Most AI research chases the next bigger model. One lab in Nanjing is asking a quieter question: when a model predicts something bad, can it also tell you how to stop it?
That lab is LAMDA, short for Learning And Mining from Data (从数据学习与挖掘), at Nanjing University. Its leader, Zhou Zhihua (周志华), is one of the most decorated machine-learning scientists in China — and in 2025 his team walked away with one of the field's most competitive honors.
Key takeaways
- Zhou Zhihua's LAMDA group won the AAAI 2025 Outstanding Paper Award for "abductive learning" (反绎学习), a method that fuses logical reasoning with data-driven learning.
- The same lab also took a best-paper award at DATE 2025, an electronic-design-automation conference — a rare double in a single year.
- Zhou's newer research pushes "avoiding undesired future" (AUF): building models that recommend actions to prevent a predicted bad outcome, not just flag it.
- He is the first researcher from mainland China to be elected a Fellow of all five major international AI societies (ACM, AAAI, AAAS, IEEE, IAPR).
Who is Zhou Zhihua?
Zhou, born in November 1973 in Guangzhou to a Yancheng family, is a homegrown academic product of Nanjing University, where he earned his bachelor's, master's and PhD (1996, 1998, 2000) and never really left. He joined the faculty in 2001 and founded LAMDA in 2004. Today he is a vice president of the university and dean of its School of Artificial Intelligence.
To students across China he is better known by a nickname: the author of the "Watermelon Book" (西瓜书), a 2016 Chinese textbook titled Machine Learning that became the standard entry point for a generation of Chinese AI learners. It has since been translated into English, Japanese and Korean.
His standing is institutional as well as popular. In 2023 he became president of the International Joint Conference on Artificial Intelligence (IJCAI) council, and in November 2025 he was elected an academician of the Chinese Academy of Sciences — the country's highest scientific honor.
What the award was actually for
The AAAI 2025 paper tackles a problem most deep-learning systems ignore. Neural networks are brilliant at prediction but weak at explaining why, and they struggle when a logical rule contradicts a pattern in the data.
Zhou's group proposed a refinement of "abductive learning" (反绎学习), a framework they first introduced in 2019 that combines a neural network with a symbolic reasoning engine. Instead of treating the two as rivals, it lets them correct each other: the network handles messy real-world signals, the logic engine enforces rules, and inconsistencies get resolved through what the authors call "abductive reflection."
The payoff is a system that is both data-driven and rule-aware — useful anywhere a wrong prediction carries a real cost, from medical triage to industrial control. Out of 12,957 submissions to AAAI 2025, only three papers received the Outstanding Paper distinction.
The bigger idea: prevention, not just prediction
A separate thread of Zhou's work, presented at NeurIPS 2025, goes further. His group formalizes "avoiding undesired future" (AUF): imagine a portfolio model that predicts a bad quarter. A normal system stops at the warning. An AUF system asks what allocation changes would keep the outcome inside the safe zone — and proves, with a mathematical bound, that the recommended action actually helps.
That reframing matters because most deployed AI today is a forecaster, not an advisor. Shifting from "here is the risk" to "here is the move that lowers the risk" is the difference between a dashboard and a co-pilot.
Why a global reader should care
China produces a large share of the world's AI papers, but the work that travels internationally is often judged by whether it opens a new direction others can build on. Abductive learning is one such direction: it has already spawned follow-up work at other labs, and the AAAI honor puts a Chinese-led framework on the same stage as research from Google, Meta and the top US universities.
For ordinary readers, the practical takeaway is modest but real: the safer AI systems of the next decade may come less from bigger models and more from methods that pair learning with old-fashioned logic — exactly the bet Zhou's lab is making.
Honest limitations
This article is based on Nanjing University's official CV of Zhou Zhihua, the AAAI 2025 program records, LAMDA's published papers, and coverage of the awards. I have not independently reproduced the mathematical proofs behind abductive learning or AUF; the performance claims are the authors' as reported in peer-reviewed venues. The "first Fellow of all five societies" claim reflects listings current as of the award period and could shift as societies elect new members. I found no verified public quote from Zhou specifically about the 2025 paper's commercial timeline, so I have not implied one.
What readers can do now
- If you study or teach AI, skim the open-access AAAI 2025 abductive-learning paper to see how logic and neural nets can be combined.
- Follow LAMDA's publication page for the AUF line of work if you build systems where wrong predictions are costly.
- Treat breathless "China vs US" AI headlines with caution — much of the real progress is collaborative and method-driven, not a horse race.