对话上交大蔡登：比RSI更本质的是部署后自进化，关键是让上下文学…

Conversation with Shanghai Jiao Tong University’s Cai Deng: More fundamental than RSI is post‑deployment self‑evolution; the key is to let the contextual literature…

Currently, most large‑model “memory” relies on a single mechanism: storing information in a database outside the model’s weights, retrieving it when needed, and injecting it into the context window for the model to use. This approach was formally named Retrieval‑Augmented Generation (RAG) by Facebook AI Research in 2020. From web‑connected search to various agents’ memory systems, this framework is essentially the same.

Even before the RAG concept was formally established, Cai Deng had already begun research in this area. In 2019, while pursuing his PhD at The Chinese University of Hong Kong, he proposed using retrieved historical dialogues as “memory” to guide a dialogue model in generating responses.

Seven years later, his assessment of this approach has changed. Information retrieved from external sources ultimately still has to fit into a limited context window, many experiences that are hard to verbalize cannot be stored in a database, and the knowledge required for complex tasks is difficult to obtain fully in a single retrieval.He believes that true continual learning ultimately requires the model to update its own parameters during use.

This judgment is closely related to the two current industry mainlines.The first is Recursive Self‑Improvement (RSI):This direction advocates having AI participate in designing and training the next generation of AI, and there have been many recent advances. However, Cai Deng argues that this work is essentially pre‑deployment optimization, optimizing for the average across all deployments, while private data, tacit experience, new insights generated through trial and error, and capability gaps created by rapidly changing environments can only be addressed after deployment.

The second is Personal AI:The general capabilities of foundational models are already sufficient for most everyday tasks; what hinders real-world deployment is more often the model’s lack of understanding of specific users, organizations, and work environments.

Cai Deng has worked at Tencent AI Lab and ByteDance Seed, where at ByteDance he mainly focused on research into large‑model memory mechanisms to enhance long‑context capabilities. He received the Outstanding Paper Award at ACL 2021 as first author and has long served as a domain chair for major conferences such as NeurIPS, ICML, ICLR, and ACL.In November, he will join the School of Artificial Intelligence at Shanghai Jiao Tong University as a tenured associate professor, establishing the ContinuousAI Lab to study AI systems that can continuously learn and evolve after deployment.

Recently, DeepTech had an in‑depth conversation with Cai Deng, discussing why he returned to academia from industry, the ultimate form of AI memory, and the technical path for self‑evolution after model deployment.

When discussing the next growth point for large models, his assessment is that after expanding parameters and data in pre‑training, expanding annotations and feedback in fine‑tuning, and extending thinking time during inference, the fourth scaling axis will be the interaction history and feedback accumulated by the model in real work, which he calls Post‑Deployment Scaling.

Regarding learning methods, he believes In‑Context Learning (ICL) is “the ultimate learning paradigm,” and the real challenge is making it scalable.In his view, the endpoint of post‑deployment self‑evolution is that every individual and organization has an AI that continuously evolves alongside them.

The following is the interview transcript between DeepTech and Cai Deng, edited and compiled.

From ByteDance to Jiaotong University: Seeing the “Fourth Scale Axis”

DeepTech: Please briefly introduce your previous academic and professional background.

Cai Deng: I graduated with a PhD from The Chinese University of Hong Kong in 2022, and my first job was at Tencent AI Lab. Tencent AI Lab is more like a traditional industrial lab, focusing on academic research. Later I went to ByteDance Seed, participating in cutting‑edge model development, focusing on long context and memory capabilities of models.

DeepTech: At ByteDance Seed, what area did you work on the most deeply?

Cai Deng:First, the long-context capability of large models, which can be viewed from two angles: performance and efficiency. In terms of performance, we mainly start from the data side. Naturally occurring data with long-range dependencies is very scarce, so to enable the model to better integrate information scattered across long texts, we often need to manually synthesize similar data.

In terms of efficiency, the focus is on improving the model architecture, such as using new structures like Sparse Attention to reduce latency. These are foundational model-level efforts. The second area is research on memory mechanisms, including building memory systems. After doing all this, my biggest takeaway is: the context window can keep extending, but the model still cannot achieve true knowledge integration, and both training and inference remain inefficient. Memory systems can alleviate the issue but cannot solve continual learning.

DeepTech:Returning from industry to academia, how has your research mindset changed?

Cai Deng:I’ve been pondering the same question: how to enable models to keep learning and evolving.During my PhD and at Tencent AI Lab, I mainly worked on RAG, and I can say I was one of the earliest researchers in this area; in 2022 I even gave a RAG tutorial at SIGIR and IJCAI. At ByteDance, I’m responsible for long-context and memory mechanisms.

Looking back, RAG is external memory, long context is short‑term memory; I have worked on both at the cutting edge and have seen their ceilings. The only path left is to write experience into the parameters, turning it into long‑term memory. I started pursuing this route in 2024, and I believe the future breakthrough lies here.

Figure | Cai Deng et al. presenting a tutorial on retrieval‑augmented text generation at SIGIR 2022 (source: SIGIR 2022)

You can think of it this way: large models already have three scaling axes. Pre‑training expands the model parameters and data volume, acquiring world knowledge; post‑training expands annotations and feedback, aligning with human preferences; inference‑time scaling expands the tokens consumed during inference, enhancing inference capability.

What is the next scaling axis? I think it is the model’s interaction history in real work, along with feedback from the environment and users.I call it Post‑Deployment Scaling, i.e., self‑evolution after deployment. The product of the first three axes is a more powerful general model; the product of the fourth axis is “your” model. This is the first growth curve that belongs to every individual and every organization.

Today I realized that self‑evolution after model deployment is a huge opportunity. In intelligence‑focused contests like math competitions, AI has already left humans behind, even showing a trend of replacing human mathematicians in research; but in the everyday work and life of the general public, when you ask AI to write a research report, it struggles to produce something that quickly satisfies you.

Once a general model’s intelligence crosses a certain threshold, the marginal returns of further scaling diminish rapidly. For most everyday tasks, a general model’s capabilities are already sufficient; being a bit smarter is less valuable than being a bit more attuned to you. The bottleneck limiting AI adoption is shifting from “not smart enough” to “not understanding you enough”: lacking knowledge of the specific deployment environment, goals, and preferences.

This implies two things.First, foundational models are continuously getting stronger,and the stronger they become, the greater the need for post‑deployment self‑evolution and the greater the value they can deliver.Second, swapping to another general‑purpose model incurs almost no cost,but it’s hard to abandon an AI that has been learning in your environment for months, so the value shifts from the model itself to what it learns after deployment.

DeepTech:Are there any tasks you feel can only be done by returning to academia?

Cai Deng: I don’t think you have to go back to school to do it; it’s just that leaving a big tech company may give you more freedom to explore more aggressive solutions and pursue a win‑in‑the‑long‑run approach.

The first three axes are about resources: parameters, data, annotations, and compute power—only big companies can afford that. Their resources and evaluation metrics are aligned with making general models stronger, which is correct, but it also means they find it hard to focus on post‑deployment issues. The fourth axis is just beginning, and it doesn’t require training a foundation model from scratch: by standing on top of the best open‑source models, you can directly conduct cutting‑edge research.

DeepTech: This time I’m joining the newly established Shanghai Jiao Tong University AI Institute in 2024. What attracts you about it?

蔡登: The AI Institute’s personnel composition is more diverse, including senior professors, newly minted PhD scholars, and people like me who have worked in industry. Today, in foundational model research, academia and industry are somewhat disconnected, with different focuses, so diversity in personnel is important.

In addition, the university can continuously bring in young people who are willing to explore cutting‑edge topics with me, so we can truly focus without distraction on important research problems over the long term. My role as a professor also allows me to engage in broader collaborations both inside and outside the campus. This school is relatively young, and many faculty members choose to start companies; the school provides some support for entrepreneurship, creating an overall open atmosphere.

The Ceiling of External Memory

DeepTech: You worked on RAG early on, and the industry is now discussing various memory systems. In your view, what is the relationship between the two? How far can this path go?

Cai Deng:The so‑called memory systems today are essentially an extension of RAG. Whether it’s a startup dedicated to memory or the built‑in memory features of products like ChatGPT and Claude, they all store memory outside the model and retrieve it when needed. How information is organized for maximum efficiency and how queries are performed are continuously evolving, but the core idea remains unchanged: the model’s memory resides outside the weights and is injected back into the context window when required.

This approach has strong limitations. Since the information ultimately still needs to be injected into the context window, complex long‑running tasks in real scenarios require many knowledge points and skill sets. The more content you inject, the more you have to contend with various issues arising from long contexts.

Fundamentally, I think this paradigm has two inherent limitations.

The first is “cannot be stored.” Some things cannot be fully captured in words, such as an investor’s feel, market sense, and judgment, a researcher’s long‑term cultivated academic taste, and even a person’s personality preferences. If you cannot obtain the complete original experience and only retain fragmented snippets, it is difficult to reproduce what has been distilled from the full experience. As AI moves toward multimodal, embodied intelligence and world models, information like audio‑video, sensory streams, and motion streams becomes increasingly hard to summarize in text, which is an unavoidable limitation.

The second issue is “cannot retrieve”. In real‑world complex tasks, you often rely on knowledge or skill points that are scattered across different places and interrelated. For example, completing a task may require knowledge points A, B, C, and D, and you are unlikely to find them all on the first query: B is relevant to the task because you first encountered A; C is relevant because you first saw A and B. The information‑retrieval chain is necessarily long and complex; the longer the chain, the lower the efficiency and the more errors tend to accumulate.

This is also not how the human brain learns. The brain’s capacity does not expand infinitely as experiences increase; external memory only becomes more bloated and harder to retrieve. More importantly, preserving the original context is not learning, and textual summaries cannot form intuition.Therefore, external memory will never reach the end; true continuous learning must rewrite the model’s weights.

Three-layer architecture of memory: facts rely on retrieval, abilities rely on internalization

DeepTech: External retrieval, long context, and parameter memory—will these forms of memory eventually develop a division of labor?

Cai Deng:You can think of it this way: the external memory stored in a database outside the model is external memory, what is placed in the context is short‑term memory, and the knowledge and skills stored in the model parameters are long‑term memory. All training of the model can be seen as the transformation from external memory to short‑term memory, then to long‑term memory. Pre‑training works the same way: extract content from web data, put it into the model’s context, and perform a gradient back‑propagation with the goal of predicting the next token (Next Token Prediction), thereby writing it into the model parameters.

Many people discuss the division of memory—what should be kept external and what should be written into parameters. How to organize information in external memory, summarize and abstract it, and manage it hierarchically are indeed worth doing well. My view is: intuition, ability, experience, and judgment should all be encoded into the parameters; precise factual details—like an exact number in a contract or a direct quote someone said—are better kept in external retrieval, because it’s neither possible nor necessary for the model to hard‑memorize them.In short: facts rely on retrieval, abilities rely on internalization.

Figure | The six transformations between reward models, parameter updates, and context prompts organized by Cai Deng et al. (source: arXiv)

DeepTech: Where do the various memory approaches stand today?

Cai Deng: The three types of memory are at completely different stages. External memory is already quite mature; the remaining challenges are more engineering problems: how to organize information better and retrieve it faster and more accurately when needed. Short‑term memory is in an arms race, with the context window expanding from a few thousand tokens to millions, but as I mentioned earlier, even though the window is longer, the model’s ability to integrate information has not kept pace.

Research on long‑term memory is still relatively scarce, especially post‑deployment self‑evolution: how to write back the interaction history generated by the model in real work, observations of the environment, and feedback from users and the environment into the model’s interior, making it part of long‑term memory. I believe this will become a future research hotspot and the main battlefield for the next generation of AI competition.

DeepTech: Beyond memory there is also forgetting. Whether it is intentional model forgetting or catastrophic forgetting, is this still a key focus in the field? When new and old information conflict, who should decide what to keep, update, or delete?

Cai Deng: Yes, especially in continual learning, the model’s experience is constantly flowing in. After new stuff arrives, how to ensure useful old information isn’t forgotten, and how to perform deep updates when conflicts arise—these are fascinating questions. I believe letting model weights evolve continuously is precisely the right direction to solve these problems. Knowledge integration occurs within the same parameter space, making it more efficient. The key is what objective to use for the modification.

If you directly fine‑tune with new data, the model is fitting a bunch of external samples and can easily drift away from its original self; whereas internalizing with the model’s performance after reading the context sets the goal as “a more knowledgeable version of itself,” making the start and end points both the model itself, which naturally prevents it from overwriting its existing capabilities. In our experiments with hundreds of consecutive internalization rounds, we have not observed any performance collapse.

As for which content should be retained long‑term, updated, or deleted, I believe that should be left to the model’s own judgment. The development direction of RAG and memory systems is the same: continuously unleash the model’s autonomy, letting it organize its own database without human‑defined rules about where things are stored or how they are removed.What we need to do is evaluate results end‑to‑end: each update is measurable, versioned, and rollbackable; as long as the model’s performance keeps improving, its actions are correct; if performance regresses, we roll back first, then let the model reflect on the cause.

DeepTech: In privacy-sensitive scenarios, are there limits to a model’s autonomous updates?

Cai Deng: Personalization and privacy are inherently contradictory: the more AI understands you, the more it needs your data. The most straightforward approach is private deployment, or moving capabilities to the edge, keeping both model and data in the user’s own environment, which is the most thorough protection. In other words, the AI that knows you best should belong only to you.

Is contextual learning the ultimate end?

DeepTech: In a 2025 tweet you said, “All learning should be done through in-context learning, the key is to make in-context learning more effective and able to learn from infinitely long contexts.” Do you still believe this judgment today?

Cai Deng: More convinced than before. Regarding learning from experience, in-context learning (In-Context Learning, ICL) is already the optimal solution and the ultimate learning method.

There are four reasons.First, its sample efficiency is extremely high.Writing one or two examples or a brief task description—such as giving the model a skill specification—can change its behavior; whereas supervised fine‑tuning (SFT) or reinforcement learning (RL) usually requires thousands of samples.

Second, it generalizes very well.Any fine‑tuning tends to overfit. For example, if we inject a new fact into the model—“the U.S. president has changed”—and train the model on that sentence, it will memorize the fact tightly, but when asked divergent questions, such as what the next U.S. policy toward China might be, it may err. However, if you simply place that sentence in the context without any additional training, the model can reason effectively based on that information.

Third, it doesn’t cherry‑pick signals.Every user prompt, every correction—so long as it is a signal—can be leveraged by contextual learning. SFT and RL are far more picky: SFT requires a standard answer, RL needs a final right‑or‑wrong judgment. We see many people working on model self‑evolution in domains where the feedback loop is short and the feedback signal is clear, such as using AI for AI research (AI for AI) or certain sub‑problems in AI for Science. The techniques that develop in those areas, I believe, have limited transferability and are not suitable for general scenarios.

Fourth, it has another great property: it leverages model capability.Given the same context, how much the model can learn is determined by its own ability. Stronger models grow more; as long as what is learned can be retained, future learning ability also improves, forming a compounding structure.

DeepTech:So where is its bottleneck?

Cai Deng:The only problem right now is that we can’t scale it. When the context gets long, the model loses focus, scattered information can’t be integrated, and the benefit quickly diminishes; the memory and compute cost of inference keep growing. What we really need to conquer is scalable contextual learning.

DeepTech:How can we make contextual learning scalable?

Cai Deng: The idea is simple: long ones are done poorly, short ones are done well, and in short contexts it’s hard to find a better approach than contextual learning. So we split the long into short pieces, internalize the effect of each context segment into the model’s weights, then discard that segment of context, freeing the context window for new content. Each time the context is cleared and we start from scratch, contextual learning can continue indefinitely. Therefore I say contextual learning is the ultimate goal, not that everything must remain in the context: contextual learning decides what to learn and how it turns out, while the parameters are responsible for retaining what has been learned.

The human brain has actually been doing this for a long time. During the day, the hippocampus makes a quick but fleeting record of experiences, much like context; real growth happens during sleep: the brain replays the day’s experiences in batches, and through replay, reflection, reorganization, and imagination it rewrites neural connections, turning short‑term experiences into long‑term abilities.What AI lacks most today is a single “sleep.”

We have already validated this approach in several rounds. The early‑2024 Temp‑LoRA was one of the first works to internalize context into parameters; last year’s InfiniteICL (ACL 2025) transferred 90% of the context into parameters, keeping only 10%, achieving an average performance of 103% of full‑context, and multi‑hop reasoning reaching 117%. Internalization can fuse scattered information, making inference better than having the entire text in view. In scenarios with up to 2 million tokens, using only 0.4% of the original context already surpasses the 128K full‑context baseline—and the longer the context, the clearer the advantage.

Figure | InfiniteICL (ACL 2025): The context window is refreshed after each transformation, allowing infinite context to be fed in a streaming manner. (Source: arXiv)

DeepTech: To make this path work, what is the key?

Cai Deng:Two sides. One side is to make short-context learning good enough. The definition of “short” itself is also changing; a 128K‑token context used to not be considered short, but now it is. What the base model needs to do is continuously extend the range of “short” and make it more effective.

The other side is to make the conversion from context to weights good enough, i.e., truly and effectively internalize a short context into the model parameters. The key to “good” here is that the goal is not to have the model memorize that context, but to become “the self after having read that context.” The same business plan yields different insights for different investors; what needs to be internalized is the growth after reading it.

This is also the fundamental difference between our approach and many new architectures. Architectures like linear attention and state-space models compress the context into a fixed-size state, but their goal and result are only to improve efficiency; it’s already good if performance doesn’t degrade. To break the performance ceiling of long contexts, I think the key is not the architecture; we need a new training paradigm. The paradigm I believe in uses the effectiveness of contextual learning as a metric, internalizing capability into the parameters.How to break the performance ceiling of long contexts, I think the key is not the architecture; we need a new training paradigm. The paradigm I believe in uses the effectiveness of contextual learning as a metric, internalizing capability into the parameters.

DeepTech：If we have to train every time we internalize, won’t the cost be too high?

Cai Deng:The cost of training is often overestimated. Training, like reading context, processes all tokens in parallel, so compute utilization is high. Just look at the API pricing: the price for input tokens is usually far lower than that for output tokens. Moreover, spending more on training can save you during inference: after internalization, inference no longer needs to carry a long context. In addition, model services naturally have peaks and troughs; training can be placed in the troughs, using idle compute. Our research is also continuously improving the efficiency of internalization.

One more point not to forget, the “Bitter Lesson” in machine learning tells us:Methods that can convert more compute into better performance will ultimately prevail. My view is that we should explore higher‑ceiling solutions while also reducing costs.

The answer is not a stronger general model, but continual learning

DeepTech: Besides better understanding users, what else can a model only learn after deployment?

Cai Deng: It’s not just personal AI. The tasks AI needs to perform are getting longer: from a few minutes of Q&A, to humans breaking down tasks and AI executing for several hours, to humans providing only a goal and AI working continuously for days or even weeks; later, AI will reside permanently in your system, proposing what to do on its own. The longer and more autonomous the task, the more it relies on context that exists only within that deployment. I believe this kind of context has four characteristics.

First, privacy. Internal corporate documents, meeting minutes, and workflows do not exist in public corpora, making it difficult to perform targeted optimization before deployment.

Second, tacitness. The intuition investors develop after reviewing numerous projects, and researchers’ judgment and taste regarding research questions—these cannot be written down or fully handed over to a model; they can only be accumulated through long‑term practice.

Third, dynamism. Marketing faces a constantly changing market: trends shift, competitors launch new moves every week, and strategies that worked yesterday may become ineffective today, requiring learning while doing.

The fourth is nativeness. In high‑value, complex scenarios such as scientific research, a model can hardly find the right path on the first try; it must iterate through trial and error and conduct repeated experiments, and the outcomes cannot be predicted before the experiments. This kind of knowledge can only be acquired through practice.

These contexts cannot be prepared in advance nor obtained externally; they are generated only after deployment and can only be absorbed after deployment. General intelligence determines whether AI can start working, and continual learning determines whether it can complete the work.

DeepTech:The discussion on continuous learning after deployment has been going on for many years, but actual progress has been limited. Where are the obstacles?

Cai Deng:On one hand, previous research hotspots focused on building better general models. The big trend in AI is becoming increasingly general. Before large models emerged, tasks like named entity recognition and sentiment analysis each had their own dedicated models; large models have shown that using a single general model to solve all tasks is actually better, and knowledge across different tasks can be shared and mutually reinforce each other.

But the pendulum will swing back: in the future everyone and every organization will have its own AI, which, through continuous use, will keep learning and evolving into a personalized model.The difference is that past personalized models were dedicated to a type of task, whereas current personalized models belong to a specific deployment.In the past, each personalized model was backed by a specially designed algorithm; in the future, millions of personalized models will be generated by the same algorithmic framework.

On the other hand, as I mentioned earlier, most prior research on self‑evolution was conducted in domains with clear feedback signals, such as mathematical and scientific problems, where it is obvious whether the answer is correct and how the model performs on a benchmark. In these areas, reinforcement learning for agents improves the model’s reasoning ability, which is essentially an optimization of general capability and thus returns to pre‑deployment optimization.

But feedback in real scenarios isn’t that clear. When AI helps you write a research report and some parts are insufficiently investigated, you’ll point it to another paper or website; if you’re really dissatisfied, you’ll directly edit some passages yourself. Users won’t give the whole report a single score. RL needs a clear right or wrong, yet in real‑world work, most of the time there is no such binary. Whoever can keep learning from these weak signals will obtain the key to post‑deployment self‑evolution.

DeepTech: Since the feedback signal is so weak, how can we tell whether a continuously learning system has truly improved rather than just covering up its original capabilities with local enhancements?

Cai Deng:There really isn’t a good benchmark right now. To evaluate a continual learning system, you need to look at the curve: as deployment time and experience accumulate, does performance on the dedicated task keep improving, and does general capability remain stable; moreover, the evaluation itself must come from real deployments. The first three scaling axes each have their own scaling law, and the fourth axis also needs one: the horizontal axis is no longer parameters and compute, but experience. This is something we must produce next.

DeepTech:RSI is hot in the industry right now, as is using AI to drive AI research (AI for AI). How does that relate to the post‑deployment self‑evolution you mentioned?

Cai Deng: RSI, AI for AI are not new concepts. Front‑line foundational model teams have long been using AI to write code, run experiments, synthesize training data, conduct evaluations, and analyze experimental results. This is something big companies have been doing and will continue to push forward; the level of automation will only deepen. This is not a new track.

People have different understandings of continual learning. Some think that as long as the model iterates fast enough, it counts as a form of continual learning; nowadays, many talk about RSI, which is about letting AI design AI, proposing new model architectures and training algorithms.I believe these are quite different from post‑deployment self‑evolution; they are still pre‑deployment optimizations that improve the average across all deployments, yielding a stronger general‑purpose model.The value of post‑deployment self‑evolution lies precisely in each specific deployment: every individual and organization is different and constantly changing; much expertise can only be gained on the job, and much data cannot leave its own environment. So a more powerful general model is not the answer.

In my view, post‑deployment self‑evolution is a larger and more fundamental issue than RSI or AI for AI. From another angle, AI for AI itself is just a specific deployment whose goal is AI research and development. The real question is, why can AI do AI R&D better? We can compare it to human researchers. The gap between outstanding researchers and ordinary ones is certainly not about having better experimental equipment or better tools for recording and analyzing results; the decisive difference lies in intuition and judgment—that is, taste.

Therefore, what truly matters for AI for AI is cultivating AI’s taste and judgment, which brings us back to how to learn from real work interactions: how to extract effective training signals from observations of the environment and feedback from users and the environment, continuously adjusting the model’s weights. So post‑deployment self‑evolution is the essential issue.

“Jarvis for Everyone”

DeepTech: In the field of continual learning there is also a concept called Lifelong Learning. Is the ultimate goal of ContinuousAI Lab Lifelong Learning?

Cai Deng: Lifelong learning will inevitably be required in the future, and that is our goal. But my understanding of lifelong learning differs from research over the past few decades. Past research focused on having a model learn several tasks sequentially while not forgetting the earlier ones; what we need is for a model, in real deployment, to grow like a human—working and learning simultaneously, with knowledge coming from its own experiences. My lab is called ContinuousAI, because we aim to let AI evolve continuously; all our work ultimately points to this. I believe that in the future, everyone will have their own “Jarvis,” accompanying them throughout their growth.

DeepTech: After entering the AI era, what ability do you value most when recruiting students?

Cai Deng:The development of AI makes some researcher skills less important, while other skills become increasingly important.For example, programming ability may become less crucial in the future, because everyone will be using coding agents like Claude Code.Communication skills will become more important, both with people and with AI. For students, the traits I value most are ambition, daring to experiment, and willingness to experiment.

DeepTech:For master’s and PhD students, is it necessary to do high‑intensity internships in industry?

Cai Deng: It’s necessary. Especially regarding foundational models, academia and industry are very disconnected. Students need to be exposed to the real frontier, first understand what the frontier looks like, where everyone’s focus and bottlenecks are, in order to conduct more cutting‑edge research.

DeepTech: Your current research has great commercialization potential. Will you consider starting a company in the future?

蔡登：Post-deployment self‑evolution is a huge opportunity and a core task of artificial general intelligence (AGI), with a very broad prospect. The needs of each position on the enterprise side differ, and individuals have different personality preferences on the personal side, both offering great application value. But what I care most about is getting the method right first: what constitutes proper internalization, how to learn stably from real, noisy experiences, and how to continue learning without damaging general capabilities. These questions can only be answered in real deployments; they certainly cannot be solved behind closed doors.

Reference material:

https://arxiv.org/abs/2504.01707

https://arxiv.org/abs/2406.16377

https://arxiv.org/abs/2401.11504

https://jcyk.github.io/RetGenTutorial/

Operations/Typesetting: He Chenlong

Note: Cover/featured image generated with AI assistance