Email 
Scholar 
GitHub 
Hugging Face LLM code evaluation
Adversarial tests, robust verifiers, and evaluation that reflects real programming ability.
Seeking 2027 opportunities
Preferred base · Xi’an
Expected graduation · Jun 2027
Zihan Ma · 马梓涵
I build AI systems that deserve to be trusted.
PhD researcher working across reliable code generation, agent safety, and multimodal misinformation — from better evaluation to more robust models.
PhD Student Xi’an Jiaotong University
Visiting PhD Nanyang Technological University
02
Research compass
My research focuses on reliable AI evaluation, agent safety, and multimodal misinformation.
I am a PhD student at Xi’an Jiaotong University, advised by Prof. Minnan Luo. Our research group is part of the broader team led by Prof. Qinghua Zheng, an Academician of the Chinese Academy of Engineering.
My work examines where modern AI systems fail — weak code verification, concealed agent risk, and misleading multimodal content — then builds evaluation and learning methods that make those failures visible and actionable.
I am currently a Visiting PhD Student at NTU with Asst. Prof. Wenya Wang, and collaborate with Zhuoyi Lin at the A*STAR Institute of Advanced Intelligence and Computing (IAIC).
From December 2024 to October 2025, I was a Research Intern with OpenCompass at Shanghai Artificial Intelligence Laboratory, working with Songyang Zhang on large-model evaluation and reliable code-generation benchmarks.
Agent behavior
Benchmarks and sandboxes that expose hidden intent, complex workflows, and brittle safeguards.
Multimodal misinformation
Causal, event-aware, and manipulation-sensitive models for understanding deceptive content.
03
News & updates
What I’ve been up to lately.
-
Conference acceptances
🎉 I’m delighted that our PRIMA paper was accepted to EMNLP 2026 Findings.
I’m the first author on PRIMA, and I’m a co-first author on our other accepted paper, “CompassJudger-2.”
-
Honor
🏅 I’m honored to be selected for the CAST Young Talent Cultivation Program.
中国科协青年科技人才培育工程博士生专项计划。
-
Current position
🇸🇬 I started a new chapter as a Visiting PhD Student at NTU.
I’m working on NLP and AI safety in Singapore with Asst. Prof. Wenya Wang.
-
New preprint
🛡️ We released OASIS, our new benchmark for agent safety.
Our paper “How Brittle is Agent Safety?” is now available on arXiv with code.
-
Journal acceptance
📄 We’re glad that our DICE paper was accepted by IEEE TIFS.
In “Graphing the Truth,” we study causal multimodal fake-news detection.
-
Service
🙌 I was named an Outstanding Reviewer at ACM Multimedia 2025.
I’m grateful for the recognition and the opportunity to contribute to the community.
-
Conference acceptance
🧪 We’re excited that our SAGA paper was accepted by NeurIPS 2025.
We rethink verification for LLM code generation through stronger testing.
-
Conference acceptance
🔎 We’re happy that our PRISM paper was accepted by SIGIR 2025.
We explore personalized recommendation that balances user interests with truthfulness.
04
Selected research
First & co-first author work.
Workflow / PRIMA
Beyond the Wrong Bit: Failure States and the Supervision Unit in Verifiable Reasoning
TL;DR PRIMA separates repairable failures into inheritable prefixes and divergent suffixes, turning coarse pass-or-fail feedback into localized supervision for verifiable reasoning.
Paper forthcoming
Pipeline / CompassJudger-2
CompassJudger-2: A Holistic Approach Towards Generalist Judge Model
TL;DR CompassJudger-2 unifies objective verification, subjective evaluation, and rubric refinement through task-specific data construction, producing a more robust generalist judge model.
Framework / DICE
Graphing the Truth: Harnessing Causal Insights for Advanced Multimodal Fake News Detection
TL;DR DICE learns to separate causal multimodal signals from spurious bias with graph masks, improving robustness under domain and distribution shifts.
Workflow / OASIS
How Brittle is Agent Safety? Rethinking Agent Risk under Intent Concealment and Task Complexity
TL;DR A two-axis benchmark and stateful sandbox reveal that concealed intent erodes safety, while capability limits can make harder tasks look deceptively safe.
Pipeline / SAGA
Rethinking Verification for LLM Code Generation: From Generation to Testing
TL;DR SAGA combines correct programs and human bugs to synthesize diverse adversarial tests that catch failures missed by standard code benchmarks.
Framework / PRISM
Bridging Interests and Truth: Towards Mitigating Fake News with Personalized and Truthful Recommendations
TL;DR PRISM uses veracity-aware representations and fake-news signals as negative diffusion prompts to balance user interests with truthful recommendations.
Framework / Code Triangle
Coding Triangle: How Does Large Language Model Understand Code?
TL;DR Editorials, code, and tests expose LLM self-consistency — and the systematic gap between model cognition and the diversity of human programming.
Pipeline / MGCA
Each Fake News Is Fake in Its Own Way: An Attribution Multi-Granularity Benchmark
TL;DR AMG moves beyond binary labels by identifying how multimodal news is fake, enabling fine-grained detection and attribution with aligned clues.
Pipeline / Event-Radar
Event-Radar: Event-driven Multi-View Learning for Multimodal Fake News Detection
TL;DR Event-level graphs fuse consistency, emotion, manipulation, and view credibility to detect fake news robustly in noisy social-media data.
Pipeline / MHFFD
Learning Multimodal Attention Mixed with Frequency Domain Information as Detector for Fake News Detection
TL;DR Token-level image-text alignment guides high-frequency manipulation cues, exposing deceptive news that remains semantically consistent.
* Equal contribution. Featured cards above highlight first- and co-first-author work; the complete collaborative record follows.
Extended publication record
Collaborative work.
Team science across foundation-model evaluation, multimodal reasoning, social networks, and trustworthy information systems. Citation counts update from Google Scholar.
TwiBot-22: Towards Graph-Based Twitter Bot Detection
Zihan Ma as a contributor in the TwiBot-22 team
Intern-S1: A Scientific Multimodal Foundation Model
Zihan Ma as a contributor in the Intern-S1 team
Mitigating World Biases: A Multimodal Multi-View Debiasing Framework for Fake News Video Detection
Zhi Zeng, Minnan Luo, Xiangzheng Kong, Huan Liu, Hao Guo, Hao Yang, Zihan Ma, Xiang Zhao
IMOL: Incomplete-Modality-Tolerant Learning for Multi-Domain Fake News Video Detection
Zhi Zeng, Jiaying Wu, Minnan Luo, Herun Wan, Xiangzheng Kong, Zihan Ma, Guang Dai, Qinghua Zheng
How Do Social Bots Participate in Misinformation Spread? A Comprehensive Dataset and Analysis
Herun Wan, Minnan Luo, Zihan Ma, Guang Dai, Xiang Zhao
SAGE: Scale-Aware Gradual Evolution for Continual Knowledge Graph Embedding
Yifei Li, Lingling Zhang, Hang Yan, Tianzhe Zhao, Zihan Ma, Muye Huang, Jun Liu
Understand, Refine and Summarize: Multi-View Knowledge Progressive Enhancement Learning for Fake News Video Detection
Zhi Zeng, Jiaying Wu, Minnan Luo, Xiangzheng Kong, Zihan Ma, Guang Dai, Qinghua Zheng
ATLAS: A High-Difficulty, Multidisciplinary Benchmark for Frontier Scientific Reasoning
Zihan Ma as a project contributor in the ATLAS team
Harmony in Chaos: A Progressive Noise-Resilient Network for Robust Fake News Video Detection
Xiangzheng Kong, Zhi Zeng, Chenguang Zhu, Zihan Ma, Minnan Luo
From Manipulation to Mistrust: Explaining Diverse Micro-Video Misinformation for Robust Debunking in the Wild
Zhi Zeng, Yiming Yang, Jiaying Wu, Xinyu Zhang, Xiangzheng Kong, Herun Wan, Zihan Ma, Minnan Luo
Deciphering Trajectory-Aided LLM Reasoning: An Optimization Perspective
Junnan Liu, Hongwei Liu, Linchen Xiao, Shudong Liu, Taolin Zhang, Zihan Ma, Songyang Zhang, Kai Chen
DiFaR: Enhancing Multimodal Misinformation Detection with Diverse, Factual, and Relevant Rationales
Herun Wan, Jiaying Wu, Minnan Luo, Xiangzheng Kong, Zihan Ma, Zhi Zeng
05
Journey
Education and research experience.
I started in electrical engineering and later moved into multimodal learning, large-model evaluation, and agent safety through graduate study, internships, and international collaboration.
2025.11Present
International research
Visiting PhD Student
Nanyang Technological University · Singapore
Researching natural language processing and AI safety with Asst. Prof. Wenya Wang, alongside collaboration with A*STAR IAIC.
2024.122025.10
Industry research
Research Intern · OpenCompass
Shanghai Artificial Intelligence Laboratory
Worked on large-model evaluation and reliable code benchmarking with Songyang Zhang and the OpenCompass team.
20242027
Doctoral study
PhD in Computer Science & Technology
Xi’an Jiaotong University
Advised by Prof. Minnan Luo, focusing on trustworthy multimodal intelligence, LLM evaluation, and agent safety.
20222024
Graduate study
Master–PhD Program · Direct Entry
Xi’an Jiaotong University
Began research in multimodal misinformation analysis and trustworthy social-media intelligence.
20182022
Undergraduate study
BEng in Electrical Engineering & Automation
Xi’an Jiaotong University
Built the engineering foundation that continues to shape my approach to dependable AI systems.
06
Recognition
Recognition & academic service.
Selected honor
China Association for Science and Technology Young Talent Cultivation Program
Doctoral Student Special Plan · 中国科协青年科技人才培育工程博士生专项计划
Outstanding Reviewer
ACM Multimedia 2025
CSC Visiting PhD Program
Nanyang Technological University
Reviewer
NeurIPS · ICLR · AAAI · ARR · IEEE TIFS · ACM MM · IEEE ICME
Selected support
XJTU Graduate Special Scholarship · Siyuan Electric Scholarship
Contact
Open to research collaborations and 2027 opportunities.
Currently based between Singapore and Xi’an, with Xi’an as the preferred long-term base.