Weihua Zheng
I am a Senior Research Engineer at A*STAR, Singapore. I am also a part-time PhD student supervised by Prof. Roy Ka-Wei Lee. My research focuses on multilingual large language models, cultural and multimodal evaluation, machine translation, question answering, and agentic AI systems.
My work spans both research and applied NLP system development. I have contributed to real-world multilingual translation systems, operational text understanding systems, and recent research on multilingual LLMs, cultural evaluation, multimodal reasoning, LLM safety, and agentic AI evaluation.
I am particularly interested in building reliable and culturally-aware AI systems for multilingual and multimodal scenarios. My recent work studies how large language models reason across languages, cultures, and modalities, with an emphasis on evaluation, benchmark construction, and practical NLP applications.
News
- NEWS May 2026: SemEval-2026 Task 7, which we are co-organizing, has been officially launched. We warmly welcome your participation! [task page]
- NEWS April 2026: Our work MMAC: A Multilingual, Multimodal Alignment Framework for Cultural Grounding Evaluation was accepted to ACL 2026. Many thanks to our collaborators from eight different countries for their support!
- NEWS December 2025: Our work BLEnD-Vis: Benchmarking Multimodal Cultural Understanding in Vision Language Models was accepted to EACL 2026.
- NEWS November 2025: Our work AdaMCoT: Rethinking Cross-Lingual Factual Reasoning Through Adaptive Multilingual Chain-of-Thought was accepted to AAAI 2026 as an Oral Paper.
- NEWS August 2025: Our work CCL-XCoT: An Efficient Cross-Lingual Knowledge Transfer Method for Mitigating Hallucination Generation was accepted to EMNLP 2025.
- NEWS 2023–2024: Built and deployed social media translation engines for Malay-to-English, Chinese-to-English, and Tamil-to-English under the Online Trust and Safety Research Programme.
- NEWS 2019–Present: Served as technical lead for Singapore Translate Together, developing and maintaining the translation models and backend application.
Research Interests
Multilingual LLM Cultural LLM Multimodal LLM AI Agent Machine Translation Cross-Lingual Reasoning LLM Evaluation Question Answering LLM Safety
- Multilingual and cross-lingual large language models
- Culturally-grounded evaluation of LLMs and vision-language models
- Multimodal benchmark construction across text, image, and speech
- Cross-lingual reasoning and hallucination mitigation
- Agentic AI systems and process-level evaluation
- Machine translation, social media translation, and applied NLP systems
Selected Publications
For the complete and up-to-date publication list, please see my Google Scholar profile.
Selected Projects
SG Translate Together (SGTT) is Singapore’s public translation web portal powered by the government’s SG Translate neural machine translation engine. It provides bi-directional, Singapore-context translations across the four official languages: English, Chinese, Malay, and Tamil.
The platform was developed jointly by MDDI, formerly MCI, and A*STAR’s I²R, with support from the TRANS Grant. SGTT also aggregates translation resources and community initiatives for Singapore’s multilingual society.
I served as the technical lead for this project, responsible for developing the translation models and the backend application. I continue to maintain the entire translation engine stack.
Machine Translation Large Language Models Backend Development Model Deployment Multilingual NLP
I was involved in a sub-project under the Research Pillar of Singapore’s National Multimodal LLM Programme. In this project, we formed an academic collaboration with Korea University to advance research on large language model safety.
We introduced a novel and efficient approach for mitigating hallucinations in multilingual LLMs by promoting knowledge transfer between high-resource and low-resource languages. This work has been accepted to EMNLP 2025.
Large Language Models LLM Safety Multilingual LLM Hallucination Mitigation Python PyTorch
The Online Trust and Safety (OTS) Research Programme is a funding initiative from the Ministry of Digital Development and Information (MDDI). It is part of the Smart Nation and Digital Economy domain of the Research, Innovation and Enterprise 2025 programme administered by the National Research Foundation.
The programme aims to develop technological capabilities to combat online harms and align translation efforts with high-impact use cases.
I served as the developer for the Social Media Translation Engines in this project, building and deploying Malay-to-English, Chinese-to-English, and Tamil-to-English translation engines specialized for social media content.
Machine Translation Social Media Translation Malay-to-English Chinese-to-English Tamil-to-English Python
I work on multilingual, multimodal, and culturally-grounded evaluation for large language models and vision-language models. My recent research studies how LLMs reason across languages, cultures, and modalities, with a focus on benchmark construction, cultural alignment, and robustness evaluation.
Cultural LLM Multimodal LLM Benchmark Construction LLM Evaluation Cross-Lingual Reasoning
I study multi-stage agentic systems, including process-level reliability, error localization, tool-use evaluation, and reasoning trace analysis. My interest is in understanding not only whether an AI system produces the correct final answer, but also where and why errors occur during the reasoning or tool-use process.
AI Agents Process Evaluation Error Localization Tool Use LLM Evaluation
Research Experience
-
Senior Research Engineer, A*STAR, Singapore
2019 – Present -
Part-Time PhD Student, Singapore University of Technology and Design
2024 – Present
Education
-
PhD, Singapore University of Technology and Design
2024 – Present -
Master's Degree, Nanyang Technological University, Singapore
2018 – 2019 -
Bachelor's Degree, Wuhan University, China
2014 – 2018
Academic Service
- Reviewer for NLP and AI conferences.
- Research and engineering support for multilingual NLP, machine translation, and LLM projects.