Exploring the intersection of communication, AI, and hardware.
nothing but the best.
About
Hi, I’m Wang Zijun, an undergraduate in Communication Engineering at Northeastern University at Qinhuangdao. I am drawn to one question: how can we make systems more intelligent and dependable?
My interests span three connected areas: large language models and multi-agent systems, multimodal learning and graph-based recommendation, and communication and coding theory. My co-authored work includes bureaucratic-bias detection, multimodal fake-news detection, and uncertainty-aware and cold-start recommendation, with three papers accepted to WISE 2026 and one to EMNLP 2026. At the IHUB Research Program, I also studied fault tolerance and computational efficiency in EVENODD coding.
Beyond research, I have worked on FPC/PCB hardware engineering and embedded development in industry. I value research that is theoretically sound and grounded in real systems. This site is where I collect ideas, document what I build, and connect with others.
— Open to research collaborations and internships · Qinhuangdao, China
News
A timeline of research, competitions, and industry experience.
Publications
Four accepted conference papers: three at WISE 2026 and one at EMNLP 2026, alongside earlier work on reliable computing.
Projects
Four collaborative research efforts in multi-agent governance, multimodal learning, and recommendation.
Open-set standpoint screening and multi-agent governance for bureaucratic-bias detection in government text
A closed-loop multi-agent governance workflow: legal and policy evidence retrieval, open-set target screening, a dual-jury mechanism, risk routing, and verdict rewriting for verification.
Uncertainty-aware hierarchical intent learning via flow matching for multimodal recommendation
Uses uncertainty estimates to adapt how broadly or precisely a recommender represents a user’s intent.
An interaction-aware mixture-of-experts framework for multimodal fake-news detection
Explicitly decomposes cross-modal interactions into five experts, using sample-level dynamic weights for fine-grained and interpretable decisions.
Motivation-guided topology inference for cold-start multimodal recommendation
Uses offline LLM reasoning about user motivations to reconnect cold-start items through more useful graph structure.
Experience
Beyond research, I build engineering intuition through hardware and embedded-systems work.
Worked across the flexible printed circuit (FPC) engineering lifecycle, learning design standards, process constraints, and quality-control requirements for chip flex boards. Tracked R&D samples and documented process issues across development stages.
Developed hardware-facing embedded modules and device drivers in C. Created schematics and production-ready layouts for multiple PCBs in Altium Designer, and contributed to board testing and design iteration.
Supported software optimisation for industrial systems and tools; collected and organised data in designated test scenarios for downstream analysis.
Awards
Provincial-level competition awards across communications, market analysis, and computer design.
Toolbox
A toolkit that spans model training, embedded work, and circuit design.
Writing
I write about multi-agent systems, multimodal learning, and the trade-offs behind reliable computing. View all writing →
How MARS-Gov uses a “tenth juror” to identify overlooked groups beyond predefined categories.
Read article →How IMEX-FND uses five experts to make cross-modal interactions more interpretable.
Read article →Why can two disks of redundancy deliver strong fault tolerance through XOR operations—and what is the cost?
Read article →Notes
Not always about research—just small observations on learning and life.
The longer I do research, the more I think the hardest part is not writing code. It is forcing a vague intuition into a falsifiable question. Everything after that is engineering.
# Research practiceSeeing FPC production firsthand made the idea of yield tangible. A seemingly minor fault-tolerance assumption in a paper becomes a real cost on a factory floor. Theory and engineering need each other.
# Internship noteGiving a multi-agent system a role that can disagree is often more useful than stacking ten more prompts. Diversity is not noise; it is another form of information.
# IdeaWhen two paper deadlines collided, what carried me through was not talent but a detailed to-do list and the refusal to give up. Onward.
# Daily noteLinks
People and sites that have inspired me.
To exchange links, send me your site name · short description · URL by email.
“Persistence is what makes a mind adaptable.”— A reminder to myself
For research collaborations, internship opportunities, or a conversation about multi-agent and multimodal systems, feel free to reach out.