Toward AI That Never Stops Learning

How can we build models that keep learning from experience long after their training is complete?

The world does not stop changing when a model finishes training.

People often get better at a task by doing it. A mistake shows us what we missed. A conversation gives us another way to think about a problem. An AI model can encounter the same experiences without learning anything lasting from them. The next time it faces a similar problem it may make the same mistake.

I want to build models that can learn from those encounters. A useful correction should remain available later. A lesson from one task should help with the next. That improvement should also preserve the knowledge and abilities the model still needs.

My research at OpenMIA focuses on lifelong learning for large AI models. This question matters for models that work with language as well as those that learn from images or several kinds of input together.

One foundation of this work is model editing. The editors I study make focused updates to what a model knows without repeating its full training or requiring of past training examples. This lets us work on a particular gap in the model's knowledge and study how the change affects the rest of its behavior.

Learning without starting over

Learning a new skill involves many changes that can be hard to untangle. Correcting a single fact gives us a more manageable starting point. We know what the model should learn and can examine what happens around that change.

Imagine a university uses an AI assistant to help students find their way around campus. The model behind it has learned that the Lumen library is in West Hall and the Lumen laboratory is in East Hall.

Now the library moves to North Hall. The laboratory stays where it is. We want the assistant to learn the new address so it can guide future visitors correctly. One correct answer would be encouraging. We would also want to know whether it understands other ways of asking and still remembers the things that have not changed.

The four questions below let us check these expectations one at a time. A conversation makes the effects easy to see. The underlying problem also matters beyond language: learning something new must preserve what still works.

New knowledge Lumen library: West Hall → North Hall

Apply the correction · Reliability

“Where is the Lumen library now?”

It should answer North Hall.

The original question should use the new knowledge. Getting this answer right is the first check. The other questions tell us whether the change holds up beyond this one answer.

Recognize another wording · Generalization

“Which building is the Lumen library in?”

It should still answer North Hall.

A change in wording should not hide the relevant memory. This motivates my work on finding the right memory and finding and storing knowledge separately.

Leave unrelated knowledge alone · Locality

“Where is the Lumen laboratory?”

It should keep answering East Hall.

The similar names should not cause confusion. The library's update does not apply to the laboratory. This is why a memory's scope matters as much as its ability to retrieve something.

Remember through later updates · Retention

“After many unrelated updates, where is the Lumen library?”

It should remember North Hall.

The correction should remain useful as the memory grows. Checking old answers again and again also costs time. This motivates my work on efficient editing and evaluation reuse.

The highlighted branch shows whether this question should use the correction or keep the earlier answer.

Passing all four checks would give us more reason to trust the correction. We would know that it works beyond the original wording and remains useful as other updates arrive. We would also have checked that unrelated knowledge still holds.

Doing this once is only the beginning. A model in everyday use may receive a long stream of corrections. The cost of learning matters too. Each update must be affordable to store and apply. Checking its effects cannot become so expensive that the model has to stop learning.

Can an experience make the model more capable?

The library example concerns something a model knows. I also want to understand how an experience can improve what it can do. A coding assistant gives us a way to think about that next step.

Suppose a program fails because it assumes some data will always be available. Could learning from that failure help the model check for missing data in a different program? That is the kind of transfer I want to study: an experience becoming an ability that remains useful beyond the example that taught it.

How earlier research shaped these questions

Keeping earlier knowledge useful has long been a concern in continual learning. Several lines of research have shaped how I approach the problem. They offer different ways to control what changes when a model learns.

  1. Learning something new can disrupt something old

    A central difficulty in continual learning is catastrophic forgetting: learning new information can damage earlier knowledge or abilities. Research on protecting the parts of a model that matter for earlier tasks showed one way to reduce this interference.

    The research behind this idea

    Elastic Weight Consolidation slows changes to parameters important for earlier tasks. It is one influential approach to balancing adaptation with retention. EWC · Kirkpatrick et al., 2017

  2. Make a focused change to the knowledge that needs updating

    Another way to limit disruption is to focus on the knowledge that needs changing. Knowledge editing asks how to update a particular factual association without repeating a model's full training. Researchers developed methods for editing a single fact and then many facts at once.

    But a successful correction today is not enough if tomorrow's correction makes it disappear.

    From individual facts to many updates

    ROME demonstrated targeted editing of factual associations in transformer parameters. MEMIT extended this direction to editing many facts together. Lifelong editing asks how updates can continue to coexist as more arrive. ROME · Meng et al., 2022; MEMIT · Meng et al., 2023

  3. Give new knowledge its own memory

    Repeated edits can still interfere with one another. Memory-based editors give new knowledge a separate place so the originally trained model can stay intact. These methods differ in what they store and how they apply it.

    That separation shifts some of the difficulty to retrieval. The system must find the right knowledge in a growing memory and recognize when none of it applies. Associative memory studies how an incomplete or imperfect cue can lead to a stored memory. HoReN brings this idea to models that receive long sequences of updates.

    Memory systems and associative retrieval

    SERAC stores edits explicitly; GRACE uses a local codebook of learned updates; WISE separates pretrained and edited knowledge into main and side memories; MEMOIR uses sparse updates in a residual memory. SERAC · Mitchell et al., 2022; GRACE · Hartvigsen et al., 2023; WISE · Wang et al., 2024; MEMOIR · 2025

    Modern Hopfield networks provide a mathematical foundation for associative retrieval. HoReN uses normalized Hopfield retrieval to refine queries toward relevant stored edits. This is a direct starting point for my work on interference and efficient execution. Ramsauer et al., 2021; HoReN · Fang, Xie & Ran, 2026

Within the broader field of AI, I work on continual learning: how a system can keep improving as experience arrives. My focus is on large AI models. Lifelong model editing gives me a way to study how they can learn throughout their useful lives.

Editing is one part of that effort. A model can also draw on external information or learn through further training. I am interested in how these approaches can support each other as experience accumulates.

Making learning reliable and efficient

My work with collaborators begins with the practical problems that appear as an editing system's memory grows. We want each update to remain useful without disrupting earlier knowledge or making the next update too expensive. The four studies below use language-model editing to investigate these problems within my broader research on large AI models.

These algorithms focus on in-parameter learning (IPL). They learn by changing the numbers that govern how a model processes information. Those numbers can also belong to an attached learning module. In our editors the original model can stay fixed while an attached memory learns updates that change its computation. This is where my work on lasting knowledge updates begins.

Find the right memory as memories become crowded

A memory is useful only when the model knows when to use it. As memories accumulate, a question about the library can become harder to distinguish from one about the laboratory.

We developed MH-HoReN from HoReN to compare a question with memories in several complementary ways. Combining the matching signals helps the system distinguish relevant memories from nearby ones. We also study how related edits can be grouped for more efficient learning. In our evaluated settings this improves the balance between applying new knowledge to relevant questions and leaving unrelated answers unchanged.

Mitigating Memory Interference in Lifelong Model Editing through Multi-Head Hopfield Retrieval

EMNLP 2026 Main Conference · Co-first author
A closer look at the method

Multiple retrieval heads operate in complementary representation subspaces. Their matching signals are combined to reduce memory interference. Grouping related edits also supports batched editing. Highly ambiguous queries and overly aggressive grouping remain difficult.

Understand where a memory applies

The next problem is deciding when to use no memory at all. A new question may resemble a stored one without asking for the same information.

Our work on local associative memory studies the boundary within which a memory should apply. An incomplete cue should still recover the right knowledge. A question outside that boundary should be rejected. The library and laboratory show why this matters: their similar names do not make the library's address relevant to the laboratory.

A Local Associative Memory that Knows Where It Holds

ICLR 2027 Conference Submission

A memory needs to recognize both what it knows and when that knowledge should be used.

Keep more knowledge without making each memory expensive

Accurate retrieval also needs to be affordable. A long learning history becomes expensive to maintain if every new memory occupies too much space.

BiSpace separates the representation used to find knowledge from the update used to apply it inside the model. This lets us improve how the system recognizes a relevant question while keeping the corresponding update compact. The aim is to support longer learning histories with a smaller storage cost per update.

Decoupled Knowledge Retrieval and Compact Memory Enable Reliable Lifelong Model Editing

ACL 2027 Submission
What the two spaces do

BiSpace identifies relevant edits in an external semantic representation space. It then injects their knowledge into the language model's internal representation through sparse, storage-efficient updates. Retrieval and knowledge injection can therefore be optimized for their different roles.

Make each update and its checks affordable

Storage is one cost of keeping a learning history. Updating it and checking its effects are others. An editing system may recompute unchanged parts of a model or set aside training memory for entries it is not updating.

eHoReN uses the dependencies of an update to determine what needs to be computed and stored. The same idea helps with evaluation. Under the method's stated conditions we can check whether an update leaves the computation behind a previous answer unchanged. A passing check lets us reuse that answer. Otherwise we generate it again.

Fast and Memory-Efficient Lifelong Model Editing with Certified Evaluation Reuse

ICLR 2027 Conference Submission
When can an earlier answer be reused?

The reuse check depends on conditions including a frozen backbone and deterministic decoding. It checks the effective memory contributions at every generation step of the earlier answer. A passing check certifies reuse under these assumptions; a failed check triggers regeneration.

Training state is limited to the active memory entry during an edit. Unchanged computation can also be reused. This reduces the cost of an update. The persistent memory still grows as new entries are added.

How these pieces fit together

The four projects address requirements that a learning system has to meet together. They offer complementary methods that we can draw on when building such a system.

Four connected questions behind reliable, efficient learning

A precise update is more useful when it is cheap enough to keep and check over time. Likewise, compact storage only helps if the model can still find the knowledge it needs. Improving one part should support the others.

There is a decision that comes before all of this. Much of model editing starts with a correction someone has already chosen. A model learning from everyday experience would also need to decide what is worth learning and how to learn it.

Toward self-evolving large AI models

These decisions are central to the framework my collaborators and I are building for self-evolving large AI models. We want a model to identify useful lessons from its interactions and learn from them as part of its everyday operation.

What to learn and how to learn influence each other. A failed programming attempt might expose a missing concept. Another failure might come from applying a familiar method in the wrong situation. The lessons would be different and might need different forms of learning. We want feedback on the outcome to improve both choices over time.

Two ways to learn from experience

One choice concerns how an experience reaches the model. A worked example or a useful reminder can be included with its current task. The model can use that information without changing its parameters. This is in-context learning (ICL).

The same experience could also provide material for in-parameter learning (IPL). A parameter update can carry its lessons into the model or an attached learning module. My algorithmic work focuses on this route. The framework brings IPL and ICL together so we can study how they help each other.

The learning process we are building toward
Interaction, experience & feedback
What is worth learning?How should it be learned?

Use experience in context In-context learning · ICL

Give the model relevant material to work with on its current task.

Learn from what is available now

Learn through parameters In-parameter learning · IPL

Train persistent updates to the model or its attached learning modules.

Carry selected lessons into learned state

Useful context can provide material for parameter learning. Learned abilities can help the model use future context.

Try new situations. Check what improved.New abilities · Earlier knowledge · Computation and memory cost

The results guide both what to learn next and how to learn it.

We can return to the coding assistant to see how this might work. It could begin with a worked example in context and try a new task. Feedback on that attempt might reveal a lesson worth learning through a parameter update. We would then test whether the update helps on different problems. The improved model could still use context for details specific to its next task.

Where do external memory and retrieval fit?

External information can supply context repeatedly through methods such as retrieval-augmented generation (Lewis et al., 2020). A model can use context without a parameter update. In-context learning in GPT-3 (Brown et al., 2020) is one example.

I use IPL here to include learned parameters in attached modules. Retrieved memories in HoReN-style editors contain trained values that change the model's internal computation. Retrieving these learned values differs from placing retrieved text in a prompt. A fixed base model can therefore participate in a system that learns through parameters.

Neither mechanism alone establishes that a model has acquired a transferable ability. That needs to be tested on new problems. Retention and resource costs also need to be measured.

For this process to be useful we need evidence that it helps beyond the experience it learned from. That leaves several questions I want to pursue as we build the framework.

Learning something useful for a different problem

A lesson from one failure should help with another problem that needs the same insight. We need evaluations that can distinguish this from recalling an answer. Comparing ICL and IPL under the same task and resource conditions will help us understand when to use either one and when their combination helps.

Keeping what still matters

Those gains would mean little if they disrupted earlier abilities. Even a frozen base model can give wrong answers when its memory applies the wrong update. I want to make this interference more predictable and controllable while allowing outdated knowledge to change.

Working within the resources people have

The cost of this process will determine where it can be used. Personal devices have limited memory and energy as well as limited computation. I want to develop simple algorithms that work within those constraints. An assistant could then learn the tools you use and improve through your corrections on the device where it is already helping you.

The future I hope to build

That possibility is what keeps me interested in learning after deployment. I want to understand how far a model could improve through the experiences it encounters over its useful lifetime. I am especially curious about what a small model on an ordinary device could come to do.

Making learning part of everyday use will require updates that people can rely on and afford to make often. These are problems I hope to spend my research career working on. I would like the models we build to grow alongside the people they serve.

I want the day a model is deployed to be the beginning of its learning, not the end.

Explore and connect

To make these ideas easier to explore, Xuming Ran and I put together Where Should New Knowledge Live?. The interactive experiments let you try different ways of storing and updating knowledge. There are also notebooks and teaching materials for a closer look.

You can find the papers behind this work on my publications page. I would also like to hear how these questions connect with your work or with a problem you see in another field. You are welcome to get in touch.