University of Macau
澳门大学
Aug 2025 - Present2025年8月 - 至今
Master of Science in Computer Science
计算机科学理学硕士
Macau SAR, China
中国澳门特别行政区
§ 01
About
关于
Profile 个人简介I am a Master's student in Computer Science at the University of Macau, working under the supervision of Prof. Derek F. Wong at the NLP & Portuguese–Chinese Machine Translation Laboratory (NLP2CT). My research focuses on trustworthy large language models, with an emphasis on LLM-generated text detection, adversarial robustness, and safety alignment across multilingual and low-resource settings.
I am additionally interested in LLM agents and machine-translation evaluation, as well as the robustness and reliability of detectors across languages and model families. I am fortunate to be closely mentored by Junchao Wu, whose guidance has been instrumental in shaping my research trajectory.
I am actively looking for collaborations on trustworthy LLMs, multilingual NLP, and LLM-generated content detection - please feel free to reach out via email.
我是澳门大学计算机科学专业的硕士研究生，在自然语言处理与葡中机器翻译实验室（NLP2CT）由 Derek F. Wong 教授指导开展研究。我的研究方向聚焦于可信赖大语言模型，尤其关注大语言模型生成文本检测、对抗鲁棒性，以及多语言与低资源环境下的安全对齐。
我同时关注大语言模型智能体与机器翻译评估，以及检测器在不同语言和模型族之间的鲁棒性与可靠性。我有幸得到吴俊超师兄的悉心指导，他的建议对我的研究方向起到了关键作用。
我正在积极寻找合作，方向包括可信赖大语言模型、多语言自然语言处理以及大语言模型生成内容检测，欢迎通过邮件与我联系。
§ 02
Research Overview
研究概述
Thematic pillars 研究方向I.
LLM-Generated Text Detection
大语言模型生成文本检测
Building cross-lingual benchmarks and adversarial stress tests for detectors across Brahmic-script languages, with a focus on robustness under generator shifts and adversarial perturbations.
构建面向婆罗米文字系语言的跨语言基准与对抗性压力测试，重点关注检测器在生成器切换与对抗扰动下的鲁棒性。
II.
Trustworthy & Multilingual NLP
可信赖与多语言自然语言处理
Adversarial robustness, safety alignment, and evaluation of large language models in low-resource languages (Hindi, Telugu, Tamil); machine-translation evaluation and failure triage.
大语言模型在低资源语言（印地语、泰卢固语、泰米尔语）中的对抗鲁棒性、安全对齐与评估；机器翻译评估与错误分析。
III.
LLM Agents & Evaluation
大语言模型智能体与评估
Reliable evaluation protocols for agentic LLM systems; reproducible experimentation pipelines combining Docker, MLflow, and W&B across HPC and cloud environments.
面向智能体大语言模型系统的可靠评估协议；基于 Docker、MLflow 与 W&B 的可复现实验流程，覆盖高性能计算与云端环境。
§ 03
News
动态
Latest updates 最新消息Aug 20262026年8月
IndicDetect has been accepted to the Findings of EMNLP 2026! 🎉
IndicDetect 已被 EMNLP 2026 Findings 接收！🎉
May 20262026年5月
Submitted CultMetric, a gloss-grounded cultural-fidelity evaluation framework for machine translation, to ACL Rolling Review (ARR).
向 ACL Rolling Review (ARR) 投稿 CultMetric：一套基于词条释义的机器翻译文化忠实度评估框架。
Feb 20262026年2月
Submitted IndicDetect to ACL Rolling Review (ARR).
向 ACL Rolling Review (ARR) 投稿 IndicDetect。
Sep 20252025年9月
Started MSc in Computer Science at the University of Macau; appointed Teaching Assistant for Discrete Structures and Software Project Management.
开始在澳门大学攻读计算机科学硕士学位；担任《离散结构》与《软件项目管理》课程助教。
May 20252025年5月
Graduated with B.Tech (Honors) in CSE from KL University - GPA 9.00/10.00, First Class with Distinction.
获 KL 大学计算机科学与工程荣誉工学学士学位：GPA 9.00/10.00，一等优异荣誉。
May 20242024年5月
Joined the NLP2CT Laboratory as an undergraduate exchange researcher.
加入 NLP2CT 实验室，担任本科交流研究员。
§ 04
Education
教育背景
Academic record 学业履历
Koneru Lakshmaiah Education Foundation (KL University)
Koneru Lakshmaiah 教育基金会（KL 大学）
Aug 2021 - May 20252021年8月 - 2025年5月
Bachelor of Technology (Honors) in Computer Science and Engineering
计算机科学与工程荣誉工学学士
Guntur, India
印度 贡土尔
§ 05
Experience
经历
Research & Teaching 研究与教学Research Assistant (Prev. Undergraduate Exchange Researcher)
研究助理（原本科交流研究员）
Macau SAR, China
中国澳门特别行政区
- Joined as an undergraduate exchange researcher in May 2024 and continued as a graduate researcher from August 2025, contributing continuously across two years of multilingual NLP and LLM evaluation projects.
- Conceived and led IndicDetect, a cross-lingual benchmark for LLM-generated text detection across Hindi, Telugu, and Tamil; drove dataset construction, evaluation framework design, and analysis of zero-shot and fine-tuned detector performance.
- Built CultMetric, a reference-free, gloss-grounded framework for evaluating cultural fidelity in machine translation, reproducing human rankings with a deterministic LLM-judge protocol.
- Developed and optimised Transformer-based models (BERT, RoBERTa, GPT-style) for multilingual machine translation, specialising in failure triage and configuration tuning for low-resource language pairs.
- Built automated evaluation pipelines (BLEU, ROUGE, BERTScore, WER) and containerised workflows with Git, Docker, MLflow, and Weights & Biases to streamline experimentation and ensure reproducibility.
- Co-authored manuscripts submitted to ACL Rolling Review (ARR) 2026, contributing experimental design, writing, and result analysis.
- 2024 年 5 月以本科交流研究员身份加入，2025 年 8 月起继续以研究生身份开展研究，在两年间持续参与多语言自然语言处理与大语言模型评估项目。
- 主导并提出 IndicDetect：面向印地语、泰卢固语与泰米尔语的跨语言大语言模型生成文本检测基准；负责数据集构建、评估框架设计，以及零样本与微调检测器的性能分析。
- 构建 CultMetric：一套无参考译文、基于词条释义的机器翻译文化忠实度评估框架，通过确定性的大语言模型评判协议复现人工排名。
- 针对多语言机器翻译开发并优化 Transformer 模型（BERT、RoBERTa、GPT 系列），专注于低资源语对的错误分析与配置调优。
- 构建自动化评估流水线（BLEU、ROUGE、BERTScore、WER），并基于 Git、Docker、MLflow 与 Weights & Biases 搭建容器化工作流，以简化实验流程并确保可复现性。
- 合作撰写并向 ACL Rolling Review (ARR) 2026 投稿论文，参与实验设计、写作与结果分析。
University of Macau
澳门大学
Sep 2025 - June 20262025年9月 - 2026年6月
Teaching Assistant
助教
Macau SAR, China
中国澳门特别行政区
- Facilitated weekly tutorials for undergraduate courses in Discrete Structures and Software Project Management, mentoring students in algorithmic complexity and systematic quality-assurance practice.
- Engineered assessment frameworks and grading rubrics, applying test-coverage principles and edge-case analysis to evaluate student submissions.
- Collaborated with lead faculty to design supplementary instructional materials and coordinate grading across large-scale computer-science classes.
- 为本科课程《离散结构》与《软件项目管理》主持每周辅导课，指导学生掌握算法复杂度与系统化质量保障实践。
- 设计评估框架与评分标准，运用测试覆盖率原则与边界情况分析评估学生作业。
- 与主讲教师协作设计辅助教学材料，并协调大规模计算机科学课程的评分工作。
§ 06
Selected Projects
代表项目
Research output 研究成果
IndicDetect - Cross-Lingual LLM-Generated Text Detection Benchmark
IndicDetect：跨语言大语言模型生成文本检测基准
Accepted · Findings of EMNLP 2026
已接收 · EMNLP 2026 Findings
Sep 2025 - Feb 20262025年9月 - 2026年2月
- Curated a benchmark of 84K human-written and LLM-generated samples across Hindi, Telugu, and Tamil, spanning four domains (academic, news, creative, movie reviews) using GPT-4.1, Qwen-Plus, and DeepSeek-v3.2.
- Engineered seven Brahmic-script-aware adversarial attacks (paraphrase via back-translation, character perturbation, whitespace, insert-paragraph, alternative spelling, misspelling, synonym swap) to stress-test detector robustness.
- Benchmarked eight detectors, including zero-shot statistical methods (Log-Likelihood, Log-Rank, LRR, FastDetectGPT, Binoculars) and supervised neural models (XLM-RoBERTa Base/Large, QLoRA fine-tuned Qwen 2.5-7B) across six evaluation settings.
- Demonstrated that Qwen 2.5-7B achieved top average scores of 87.16 (Telugu), 85.74 (Hindi), 87.23 (Tamil), while zero-shot detectors degraded sharply under generator shifts and adversarial perturbations.
- 构建了涵盖印地语、泰卢固语与泰米尔语的 8.4 万条人写与大语言模型生成样本基准，跨越四个领域（学术、新闻、创意写作、影评），使用 GPT-4.1、Qwen-Plus 与 DeepSeek-v3.2 生成。
- 设计了七种面向婆罗米文字系的对抗性攻击（回译改写、字符扰动、空白字符、段落插入、替代拼写、拼写错误、同义词替换），用于压力测试检测器的鲁棒性。
- 在六个评估设置下对八种检测器进行基准测试，包括零样本统计方法（Log-Likelihood、Log-Rank、LRR、FastDetectGPT、Binoculars）与监督式神经模型（XLM-RoBERTa Base/Large、QLoRA 微调的 Qwen 2.5-7B）。
- 结果表明 Qwen 2.5-7B 取得最佳平均分：泰卢固语 87.16、印地语 85.74、泰米尔语 87.23，而零样本检测器在生成器切换与对抗扰动下表现显著下降。
CultMetric - Gloss-Grounded Cultural Fidelity Evaluation for Machine Translation
CultMetric：基于词条释义的机器翻译文化忠实度评估
Under Review · ACL ARR 2026
审稿中 · ACL ARR 2026
Jan 2026 - May 20262026年1月 - 2026年5月
- Designed a reference-free MT evaluation framework that grounds an LLM judge in a curated 872-entry glossary of culture-specific items, classifies failures into five typed categories, and produces a deterministic 0–100 score with bit-exact reproducibility.
- Curated the CSI glossary from Classical Chinese source texts via LLM-assisted extraction and expert validation against authoritative scholarly translations, covering religious, social, material, ecological, and linguistic categories.
- Evaluated four MT systems (GLM-5.1, DeepSeek-V4 Flash, Llama-3, Google Translate) on ~6,400 segments, reproducing the human cultural-fidelity ranking exactly - Spearman correlation 2.7× stronger than the best non-judge baseline.
- Ran ablation studies across two independent judge models (GPT-4o, Qwen-3.6 Flash) and culturally-flattened paraphrase conditions, demonstrating ranking robustness across all configurations.
- 设计了无参考译文的机器翻译评估框架，将大语言模型评判器锚定于含 872 条文化特有项的人工词表，将翻译错误归为五类，并生成确定性的 0–100 分评分，具备逐位可复现性。
- 基于古汉语源文本，通过大语言模型辅助抽取并经专家对照权威学术译本校验，构建文化特有项（CSI）词表，覆盖宗教、社会、物质、生态与语言五类范畴。
- 在约 6,400 个句段上评估四套机器翻译系统（GLM-5.1、DeepSeek-V4 Flash、Llama-3、Google 翻译），完全复现人工文化忠实度排名，Spearman 相关性较最佳非评判器基线强 2.7 倍。
- 在两个独立评判模型（GPT-4o、Qwen-3.6 Flash）及文化扁平化改写条件下开展消融实验，验证排名在所有配置下的鲁棒性。
Emotion Aid - Emotion Speech Recognition for Disordered Speech
Emotion Aid：针对障碍语音的情感语音识别
Undergraduate · Completed
本科 · 已完成
Aug - Dec 20232023年8月 - 12月
- Built a CNN-LSTM pipeline for multi-class emotion recognition from disordered clinical speech, extracting MFCC, prosodic, and spectral features with Librosa and addressing class imbalance via pitch shifting, time stretching, and noise injection.
- Automated the end-to-end evaluation pipeline (feature extraction, inference, multi-class reporting) and conducted systematic failure analysis across emotion categories to guide targeted augmentation.
- 构建 CNN-LSTM 流水线，用于从障碍临床语音中进行多类别情感识别，使用 Librosa 提取 MFCC、韵律与频谱特征，并通过音调变换、时间拉伸与噪声注入处理类别不平衡问题。
- 自动化端到端评估流水线（特征提取、推理、多类别报告），并对各情感类别进行系统性错误分析，以指导针对性数据增强。
§ 07
Publications
论文发表
Papers 论文C = Conference · J = Journal · S = In Submission · T = Thesis · * = equal contribution
C = 会议 · J = 期刊 · S = 投稿中 · T = 学位论文 · * = 共同一作
-
C.1
Accepted · Findings of EMNLP 2026
已接收 · EMNLP 2026 Findings
IndicDetect: Evaluating Cross-Lingual LLM-Generated Text Detection for Hindi, Telugu, and TamilIndicDetect：面向印地语、泰卢固语与泰米尔语的跨语言大语言模型生成文本检测评估To appear in the Findings of the Association for Computational Linguistics: EMNLP 2026.将发表于计算语言学协会研究成果：EMNLP 2026（Findings of EMNLP 2026）。
-
S.1
In Submission · ACL ARR 2026
投稿中 · ACL ARR 2026
CultMetric: Gloss-Grounded Cultural Fidelity Evaluation for Machine TranslationCultMetric：基于词条释义的机器翻译文化忠实度评估Manuscript submitted to ACL Rolling Review (ARR).论文已投稿至 ACL Rolling Review (ARR)。
§ 08
Technical Skills
技术技能
Competencies 专业能力Programming Languages
编程语言
Python, C/C++, Java, SQL, R, Bash
Python、C/C++、Java、SQL、R、Bash
Deep Learning & LLM Frameworks
深度学习与大语言模型框架
PyTorch, TensorFlow / Keras, Hugging Face Transformers, QLoRA, Fairseq, OpenNMT; Transformer architectures (BERT, RoBERTa, XLM-RoBERTa, GPT-style); BPE, SentencePiece.
PyTorch、TensorFlow / Keras、Hugging Face Transformers、QLoRA、Fairseq、OpenNMT；Transformer 架构（BERT、RoBERTa、XLM-RoBERTa、GPT 系列）；BPE、SentencePiece。
NLP & Speech
自然语言处理与语音
NLTK, SpaCy, SpeechBrain, ESPnet, Librosa; MFCC, prosodic, and spectral feature extraction; CNN-LSTM classifiers; cross-lingual benchmarking; low-resource NLP.
NLTK、SpaCy、SpeechBrain、ESPnet、Librosa；MFCC、韵律及频谱特征提取；CNN-LSTM 分类器；跨语言基准测试；低资源自然语言处理。
Evaluation Metrics
评估指标
AUROC, F1, BLEU, ROUGE, BERTScore, WER.
AUROC、F1、BLEU、ROUGE、BERTScore、WER。
MLOps & Reproducibility
MLOps 与可复现性
MLflow, Weights & Biases, Docker, Git, CI pipeline integration; pytest and shell-based test automation.
MLflow、Weights & Biases、Docker、Git、CI 流水线集成；pytest 与基于 shell 的测试自动化。
Data & Analysis
数据与分析
pandas, NumPy, SciPy, scikit-learn, Matplotlib, Seaborn.
pandas、NumPy、SciPy、scikit-learn、Matplotlib、Seaborn。
Developer Tools
开发工具
Linux / HPC cluster environments, Google Cloud Platform, LaTeX, Jupyter, VS Code.
Linux / 高性能计算集群环境、Google Cloud Platform、LaTeX、Jupyter、VS Code。
Languages
语言
English (Fluent) · Telugu (Native) · Hindi (Fluent) · Mandarin (Conversational).
英语（流利）· 泰卢固语（母语）· 印地语（流利）· 普通话（日常交流）。
§ 09
Relevant Coursework
相关课程
Curriculum 课程内容
University of Macau
澳门大学
Graduate
研究生
Advanced Natural Language Processing.
高级自然语言处理。
University of Macau
澳门大学
Undergraduate Exchange
本科交流
Natural Language Processing · Digital Image Processing.
自然语言处理 · 数字图像处理。
§ 10
Academic Service
学术服务
Reviewing 审稿工作Student Reviewer
学生审稿人
IEEE/ACM Transactions on Audio, Speech, and Language Processing (TASLP) - reviewed three manuscripts under the supervision of Prof. Suryakanth V. Gangashetty.
2023 - 2024
IEEE/ACM Transactions on Audio, Speech, and Language Processing (TASLP)：在 Suryakanth V. Gangashetty 教授指导下审阅了三篇论文。
2023 - 2024