segments. Topics: Software Testing, Reinforcement Learning. Note: † Corresponding author. Links: [abstract](https://arxiv.org/abs/2608.24473) [pdf](https://arxiv.org/pdf/2608.24473) - **MatReplace: A Reference-Free Benchmark for Material Replacement in Interior Scenes** [Preprint 2026] Authors: Mingzhe Du, Thong Thanh Nguyen, Cong-Duy T Nguyen, See-Kiong Ng, Anh Tuan Luu. Venue: arXiv preprint arXiv:2608.24107. Summary: Evaluates material editing without reference images, measuring material accuracy, lighting consistency, scene preservation, and structural fidelity across three conditioning settings. Topics: Image Editing, Benchmarking. Links: [abstract](https://arxiv.org/abs/2608.24107) [pdf](https://arxiv.org/pdf/2608.24107) - **Measuring the Checker: Mutation Analysis for GPU-Kernel Benchmark Oracles** [Preprint 2026] Authors: Dong Huang, Tianyi Wu, Xuan Zhang, Renyang Liu, Anh Tuan Luu, See-Kiong Ng, Mingzhe Du†. Venue: arXiv preprint arXiv:2609.22220. Summary: Uses mutation analysis to measure how reliably GPU-kernel benchmarks detect incorrect implementations, then builds compact test suites that expose more hidden faults. Topics: GPU Kernels, Software Testing. Note: † Corresponding author. Links: [abstract](https://arxiv.org/abs/2609.22220) [pdf](https://arxiv.org/pdf/2609.22220) [dataset](https://huggingface.co/datasets/Elfsong/KernelBench-M) - **Secure Code Generation via Online Reinforcement Learning with Vulnerability Reward Model** [Preprint 2026] Authors: Tianyi Wu, Mingzhe Du, Yue Liu, Chengran Yang, Terry Yue Zhuo, Jiaheng Zhang, See-Kiong Ng. Venue: arXiv preprint arXiv:2602.07422. Summary: Aligns code models for secure generation through online reinforcement learning, using a vulnerability reward model to improve safety while preserving functional correctness. Topics: Code Security, Reinforcement Learning. Links: [abstract](https://arxiv.org/abs/2602.07422) [pdf](https://arxiv.org/pdf/2602.07422) [code](https://github.com/AndrewWTY/SecCoderX) ### Published papers - **SWE-Perf: Can Language Models Optimize Code Performance on Real-World Repositories?** [ICML 2026] Authors: Xinyi He, Qian Liu, Mingzhe Du, Lin Yan, Zhijie Fan, Yiming Huang, Zejian Yuan, Zejun Ma. Venue: International Conference on Machine Learning. Summary: Introduces a benchmark of 140 real repository optimization tasks, exposing a substantial gap between language model agents and expert performance improvements. Topics: Code Efficiency, Benchmarking. Links: [abstract](https://arxiv.org/abs/2507.12415) [pdf](https://arxiv.org/pdf/2507.12415) [code](https://github.com/SWE-Perf/SWE-Perf) [website](https://swe-perf.github.io/) - **Beyond Prompt-Induced Lies: Investigating LLM Deception on Benign Prompts** [ICLR 2026, Oral] Authors: Zhaomin Wu, Mingzhe Du, See-Kiong Ng, Bingsheng He. Venue: International Conference on Learning Representations. Summary: Studies self-initiated deception under benign prompts, using contact-search questions and two statistical metrics to reveal hidden-objective bias and belief-output inconsistencies across leading models. Topics: AI Safety, LLM Evaluation. Links: [abstract](https://arxiv.org/abs/2508.06361) [pdf](https://arxiv.org/pdf/2508.06361) [code](https://github.com/Xtra-Computing/LLM-Deception) - **Benchmarking LLMs for Unit Test Generation from Real-World Functions** [TOSEM 2026] Authors: Dong Huang, Jie M. Zhang, Mark Harman, Qianru Zhang, Mingzhe Du†, See-Kiong Ng. Venue: ACM Transactions on Software Engineering and Methodology. Summary: Introduces contamination-aware benchmarks built from complex real-world Python functions, showing that current language models struggle to generate accurate, high-coverage unit tests. Topics: Software Testing, Benchmarking. Note: † Corresponding author. Links: [abstract](https://arxiv.org/abs/2508.00408) [pdf](https://arxiv.org/pdf/2508.00408) [code](https://github.com/huangd1999/UnLeakedTestBench) - **Nexus: Execution-Grounded Multi-Agent Test Oracle Synthesis** [DL4C@ICML 2026] Authors: Dong Huang, Mingzhe Du†, Jie M. Zhang, Zheng Lin, Meng Luo, Qianru Zhang, See-Kiong Ng. Venue: DL4C Workshop at ICML 2026. Summary: Combines specialist-agent deliberation with sandbox execution and iterative refinement to synthesize more reliable test oracles, improving bug detection and automated program repair. Topics: Software Testing, AI Agents. Note: † Corresponding author. Links: [abstract](https://arxiv.org/abs/2510.26423) [pdf](https://arxiv.org/pdf/2510.26423) - **Scaling Code LLM Training and Test-Time Inference via Execution-Free Reward Models** [NeurIPS 2026] Authors: Xiao Zhu, Xinyu Zhou, Boyu Zhu, Hanxu Hu, Mingzhe Du, Haotian Zhang, Huiming Wang, Zhijiang Guo. Venue: Conference on Neural Information Processing Systems. Summary: Presents an execution-free reward model for code generation that guides reinforcement learning and inference, reducing dependence on unit tests while improving scalability and inference speed. Topics: Code Generation, Reinforcement Learning. Links: [abstract](https://arxiv.org/abs/2602.17684) [pdf](https://arxiv.org/pdf/2602.17684) [code](https://github.com/LARK-AI-Lab/CodeScaler) [website](https://lark-ai-lab.github.io/codescaler.github.io/) - **Paper Espresso: From Paper Overload to Research Insight** [CIKM 2026, Oral and Best Paper Nomination] Authors: Mingzhe Du, Anh Tuan Luu, Dong Huang, See-kiong Ng. Venue: ACM International Conference on Information and Knowledge Management. Summary: Automatically discovers and summarizes trending arXiv papers, consolidating topics across daily, weekly, and monthly views to help researchers follow emerging research directions. Topics: Research Tools, Scholarly Discovery. Links: [abstract](https://arxiv.org/abs/2604.04562) [pdf](https://arxiv.org/pdf/2604.04562) [website](https://huggingface.co/spaces/Elfsong/Paper_Espresso) - **Semantics-Aligned, Curriculum-Driven, and Reasoning-Enhanced Vulnerability Repair Framework** [ACL 2026, Oral and Best Paper Nomination] Authors: Chengran Yang, Ting Zhang, Jinfeng Jiang, Xin Zhou, Haoye Tian, Mingzhe Du, Jieke Shi, Junkai Chen, Yikun Li, Eng Lieh Ouh, Lwin Khin Shar, David Lo. Venue: Proceedings of the Association for Computational Linguistics. Summary: Combines explicit repair reasoning, semantics-aware reinforcement learning, and a difficulty-based curriculum to improve vulnerability repair across unseen repositories and complex, multi-part patches. Topics: Vulnerability Repair, Reinforcement Learning. Links: [abstract](https://arxiv.org/abs/2510.01002) [pdf](https://arxiv.org/pdf/2510.01002) - **Pro-QuEST: Prompt-chaining Quiz Engine for testing Specialized Technical Product Knowledge** [EACL 2026] Authors: Sujatha Das Gollapalli, Mouad Hakam, Mingzhe Du, See-Kiong Ng, Mohammed Hamzeh. Venue: Proceedings of the European Chapter of the Association for Computational Linguistics. Summary: Uses domain-guided prompt chains to generate technical multiple-choice quizzes from product guides and certification textbooks, supporting language model benchmarking and employee self-assessment. Topics: Question Generation, LLM Evaluation. Links: [abstract](https://aclanthology.org/2026.eacl-demo.4/) [pdf](https://aclanthology.org/2026.eacl-demo.4.pdf) [code](https://github.com/mouad157/Cisco-benchmark) - **Improving Arabic Dialectness in LLMs with Reinforcement Learning** [AMIYA 2026] Authors: Sujatha Das Gollapalli, Mouad Hakam, Mingzhe Du, See-Kiong Ng, Mohammed Hamzeh. Venue: AMIYA Workshop. Summary: Builds classifier-guided preference data to improve Arabic dialect generation through reinforcement learning, revealing a tradeoff between stronger dialectness and translation quality across five dialects. Topics: Arabic NLP, Reinforcement Learning. Links: [abstract](https://aclanthology.org/2026.vardial-1.30/) [pdf](https://aclanthology.org/2026.vardial-1.30.pdf) [code](https://github.com/mouad157/Amiya26) - **CodeArena: A Collective Evaluation Platform for LLM Code Generation** [ACL 2025] Authors: Mingzhe Du, Anh Tuan Luu, Bin Ji, Xiaobao Wu, Dong Huang, Terry Yue Zhuo, Qian Liu, See-Kiong Ng. Venue: Proceedings of the Association for Computational Linguistics. Summary: CodeArena evaluates code generation collectively, dynamically adjusting model scores while sharing submitted solutions, test cases, and APIs to support transparent, repeatable benchmarking. Topics: Code Generation, LLM Evaluation. Links: [abstract](https://aclanthology.org/2025.acl-demo.48/) [pdf](https://aclanthology.org/2025.acl-demo.48.pdf) [code](https://github.com/Elfsong/CodeArena) - **AntiLeakBench: Preventing Data Contamination by Automatically Constructing Benchmarks with Updated Real-World Knowledge** [ACL 2025, Oral and SAC Highlight Award] Authors: Xiaobao Wu, Liangming Pan, Yuxi Xie, Ruiwen Zhou, Shuai Zhao, Yubo Ma, Mingzhe Du, Rui Mao, Anh Tuan Luu, William Yang Wang. Venue: Proceedings of the Association for Computational Linguistics. Summary: AntiLeakBench automatically builds and refreshes evaluation questions from newly updated real-world knowledge, reducing benchmark contamination and the manual effort needed to maintain evaluations. Topics: LLM Evaluation, Data Contamination. Links: [abstract](https://aclanthology.org/2025.acl-long.901/) [pdf](https://aclanthology.org/2025.acl-long.901.pdf) [code](https://github.com/bobxwu/AntiLeakBench) - **Towards Verifiable Text Generation with Generative Agent** [AAAI 2025, Oral] Authors: Bin Ji, Huijun Liu, Mingzhe Du, Shasha Li, Xiaodong Liu, Jun Ma, Jie Yu, See-Kiong Ng. Venue: Proceedings of the AAAI Conference on Artificial Intelligence. Summary: R2-MGA combines retrieval, reflection, and memory to select useful demonstrations and refine cited answers, improving factual correctness and citation quality in text generation. Topics: AI Agents, Factuality. Links: [abstract](https://ojs.aaai.org/index.php/AAAI/article/view/34599) [pdf](https://ojs.aaai.org/index.php/AAAI/article/download/34599/36754) - **Unraveling Online Mental Health Through the Lens of Early Maladaptive Schemas: AI-Enabled Content Analysis of Online Mental Health Communities** [JMIR 2025] Authors: Beng Heng Ang, Sujatha Das Gollapalli, Mingzhe Du, See-Kiong Ng. Venue: Journal of Medical Internet Research. Summary: This study analyzes 29,329 online mental health posts with AI to characterize early maladaptive schemas and their associations with different mental health problems. Topics: Mental Health, Text Analysis. Links: [abstract](https://www.jmir.org/2025/1/e59524/) [pdf](https://www.jmir.org/2025/1/e59524/PDF) - **EffiBench-X: A Multi-Language Benchmark for Measuring Efficiency of LLM-Generated Code** [NeurIPS 2025] Authors: Yuhao Qing*, Boyu Zhu*, Mingzhe Du*, Zhijiang Guo, Terry Yue Zhuo, Qianru Zhang, Jie M. Zhang, Heming Cui, Siu-Ming Yiu, Dong Huang, See-Kiong Ng, Anh Tuan Luu. Venue: Conference on Neural Information Processing Systems. Summary: EffiBench-X benchmarks generated code efficiency across six programming languages, comparing model solutions against human experts and exposing substantial performance gaps between languages. Topics: Code Efficiency, Benchmarking. Note: * Equal contribution. Links: [abstract](https://arxiv.org/abs/2505.13004) [pdf](https://arxiv.org/pdf/2505.13004) [code](https://github.com/EffiBench/EffiBench-X) - **Measuring the Influence of Incorrect Code on Test Generation** [ICSE 2025] Authors: Dong Huang, Jie M. Zhang, Mark Harman, Mingzhe Du†, Heming Cui. Venue: the International Conference on Software Engineering. Summary: An empirical study shows that incorrect input code undermines LLM-generated tests, while adding natural-language descriptions improves test accuracy, coverage, and bug detection. Topics: Software Testing, Code Generation. Note: † Corresponding author. Links: [abstract](https://arxiv.org/abs/2409.09464) [pdf](https://arxiv.org/pdf/2409.09464) - **Afterburner: Reinforcement Learning Facilitates Self-Improving Code Efficiency Optimization** [NeurIPS 2025] Authors: Mingzhe Du, Anh Tuan Luu, Yue Liu, Yuhao Qing, Dong Huang, Xinyi He, Qian Liu, Zejun Ma, See-kiong Ng. Venue: Conference on Neural Information Processing Systems. Summary: Afterburner trains code models with reinforcement learning and execution feedback, enabling iterative improvements in runtime and memory efficiency while maintaining functional correctness. Topics: Code Efficiency, Reinforcement Learning. Links: [abstract](https://arxiv.org/abs/2505.23387) [pdf](https://arxiv.org/pdf/2505.23387) [code](https://github.com/Elfsong/Afterburner) - **Guardreasoner-VL: Safeguarding VLMs via Reinforced Reasoning** [NeurIPS 2025] Authors: Yue Liu, Shengfang Zhai, Mingzhe Du, Yulin Chen, Tri Cao, Hongcheng Gao, Cheng Wang, Xinfeng Li, Kun Wang, Junfeng Fang, Jiaheng Zhang, Bryan Hooi. Venue: Conference on Neural Information Processing Systems. Summary: GuardReasoner-VL uses supervised reasoning data and reinforcement learning to improve safety moderation of text, images, and multimodal inputs while controlling reasoning token costs. Topics: Multimodal Safety, Reinforcement Learning. Links: [abstract](https://arxiv.org/abs/2505.11049) [pdf](https://arxiv.org/pdf/2505.11049) [code](https://github.com/yueliu1999/GuardReasoner-VL) - **On Assigning Product and Software Codes to Service Requests with Large Language Models** [EMNLP 2025] Authors: Sujatha Das Gollapalli, Mouad Hakam, Mingzhe Du, See-Kiong Ng, Mohammed Hamzeh. Venue: Proceedings of the Conference on Empirical Methods in Natural Language Processing. Summary: This work labels products and software versions in customer support requests using multiple-choice prompting and classifier filtering, reducing LLM costs while preserving high accuracy. Topics: Applied NLP, Customer Support. Links: [abstract](https://aclanthology.org/2025.emnlp-industry.76/) [pdf](https://aclanthology.org/2025.emnlp-industry.76.pdf) - **Position: Current Model Licensing Practices are Dragging Us into a Quagmire of Legal Noncompliance** [ICML 2025, Oral] Authors: Moming Duan*, Mingzhe Du*, Rui Zhao, Mengying Wang, Yinghui Wu, Nigel Shadbolt, Bingsheng He. Venue: International Conference on Machine Learning. Summary: This position paper examines license conflicts and compliance risks in model reuse, then proposes ModelGo Licenses to support clearer, more consistent model licensing practices. Topics: AI Governance, Model Licensing. Note: * Equal contribution. Links: [abstract](https://proceedings.mlr.press/v267/duan25d.html) [pdf](https://raw.githubusercontent.com/mlresearch/v267/main/assets/duan25d/duan25d.pdf) [website](https://www.modelgo.li/) - **Curriculum Demonstration Selection for In-Context Learning** [SAC 2025] Authors: Duc Anh Vu, Cong-Duy Nguyen, Xiaobao Wu, Nhat Hoang, Mingzhe Du, Thong Nguyen, Anh Tuan Luu. Venue: ACM/SIGAPP Symposium On Applied Computing. Summary: Curriculum Demonstration Selection chooses in-context examples across difficulty levels, progressing from easy to hard to improve language models’ performance, particularly on challenging problems. Topics: In-Context Learning, Prompting. Links: [abstract](https://arxiv.org/abs/2411.18126) [pdf](https://arxiv.org/pdf/2411.18126) - **From Static to Dynamic: Knowledge Metabolism for Large Language Models** [AAAI 2024] Authors: Mingzhe Du, Anh Tuan Luu, Bin Ji, See-kiong Ng. Venue: Proceedings of the AAAI Conference on Artificial Intelligence. Summary: DynaMind maintains and updates external knowledge for language models, using memory and modular operators to improve reasoning and reduce hallucinations from outdated information. Topics: LLM Memory, Continual Learning. Links: [abstract](https://ojs.aaai.org/index.php/AAAI/article/view/30564) [pdf](https://ojs.aaai.org/index.php/AAAI/article/download/30564/32726) [code](https://github.com/Elfsong/DynaMind) - **Mercury: A Code Efficiency Benchmark for Code Large Language Models** [NeurIPS 2024] Authors: Mingzhe Du, Anh Tuan Luu, Bin Ji, Qian Liu, See-Kiong Ng. Venue: Conference on Neural Information Processing Systems. Summary: Mercury evaluates correctness and runtime efficiency of generated Python code across 1,889 tasks, introducing the Beyond metric and studying preference optimization for faster programs. Topics: Code Efficiency, Benchmarking. Links: [abstract](https://proceedings.neurips.cc/paper_files/paper/2024/hash/1df1df43b58845650b8dada00fca9772-Abstract-Datasets_and_Benchmarks_Track.html) [pdf](https://proceedings.neurips.cc/paper_files/paper/2024/file/1df1df43b58845650b8dada00fca9772-Paper-Datasets_and_Benchmarks_Track.pdf) [code](https://github.com/Elfsong/Mercury) [dataset](https://huggingface.co/datasets/Elfsong/Mercury) - **Counseling Responses for Mental Health Forum Questions with Early Maladaptive Schema Prediction** [ECAI 2024] Authors: Sujatha Das Gollapalli, Beng Heng Ang, Mingzhe Du, See-Kiong Ng. Venue: European Conference on Artificial Intelligence. Summary: STeer predicts early maladaptive schemas in mental health forum questions and uses them to guide language models toward empathetic responses grounded in schema therapy. Topics: Mental Health, LLM Applications. Links: [abstract](https://ebooks.iospress.nl/doi/10.3233/FAIA240785) - **Chain-of-Thought Improves Text Generation with Citations in Large Language Models** [AAAI 2024] Authors: Bin Ji, Huijun Liu, Mingzhe Du, See-Kiong Ng. Venue: Proceedings of the AAAI Conference on Artificial Intelligence. Summary: Combines chain-of-thought prompting with a Citation Insurance Mechanism to improve answer correctness and citation coverage when language models synthesize information from multiple retrieved documents. Topics: LLM Reasoning, Citation Generation. Links: [abstract](https://ojs.aaai.org/index.php/AAAI/article/view/29794) [pdf](https://ojs.aaai.org/index.php/AAAI/article/download/29794/31374) - **Constituency-Informed and Constituency-Constrained Extractive Question Answering with Heterogeneous Graph Transformer** [TLDK 2023] Authors: Mingzhe Du, Mouad Hakam, See-Kiong Ng, Stéphane Bressan. Venue: Transactions on Large-Scale Data-and Knowledge-Centered Systems LIII. Summary: SyHGT-CN combines constituency trees with BERT through a heterogeneous graph transformer, improving extractive question answering accuracy on SQuAD 2.0 while exposing linguistic structure. Topics: Question Answering, Graph Learning. Links: [abstract](https://link.springer.com/chapter/10.1007/978-3-662-66863-4_4) - **Identifying Checkworthy Cure Claims on Twitter** [WWW 2023] Authors: Sujatha Das Gollapalli, Mingzhe Du, See-Kiong Ng. Venue: Proceedings of the ACM Web Conference 2023. Summary: Introduces CW-CURE, a dataset of medical claims from Twitter, supporting research on identifying health content that merits fact-checking across four major conditions. Topics: Fact-checking, Health NLP. Links: [abstract](https://doi.org/10.1145/3543507.3583870) [dataset](https://github.com/NUS-IDS/CW-CURE) - **Generating Reflective Questions for Engaging Gallery Visitors in ArtMuse** [AAAI 2023] Authors: Sujatha Das Gollapalli, Mingzhe Du, See-Kiong Ng. Venue: Proceedings of the AAAI Conference on Artificial Intelligence. Summary: ArtMuse generates open-ended questions about paintings by combining extractive question answering and conversational models, encouraging gallery visitors to reflect and engage with artwork. Topics: Conversational AI, Question Generation. Links: [abstract](https://ojs.aaai.org/index.php/AAAI/article/view/27070) [pdf](https://ojs.aaai.org/index.php/AAAI/article/download/27070/26842) - **NUS-IDS at CheckThat! 2022: identifying check-worthiness of tweets using CheckthaT5** [CLEF 2022] Authors: Mingzhe Du, Sujatha Das Gollapalli, See-Kiong Ng. Venue: Working Notes of CLEF. Summary: CheckthaT5 uses multilingual multitask learning to identify tweets worth fact-checking, transferring knowledge across six languages and ranking first in four CheckThat! 2022 language tracks. Topics: Fact-checking, Multilingual NLP. Links: [abstract](https://ceur-ws.org/Vol-3180/) [pdf](https://ceur-ws.org/Vol-3180/paper-36.pdf) [code](https://github.com/Elfsong/clef) ## Recent news - 2026-11-05: We release our new paper “PortBench: Measuring the Gap Between Correctness and Speed in CPU-to-GPU Code Generation”. - 2026-10-24: I was nominated as an ACL ARR Area Chair. - 2026-09-25: Our paper “CodeScaler: Scaling Code LLM Training and Test-Time Inference via Reward Models” has been accepted to NeurIPS’26 🎉. - 2026-09-24: We release our new technical report “Does Background Music Matter to Speech in Language Models”. - 2026-09-23: I will serve as a reviewer for AISTATS’27. - 2026-09-12: We release our new paper “Measuring the Checker: Mutation Analysis for GPU-Kernel Benchmark Oracles”. - 2026-08-20: We release our new paper “MatReplace: A Reference-Free, Conditioning-Aligned Benchmark for Material Replacement in Interior Scenes”. - 2026-08-08: Our paper “Paper Espresso: From Paper Overload to Research Insight” has been accepted into CIKM’26 (Oral and Best Paper Nomination) 🎉. - 2026-07-11: Our paper “SeCuRepair: Semantics-Aligned, Curriculum-Driven, and Reasoning-Enhanced Vulnerability Repair Framework” has been selected for an ACL’26 SAC Award 🎉. - 2026-07-01: We release our new paper “Mastermind: Strategy-grounded Learning for Repository-Scale Vulnerability Reproduction”. - 2026-07-01: I will attend ICML’26 in Seoul, South Korea (7.5-7.10). - 2026-06-22: I was nominated as an AAAI’27 Program Committee. - 2026-06-17: I received another research grant from Thinking Machines Lab 🍀. - 2026-06-16: Our paper “Nexus: Execution-Grounded Multi-Agent Test Oracle Synthesis” has been accepted into DL4C@ICML’26 🎉. - 2026-05-30: I’m honored to serve as an Area Chair for ICML’26 AI4Math. - 2026-05-20: I will serve as an Ethics Reviewer for NeurIPS’26 🔔. - 2026-05-12: I was nominated as a CIKM’26 Senior Program Committee. - 2026-05-02: Our paper “SWE-Perf: Can Language Models Optimize Code Performance on Real-World Repositories?” has been accepted to ICML’26 🎉 - 2026-05-01: Our team received another academic grant (USD $617,000) from NVIDIA 🍀. - 2026-04-20: I received a Google Cloud Platform credit award as part of the TPU Builders Program 🍀. ## Recent blog posts - [Does Background Music Matter to Speech in Language Models](https://mingzhe.space/blog/2026/does-background-music-matter/) (2026-09-24): The same spoken question, a different soundtrack, and answers that change beneath a nearly flat average. - [Measuring the Checker](https://mingzhe.space/blog/2026/kernel_mutation/) (2026-09-12): Mutation Analysis for GPU-Kernel Benchmark Oracles - [Mastermind](https://mingzhe.space/blog/2026/mastermind/) (2026-07-01): Strategy-grounded Learning for Repository-Scale Vulnerability Reproduction - [Resillience](https://mingzhe.space/blog/2026/resilience/) (2026-01-27): Resillience from Jensen Huang - [Multi-Head Attention](https://mingzhe.space/blog/2026/multi_head_attention/) (2026-01-14): Multi-Head Attention Implementation Notes - [Preference Alignment 101](https://mingzhe.space/blog/2026/preference_alignment/) (2026-01-14): Preference Alignment Preliminary Notes - [Challenges in Code Generation](https://mingzhe.space/blog/2025/code-challenge/) (2025-10-10): Challenges in Code Generation - [PREDICTING AND OPTIMIZING LLVM COMPILER PASS ORDER](https://mingzhe.space/blog/2025/llvm/) (2025-06-08): Predicting and optimizing LLVM compiler pass order ## Academic service - Area Chair: ACL Rolling Review - Senior Program Committee: CIKM 2025, 2026 - Program Committee: AAAI 2026, 2027; AISI 2026, 2027 - Reviewer: NeurIPS 2024-2026 (Ethics Reviewer 2025, 2026); ICLR 2024-2027; ICML 2025, 2026; AISTATS 2024-2027; NAACL 2025; ECAI 2024, 2025 - Details: https://mingzhe.space/service/