Computing’s Top 30: Quanming Yao

Quanming Yao is one of our "Computing's Top 30 Early Career Professionals" for 2025. This program seeks to highlight an esteemed group of rising stars who earned this honor for their exceptional early-career achievements and role in driving advancements across the computing landscape. 

Introduction

My name is Quanming Yao. I am an Associate Professor in the Department of Electronic Engineering at Tsinghua University. Our department is one of Tsinghua’s core engineering units, where foundational research meets real-world system innovation across areas such as communications, hardware, and AI-driven computing.

My work focuses on building machine learning systems that are not only powerful, but also data-efficient and reliable — especially for structured data such as graphs, knowledge networks, tabular systems, and biomedical data. Before returning to academia, I worked in industry at 4Paradigm, an AI unicorn, where I helped build and lead research teams translating machine learning ideas into large-scale production systems. That experience shaped how I view research: theoretical elegance matters, but robustness under real constraints matters just as much.

People sometimes describe me as a “hardcore machine learning researcher.” I take that as a compliment. I enjoy diving deep into learning theory, optimization, and representation learning — while always asking how these ideas can make an impact in industry and scientific discovery.

What inspired you to pursue a career in technology?

What drew me to technology was the realization that computation turns abstract ideas into working systems. I still remember the first time I saw a model trained on data generalize correctly to unseen examples. It felt like building a reusable reasoning engine.

Technology also offers a rare bridge between theory and impact. In machine learning, you can begin with a mathematical formulation, design an algorithm, and eventually see it deployed in systems that influence scientific research, medical discovery, or industrial decision-making. That combination of intellectual depth and tangible impact has always motivated me.

Machine learning in particular sits at a fascinating intersection — statistics, optimization, computer science, and increasingly, domain sciences like biology and chemistry. I enjoy working in a field where progress depends equally on clarity of thinking and disciplined engineering.

Who do you draw inspiration from and how did that motivate you in your education or career?

I have been fortunate to learn from two influential mentors: Prof. James Kwok and Prof. Qiang Yang. Those two influences — rigor and impact — continue to guide my work today.

During my Ph.D., Prof. James Kwok trained me to think like a method-driven machine learning researcher. My doctoral work focused on optimization algorithms for sparse and low-rank modeling. That experience strengthened my appreciation for clean formulations, convergence guarantees, and principled algorithm design. From him, I learned that strong results should be explainable — not just impressive.

Later, at 4Paradigm, I worked closely with Prof. Qiang Yang, one of the company’s co-founders. He encouraged me to connect rigorous methods with real-world needs. Together, we worked on AutoCross (KDD 2019), an automatic feature-crossing system deployed in large-scale industrial tabular learning systems. The philosophy was simple: start from a real production bottleneck, and design a solution that is both theoretically grounded and practically deployable.

Is there any emerging technology or industry segment you find exciting or interesting?

One emerging direction I find especially compelling is scalable agentic learning with limited supervision.As we move from training language models on massive curated datasets to deploying autonomous agents in the real world, supervision becomes scarce. Feedback is delayed, outcomes may only become clear after long cycles, and humans cannot label every intermediate decision. In this setting, intelligence must come not only from data scale, but from interaction, adaptation, and internal consistency.

This shift is particularly important in AI for Science and industry. In drug discovery and biomedical research, labeled data is expensive and often mechanism-dependent. Systems must reason across molecular structures, biological networks, and scientific literature, while refining hypotheses under limited feedback. In industrial environments, such as recommender systems or automated decision platforms, feedback is noisy and constantly evolving. Models must adapt under distribution shift while remaining stable and reliable.

In my research experience, which spans earlier industry work to current academic efforts, we explore this direction through structured learning systems for large-scale tabular data, multi-agent reasoning frameworks for scientific prediction (such as drug–drug interaction modeling), and self-evolving alignment methods that couple generation and discrimination within large language models. Across these efforts, the underlying theme is the same: leverage structure, interaction, and internal learning signals to build intelligence that scales even when supervision does not.

If you have ever worked cross-discipline, how did that influence your way of thinking or the way you approach your work?

Working across disciplines — especially with biomedical scientists and chemists — has significantly shaped my thinking.

In cross-disciplinary projects, benchmarks become secondary to real constraints. Data may be incomplete. Labels may take years to obtain. Predictions must be interpretable and decision-relevant. That environment forces you to think differently.

I became more “problem-first.” Instead of starting from a method and searching for applications, I now begin with the scientific bottleneck and the downstream decision need. Then I design the right abstraction and modeling framework around it.

For example, in work on emerging drug interaction prediction (Nature Computational Science, 2023), we embedded biomedical network structure directly into the learning process to enable prediction for new drugs with limited supervision. In property-aware molecular modeling (IEEE TPAMI, 2024), we incorporated substructure-level patterns aligned with chemical properties to enhance interpretability and generalization.

Cross-disciplinary collaboration has reinforced my belief that principled modeling and domain knowledge must evolve together.

What advice would you give to young professionals or recent graduates who are trying to enter your field?

First, build strong fundamentals. AI moves quickly, but mathematics and core principles endure. A solid foundation allows you to adapt across waves of new models and tools.

Second, learn to ask good questions. When reading papers or starting projects, ask not only “what did they do?” but “what assumption made this possible?” and “what prevents this from working in the real world?”

Third, cultivate clarity. If you cannot explain your idea simply, you probably do not understand it deeply enough. The best solutions are often those that are both powerful and easy to articulate.

Machine learning is evolving rapidly, but thoughtful, principled researchers will always be needed.

Make Connections

You can find Quanming Yao on LinkedIn.

Your Effort, Your Community

Your successful career in computing begins with reaching out to your community. Active community engagement and networking create professional growth. The IEEE Computer Society provides various platforms for contribution, such as:

• Becoming a Distinguished Visitor
• Mentoring at conferences
• Acting as a judge for hackathons and challenges

Discover how you can participate with the IEEE Computer Society and volunteer to build your community.