Chengwei Wei魏承威
AI Scientist · Singapore
My focus is on natural language processing, and increasingly on what lies beyond it: how large language models reason, how efficiently they do it, and whether that reasoning survives when the input is no longer text.
Before A*STAR, I did my Ph.D. at the University of Southern California with Prof. C.-C. Jay Kuo, and interned at Amazon working on LLMs for e-commerce with Karim Bouyarmane.
Research
My work spans the pipeline, from what a model is trained on, to how it decodes, to how we find out where it breaks.
LLM reasoning
Two questions here: one about ability, one about cost. CoinMath asks what coding data teaches mathematical reasoning. The style of code-based rationales, not just their correctness, shapes what a model learns, and diversifying those styles beats adding more general-domain code. InfoDensity asks how briefly a model can reason and still be right, rewarding information-dense traces so it reaches the answer in fewer tokens instead of padding its way there.
Decoding and structured generation
CABS estimates confidence at the sub-structure level rather than per token, then uses it to steer beam search and refine prompts. That cuts hallucination in structured data generation, where an entry can be perfectly well-formed and still wrong.
Evaluation and robustness
Spoken-MQA and IFEval-Audio ask whether mathematical reasoning and instruction-following survive in audio LLMs, rather than testing transcription alone. Resilience measures how far models degrade when instructions arrive with ASR slips, OCR errors, typos, or distracting content.
Selected work
InfoDensity
EMNLP 2026Rewarding information-dense reasoning traces, so models reason efficiently instead of padding.
CoinMath
ACL 2025 FindingsHow coding style in code-based rationales shapes math reasoning, and why diversifying that style beats adding more code.
Spoken-MQA
Benchmark · 2025Can speech-based models do math? A benchmark for multi-faceted mathematical reasoning delivered as audio.
CABS
Amazon · 2024Confidence-aware sub-structure beam search that cuts hallucination when LLMs generate structured data.
IFEval-Audio
AACL 2025Can audio LLMs follow an instruction they hear? A benchmark built with my mentee Yiming Gao.
Resilience
EMNLP 2024 FindingsHow far LLMs degrade when instructions arrive with ASR slips, OCR errors, typos, or distracting content.
Publications
Selected papers
The complete and most up-to-date list is on Google Scholar.
LLMs & Reasoning
InfoDensity: Rewarding Information-Dense Traces for Efficient Reasoning
Chengwei Wei, Jung-jae Kim, Longyin Zhang, Shengkai Chen, Nancy F. Chen
EMNLP 2026 paper
CoinMath: Harnessing the Power of Coding Instruction for Math LLMs
Chengwei Wei, Bin Wang, Jung-jae Kim, Guimei Liu, Nancy F. Chen
ACL 2025 Findings paper code + data + model
Confidence-Aware Sub-Structure Beam Search (CABS): Mitigating Hallucination in Structured Data Generation with Large Language Models
Chengwei Wei, Kee Kiat Koo, Amir Tavanaei, Karim Bouyarmane
Resilience of Large Language Models for Noisy Instructions
Bin Wang, Chengwei Wei, Zhengyuan Liu, Geyu Lin, Nancy F. Chen
EMNLP 2024 Findings paper
CRAFT: Extracting and Tuning Cultural Instructions from the Wild
Bin Wang, Geyu Lin, Zhengyuan Liu, Chengwei Wei, Nancy F. Chen
Bias and Fairness in Chatbots: An Overview
Jintang Xue, Yun-Cheng Wang, Chengwei Wei, Xiaofeng Liu, Jonghye Woo, C.-C. Jay Kuo
APSIPA Trans. SIP, 2024 paper
An Overview on Generative AI at Scale with Edge-Cloud Computing
Yun-Cheng Wang, Jintang Xue, Chengwei Wei, C.-C. Jay Kuo
IEEE OJ-COMS, 2023 paper
An Overview of Language Models: Recent Developments and Outlook
Chengwei Wei, Yun-Cheng Wang, Bin Wang, C.-C. Jay Kuo
APSIPA Trans. SIP, 2023 paper
Speech & Audio
Advancing Singlish Understanding: Bridging the Gap with Datasets and Multimodal Models
Bin Wang, Xunlong Zou, Shuo Sun, Wenyu Zhang, Yingxu He, Zhuohan Liu, Chengwei Wei, Nancy F. Chen, AiTi Aw
ASRU 2025 paper code + data + model
IFEval-Audio: Benchmarking Instruction-Following Capability in Audio-based Large Language Models
Yiming Gao, Bin Wang, Chengwei Wei, Shuo Sun, AiTi Aw
AACL 2025 paper code + data
Towards Spoken Mathematical Reasoning: Benchmarking Speech-based Models over Multi-faceted Math Problems
Chengwei Wei, Bin Wang, Jung-jae Kim, Nancy F. Chen
arXiv:2505.15000, 2025 paper code + data
A Green Learning Approach to Spoofed Speech Detection
Chengwei Wei, Runqi Pang, C.-C. Jay Kuo
ICASSP 2024 paper
Fundamental NLP
Word Embedding Dimension Reduction via Weakly-Supervised Feature Selection
Jintang Xue, Yun-Cheng Wang, Chengwei Wei, C.-C. Jay Kuo, et al.
APSIPA Trans. SIP, 2024 paper
SynWMD: Syntax-aware Word Mover's Distance for Sentence Similarity Evaluation
Chengwei Wei, Bin Wang, C.-C. Jay Kuo
Task-specific Dependency-based Word Embedding Methods
Chengwei Wei, Bin Wang, C.-C. Jay Kuo
Pattern Recognition Letters, 2022 paper
Vision & Multimodal
Crowdsource, Crawl, or Generate? Creating SEA-VL, a Multicultural Vision-Language Dataset for Southeast Asia
Samuel Cahyawijaya, Holy Lovenia, …, Chengwei Wei, et al.
ACL 2025 paper
Efficient Human-Object-Interaction (EHOI) Detection via Interaction Label Coding and Conditional Decision
Tsung-Shan Yang, Yun-Cheng Wang, Chengwei Wei, Suya You, C.-C. Jay Kuo
CVIU, 2024 paper
GHOI: A Green Human-Object-Interaction Detector
Tsung-Shan Yang, Yun-Cheng Wang, Chengwei Wei, C.-C. Jay Kuo
IEEE MIPR 2024 paper
ExpressionHop: A Lightweight Human Facial Expression Classifier
Chengwei Wei, C.-C. Jay Kuo, Rafael Luiz Testa, Ariane Machado-Lima, Fátima L. S. Nunes
IEEE MIPR 2022 paper
Mentees
I have been fortunate to work with these students
Yiming Gao
Undergraduate Research Intern
NTU, Singapore
Jan 2025 – May 2025
Multimodal LLMs · co-advised with Bin Wang and Shuo Sun · AACL 2025
Pham The Binh Minh
Undergraduate Research Intern
NTU, Singapore
Jan 2025 – May 2025
Multimodal LLMs · co-advised with Bin Wang and Shuo Sun
Runqi Pang
Graduate Research Intern
University of Southern California
May 2023 – Aug 2023
NLP and speech processing · ICASSP 2024
Teaching