Fangyanuo Zhou – Research

Publications

Patents

Research Experience

Focus: Interdisciplinary research at the intersection of news communication and large models, including LLM value alignment, multimodal fake news detection, multi-task fake news detection models, agent-based fake news dissemination simulation, sentiment analysis, hallucination research, and text readability analysis. It also includes recommender-system-related fake news studies, such as text-level comparisons between generative recommender systems and traditional recommender systems, and the propagation effects of fake news within recommender systems.

Responsibilities:

• Core work: Built an agent-based simulation for the news dissemination process; designed intervention and ablation experiments; implemented a multi-task fake news detection model.
• Supporting work: Investigated differences in value alignment between generative recommender systems and traditional recommender systems; studied whether fake news generated by large models, after passing through recommender systems, may lead to truth decay; used Python for web crawling on domestic and international news websites; leveraged large models for batch summarization and value system analysis; ran batch experiments via external APIs; explored diverse neural network frameworks and popular deep learning architectures; contributed to research background and related knowledge sections in papers.
• Competition: Participated in a social simulation competition, contributing to prompt design and cloud-based distributed deployment.

Achievements: Co-authored a multimodal MoE fake news detection paper accepted by IEEE TCE (doi: 10.1109/TCE.2026.3677445); co-authored “Towards Better Transparency and Reliability in Smart Contract Vulnerability Detection via Counterfactual Contrastive Explanations” (under review) and a multi-task fake news detection paper (in progress), plus a tourism trajectory simulation and visitor-flow prediction study for travel routes (in progress); open-source outputs: WeSpeak, Reptile; patent pending (expected approval in 2026.9 – 2026.10).

Introduction: Developing a multimodal AI psychological coaching system for college students, with a rigorous randomized controlled trial to evaluate its causal intervention effect on students’ mental health. The project integrates large language models, multimodal perception technology, and evidence-based psychological intervention methods.

Responsibilities:

• Developed a Linux-deployed AI coaching system integrating multiple RAG knowledge bases and MCP-based tool interfaces.
• Contributed to the randomized controlled trial design evaluating the causal effects of AI-assisted interventions on university students’ mental health.
• Contributed to system implementation and manuscript preparation.

Introduction: Built generative AI to produce structured medical records, reducing repetitive work, delays, and quality inconsistency. Focused on medical dialogue modeling to enable automatic outpatient record generation.

Responsibilities:

• Early-stage speech-to-text (Whisper), Qwen fine-tuning, and prompt engineering. As a core developer, iterated on models to improve the standardization and practicality of generated medical records.
• Batch data testing and effect tuning by building automated test pipelines, significantly improving testing efficiency.
• Assisted environment deployment, local proxy setup, and firewall access.

Achievements: The project later shifted to mental health departments with potential hospital adoption in Guangzhou.

Introduction: Built a visitor behavior simulation platform using Python and Vue with official platform data plus web-crawled data, exploring social media impact on travel route choices.

Responsibilities: Implemented time filtering and improved data visualization modules; delivered key functions such as multi-day repeated population addition and temporary scenic spot closure.

Achievements: Expected to publish at least one paper (target submission before 2026.12).

Introduction: Applied computer vision to evaluate livestock breeding effects; built a custom data annotation platform; trained YOLO models for target weight recognition.

Responsibilities: Researched and implemented keyframe recognition algorithms; used YOLOn11pose to address livestock body keypoint recognition challenges.

Achievements: Data annotation platform delivered for enterprise acceptance in 2025.12; algorithms are still iterating.

Projects

• RecRedTeam — red-team audit framework for LLM/agentic recommender systems. GitHub
• Crusaders — human–machine collaboration framework. GitHub
• Open Stethoscope — heart sound AI murmur detection for grassroots healthcare. GitHub
• OnChainGov — on-chain DAO governance research toolchain with causal inference. GitHub
• WeSpeak — open-source output of the news communication and large models research. GitHub
• Reptile — open-source output of the news communication and large models research. GitHub
• DLFaceDetection — RPC / remote function call sample. GitHub