For much of machine learning’s history, the central assumption has been convenient: training data describe the world a model will later encounter. A research group at Nanjing University has spent years working on the opposite premise. In an August 31 profile, Sina Tech, republishing China Science Daily, described Chinese Academy of Sciences academician Zhou Zhihua’s effort to build machine-learning theory for systems that must function after the environment has shifted, new categories have appeared, and information remains incomplete.
The project is called “Machine Learning for Open Dynamic Environments.” It was approved in 2019 under a National Natural Science Foundation of China Innovation Research Group program and was the first such group funded under the foundation’s artificial intelligence primary discipline code created in 2018, according to the report. That history makes this more than a new laboratory announcement. It is a snapshot of how a Chinese research program has framed a long-term problem that is becoming more visible as AI moves from demonstrations into workflows and physical systems.
Zhou’s team condenses the challenge into four ideas: small samples, changing distributions, frequent updates and interaction. Each describes a failure mode familiar to developers. A model may have very little relevant data for a new situation. The statistical pattern it learned may no longer hold. New information may arrive continually. Or the system may need to act and learn from the consequences, rather than merely classify a static record.
The problem becomes clearer beside the new generation of agent benchmarks. EastFrontier recently examined how Baidu’s DuMateBench tests whether China’s AI agents finish work. Benchmarks can measure a system’s performance on defined tasks, but they cannot remove the difficulty of a changing setting. Nanjing’s work asks what happens when the task itself begins to change around the model.
Four Problems Define AI Outside the Training Set
Small samples are the first obstacle. In a newly changing environment, the data that matter most may be scarce precisely because the condition is new. A conventional system trained on a large historical collection can be ill-equipped when it meets an unfamiliar signal or a rare event. Zhou’s group is investigating how machines can extract useful structure under those limits rather than assuming that more historical data will always be available.
The second problem is changing distributions. The profile uses environmental monitoring as one example: shifts in climate or industrial activity can change how pollutants spread, and a model that keeps applying old regularities can deliver distorted results. The broader lesson applies to supply chains, factories, markets, and mobility systems. A model may be accurate at one point in time and less useful later without an obvious software failure.
Frequent updates create a third challenge. In the group’s formulation, a system needs a way to keep learning when the underlying conditions are moving. That is not simply a matter of retraining on a schedule. Updates can introduce new data biases, erase useful prior behavior or consume expensive computing resources. The research program’s interest in online learning is therefore central: the model must adapt without treating every new observation as a reason to start from scratch.
The fourth issue is interaction. A model operating in an open environment cannot be purely passive. It needs to collect the information required for a decision, then interpret what happens after it acts. That is one reason the theory has relevance beyond traditional classification. In a physical setting, an agent may need to navigate a new obstacle. In a service system, it may need to recognize that a user’s problem does not fit categories present in the original dataset.
From Dynamic-Regret Theory to Robust AI Systems
The Nanjing team’s reported work includes an optimal dynamic-regret online-learning mechanism and an open-environment learning framework. The article also says the group resolved a causal-identification question that had challenged researchers for two decades. Those are substantial claims, but they should be read as descriptions of the group’s research results, not as proof that deployed products have solved every real-world reliability issue.
The program’s stated approach is valuable because it separates foundational theory from product headlines. Generative AI has made model performance highly visible, yet an impressive answer on a fixed test does not guarantee a dependable response once the data distribution changes. Tencent’s PatternEval work, another research story in today’s file, makes a related point by examining how hybrid-thinking systems can display response-pattern failures even when ordinary correctness metrics are insufficient.
The relationship between these areas is not that one paper validates the other. Rather, both show why evaluation has to account for behavior that standard scores can miss. PatternEval focuses on how a model’s thinking and non-thinking modes affect delivered responses. Nanjing’s framework focuses on the external setting in which a model operates. Together, they illustrate why AI reliability cannot be reduced to a single leaderboard number.
The project’s history also reflects China’s preference for long research cycles in strategically important technical fields. Zhou’s work in machine learning began in the 1990s, the report says, and the team started laying out the open-environment direction around 2012. Funding under the 2019 Innovation Research Group project gave the researchers a multi-year structure for problems that do not yield easily to short product cycles.
Why Open-Environment Learning Matters for China’s AI Strategy
The research team argues that basic machine-learning theory and applied engineering play different roles. Theory investigates the underlying mechanisms and algorithms, while application work adapts existing methods to particular tasks. In today’s AI market, the second category often receives faster commercial attention because it can produce visible results sooner. Zhou’s group is making the case that the first category determines how far systems can eventually extend beyond their original use cases.
That argument has practical implications for China’s industry. Chinese companies are rapidly deploying models in factories, public services, logistics and robotics. The more varied those environments become, the less safe it is to assume that a static training corpus will remain representative. Amap’s long-range 3D AI work offers one applied example of why systems must interpret changing environments. Open-environment learning addresses a technical limitation that may become more expensive as deployments spread.
The report says the Nanjing team has established a joint laboratory with a domestic leading company and later an advanced research center, but it does not identify the partner. What matters is the model of collaboration: research directions are meant to influence industrial work, rather than only accepting a narrowly specified corporate problem. That arrangement will be watched closely as China tries to translate foundational research into commercial AI systems.
No theoretical framework can guarantee that an AI system will behave correctly in every unknown setting. The Nanjing program instead offers a more disciplined goal: identify the assumptions that fail outside a controlled dataset, then design methods to recognize new categories, adapt to drift and interact safely. As Chinese AI expands into more complex environments, that may become a more meaningful measure of progress than whether a model can merely perform well in the conditions it has already seen.