ICML 2026 Closed-Door Meeting Reveals 10 Major AI Research Trends

What happens to scientific research itself when agents begin to participate in the research process?

Author and source: Quantum Bit

ICML 2026 has concluded.

But the most interesting discussions this year didn't take place in the sessions.

In the hallways of Coffee Break, in front of the poster boards, around the dinner tables—researchers from Frontier AI Labs, top universities, startups, and investment firms around the world are no longer just discussing the papers themselves:

• How much room is left for the next-generation diffusion language models?
• Where are the opportunities in world models?
• What happens to scientific research itself when agents begin to participate in the research process?

These thoughts, not yet included in papers, are often the starting point for the next consensus.

During this ICML, Nine Kun Ventures' Global AI Bridge organized three small-scale closed-door discussions.

Among them are recipients of the ICML Outstanding Paper Award, researchers from Frontier Lab, Huawei’s Genius Young Talent program, and practitioners actively advancing AI in finance, robotics, and scientific discovery.

We have extracted ten of the most significant signals.

They may not yet be consensus, but they are likely to become the most important new benchmarks for AI in the coming years.

In the past, we focused on designing better models; today, we are beginning to rethink how intelligence is created.

From model architecture to data paradigms, and from AI deployment in high-value industries, the most significant shift at this year’s ICML is that AI is moving from pursuing model capabilities toward building truly sustainable intelligent systems.

Takeaway 01: What redefines the next generation of diffusion language models is not only decoding speed, but also the generation method.

Ni Zanlin

First author of the Outstanding Paper award at ICML 2026, undergraduate and Ph.D. graduate of the Department of Automation at Tsinghua University, advised by Associate Professor Huang Gao, recipient of the National Scholarship, Future Scholar Scholarship, SenseTime Scholarship, and others.

This year, two things at ICML made me especially happy:

First, "The Flexibility Trap" received the Outstanding Paper award; second, during oral and poster presentations, I noticed an increasing number of researchers beginning to reconsider the true value and capabilities of diffusion language models.

Our work aims to answer a core question:

What exactly does arbitrary-order decoding bring, and what are its boundaries?

If the advantages of DLM are limited only to the efficiency gains from parallel decoding, its distinction from Multi-token Prediction may not be as fundamental.

Therefore, what truly matters is whether it can support a paradigm of “writing a draft first, then revising already-written text,” similar to human writing—a capability that today’s most popular diffusion-based large language models still lack.

In the future, I am more focused on exploring next-generation diffusion architectures, such as continuous-space diffusion language models (like MIT’s ELF) and Google’s DiffusionGemma based on uniform-state diffusion, which may open new pathways for the development of language models.

Takeaway 02: Four years ago, models defined data; four years later, data began to define models.

Zhao Bo

ICML 2022 Outstanding Paper Award recipient, Associate Professor at Shanghai Jiao Tong University, National Young Talent, former Head of the Data Intelligence Center at Beijing Academy of Artificial Intelligence, Visiting Scholar at Alibaba Embodied Intelligence, Master’s degree from Peking University, Ph.D. from the University of Edinburgh

Four years ago, Data-Centric AI was a niche focus, often questioned for its value; today, it has become a foundational paradigm shared by cutting-edge fields such as large models, embodied intelligence, and AI for Science.

AI is entering a new phase, with models increasingly becoming carriers of data, where data equals intelligence.

For embodied intelligence, the real bottleneck lies not only in data scarcity but also in the coordination mechanism between data and models.

The core of the next phase of competition will not be who has more data, but who can use data to define model architecture, training strategies, and capability boundaries (Data-Driven Model Design).

As embodied intelligence enters the big data era, efficient data utilization and data-driven model design will become the most important research directions in the coming years, reshaping how the next generation of physical AI systems are built.

Takeaway 03: AI is entering a new phase in finance, moving from technical validation to commercial realization.

Chen Xi

Andre Meyer Chair Professor of Finance at New York University’s Stern School of Business, Adjunct Professor in the Department of Computer Science and the School of Data Science, Head of Artificial Intelligence for Quantitative Research at J.P. Morgan, Ph.D. from CMU, Postdoctoral Fellow at Berkeley, advised by Michael I. Jordan

This year, an exciting development at ICML is that AI and finance are truly converging.

AI in finance is transitioning from technical validation to commercial realization, becoming one of the most valuable application scenarios.

Foundation models, inference models, and agents are transforming Alpha Research, market microstructure modeling, and automated research workflows.

I believe what’s truly noteworthy isn’t just AI companies entering finance, but the growing number of top researchers with Frontier AI backgrounds applying cutting-edge large model research to investment analysis, risk management, trading systems, and other areas.

Over the next five years, AI-native financial institutions and research teams are expected to become one of the most significant areas for AI commercialization.

If AI were a new operating system, what would its "kernel" be?

The next three takeaways provide three responses:

An autonomously planned, continuously evolving agent; the first intelligent system to free scientific accumulation from the constraints of biological lifespan; and an interactive architecture that listens and responds in real time, just as humans do with each other.

Takeaway 04: What truly determines the value of a world model is not the model itself, but the agent.

Wu Jiahong

Head of the DreamX World world model at Alibaba, former Senior Algorithm Expert in Multimodal Understanding at Kuaishou's MMU, former Technical Expert at Innovation Works' Chiqi Intelligence, Master's graduate from Peking University

The endpoint of agents is not merely a tool, but the new infrastructure of the AI era.

The biggest change over the next year won't just be that agents themselves become stronger, but that they will be deeply integrated with world models, AIGC, and embodied intelligence.

Currently, digital world models still face challenges in long-term consistency, memory capacity, and stable output, while agents' capabilities in autonomous planning, persistent memory, and multi-agent collaboration hold promise as key pathways to overcoming these bottlenecks.

In the field of physical world modeling, I believe the greatest opportunity lies not in creating a single, all-encompassing model, but in building specialized capabilities for various vertical use cases.

Since a general world model will still struggle to cover all downstream tasks in the short term, this creates a window of opportunity for entrepreneurs in various niches.

However, this window won’t last long; over the next one to two years, as foundational models continue to improve, general-purpose models will steadily expand their capabilities and gradually encroach upon various domains.

The future competition in AI will no longer be limited to model capabilities, but rather who can build agent systems capable of autonomous perception, continuous decision-making, autonomous execution, and ongoing evolution.

The agent will also become a key infrastructure connecting foundational models to the real world and driving the intelligent transformation of industries.

Takeaway 05: AI4AI × AI4Science: The New Operating System for Scientific Breakthroughs

Liu Rui

Director of the Xiao Yi Large Model Application Laboratory at Huawei's Hong Kong Research Institute, Ph.D. from CUHK's MMLab, participant in Huawei's "Genius Youth" program, currently engaged in cutting-edge research in multimodal and agent technologies, with practical applications deployed at scale for hundreds of millions of end users.

The ultimate paradigm of scientific breakthroughs is not just creating smarter models, but building systems that enable faster scientific discoveries.

This will also lead to a deep convergence of AI for AI and AI for Science.

In the future, the advancement of human civilization will come from the top 1% of scientists, leading countless tireless AI agents.

The integration of AI4AI and AI4Science will overcome a major bottleneck in scientific exploration—data silos and domain barriers—essentially enabling the decentralization and reintegration of knowledge across all fields.

In the past, scientific accumulation relied on the limited lifespan of humans, resulting in significant decay (in books, education, and experience);

In the future, AI agents will be able to transmit knowledge without loss and continuously inherit experience, giving scientific breakthroughs nearly unlimited cumulative potential for the first time.

This ensures that the accumulation of scientific knowledge and experience is no longer limited by the death of individual organisms, allowing technological advancement to continuously reach new heights.

Takeaway 06: The next generation of AI interaction isn't just about "completing tasks"—it's about completing them comfortably.

Ma Ziyang

Ph.D. jointly awarded by Shanghai Jiao Tong University and Nanyang Technological University, and a Puri Scholar at Shanghai Chuangzhi Institute. Recipient of the Interspeech 2023 Best Student Paper nomination, Best Student Paper at the Nanyang Speech Technology Forum, and has published over ten first-author papers with more than 5,900 citations. His published LSLM was among the earliest to explore the modeling paradigm for full-duplex speech interaction. He led the open-sourcing of the emotion2vec and SLAM-LLM series, and is a core contributor to the Qwen-Omni series.

One of the most exciting things at this year's ICML was OpenAI's release of GPT-Live during the conference.

I’m more concerned with the signal it sends: that the next generation of human-computer interaction is gradually becoming a consensus.

During the conference, many researchers and investors discussed the same question: What will be the next-generation interaction paradigm for AI?

I'm pleased to see that Mark Chen, Chief Research Officer at OpenAI, also mentioned during his ICML live talk that interaction will be a key focus area for them over the coming year.

The greatest technical challenge for the next generation of AI is no longer just model capability, but a full-duplex, native interaction architecture.

In the future, whether AI runs on a PC, headphones, glasses, a robot, or an immersive space, it must be able to listen, understand, and respond in real time—just as humans do in conversation—rather than waiting for one statement before giving one reply.

Focusing on this trend, I am most interested in three areas: AI-native interaction architectures, innovations in hardware and hardware-software collaboration, innovations in the human-AI relationship (such as long-horizon tasks and model memory capabilities), and the continuous self-evolution of models during human-AI collaboration.

The next generation of AI, whether in model architecture or product experience, will inevitably give rise to new native interaction paradigms.

AI is no longer just an object of scientific research, but has begun to participate in scientific research.

From AI4Science to AI Research Workflow, and further to Evaluation and Verification, more researchers this year are beginning to discuss:

When AI begins to assist scientists in discovering knowledge, how should we redefine research itself?

Takeaway 07: The true bottleneck in recursive self-improvement is verification bandwidth.

Fu Jie

IQuest Research Scientist, 2021 ICLR Outstanding Paper, 2024 NAACL Outstanding Paper, Ph.D. from National University of Singapore (advisor: Tat-Seng Chua), Postdoctoral Researcher at Mila (advisor: Yoshua Bengio)

A series of articles from ICML 2026 (particularly the position: AI Agents Should Be Evaluated as Behavioral Systems) remind us that:

Evaluating an agent cannot rely solely on final task scores; it requires systematic observation, perturbation, and interpretation of its behavior.

This is especially critical for recursive self-improving AI—each cycle of self-generated data, trajectory filtering, and strategy updates may improve measurable metrics, but does not necessarily mean the system has become more reliable or better aligned.

This directly echoes the core assessment of "Some Simple Economics of AGI":

When AI reduces measurable execution costs to near zero, the bandwidth for human verification, auditing, and accountability guarantees will become the new hard constraint, creating an ever-widening Measurability Gap between what can be executed and what can be verified.

Therefore, the security issue with RSI is not just about how to measure whether the model is becoming stronger, but how to verify whether each improvement is genuine and whether anomalous strategies are being amplified during iterations.

Takeaway 08: When models begin to discover science, evaluation must first become a science.

Liu Hongxuan

Bachelor's degree from Tsinghua University; first-year Ph.D. student in Computational Science and Engineering at MIT; first author of the best paper at the ICML 2026 GenBio Workshop.

Over the past few years, progress in AI for Science has often been defined by higher prediction accuracy, larger models, and more impressive benchmark numbers.

But as AI gradually integrates into real scientific research processes, the rigor and fairness of evaluation are becoming key issues determining whether the technology can truly be deployed.

In scientific tasks, "high scores" may result from data leakage, unreasonable data splits, disparities in computational resources, or inadequate comparison with strong baselines.

A model performing well on random splits does not mean it can handle new molecular systems, material spaces, or experimental conditions.

In the next phase, AI4Science will require more than just larger-scale pretraining, more sophisticated architectures, or harder benchmarks—it needs evaluation systems that better reflect the scientific discovery process: unified comparison protocols, transparency in data and computational costs, and emphasis on out-of-distribution generalization, experimental validation, reproducibility, and failure cases.

When models begin to participate in formulating hypotheses, designing experiments, and drawing scientific conclusions, evaluation becomes more than just a leaderboard—it becomes part of scientific credibility.

Whoever establishes a more rigorous and fair evaluation standard will be the one to truly define the boundaries of AI for Science.

Takeaway 09 Robot Safety: It’s not just about preventing accidents—it’s about building trust.

Li Jiachen

Assistant Professor at Georgia Tech, Postdoctoral Fellow at Stanford University, Ph.D. from the University of California, Berkeley, Director of the Trustworthy Autonomous Systems Laboratory (TASL), Co-Chair of the IEEE RAS Technical Committee on Robot Learning, Associate Editor for IEEE Transactions on Robotics and IEEE Robotics and Automation Letters.

This year at ICML, one of the most striking impressions I had was that safety has gradually evolved from a niche research direction into a core issue of shared concern across the entire AI community.

From large language models to agents, and now to foundation models, an increasing number of efforts are focusing on whether models are reliable, aligned with human intentions, and how to maintain stable and safe behavior in open-world environments.

However, this also leads me to continually ponder a question: When AI steps out of the virtual world and gains a physical form, how should safety be redefined?

In the past, when discussing robot safety, we focused more on physical safety, such as how to avoid collisions, ensure stable control, and guarantee that robots do not harm people under various abnormal conditions.

These questions have always been fundamental to robots entering the real world and have been the most important topic in robotics safety research for decades.

But as robots increasingly enter human environments such as homes, hospitals, and office buildings, I increasingly feel that simply “not harming people or colliding” is no longer sufficient to address the question of safety.

Even if a robot always complies with all safety constraints, people may still feel tense or instinctively move away if it suddenly accelerates, approaches from behind a person, or makes unpredictable movements.

This sense of insecurity does not come from real risks, but from a lack of trust between humans and robots.

Therefore, I believe the focus of future Robot Safety research will gradually shift from Physical Safety to Perceived Safety.

Robots must not only be truly safe but also make people feel safe; they must not only avoid danger but also ensure their behavior is always understandable, predictable, and aligned with human psychological expectations.

I believe this is a noteworthy research direction in the era of Physical AI.

For robots to truly integrate into human society, it will depend not just on better algorithms, higher precision, or faster reasoning speeds, but on the ability to establish ongoing, natural trust between humans and robots.

The ultimate question a robot must answer is not just “Can I do this safely?” but “Will people feel comfortable when I do this?”

Takeaway 10: The Agent may change not just AI, but the way we conduct research.

Jeremy

Top global researcher at a Frontier AI Lab, graduated from Stanford University, with long-term research experience in large language models (LLMs), agents, and machine learning systems, focusing on AI reasoning, agent systems, and the application of AI in complex real-world scenarios.

After attending ICML 2026, I've been wondering what the biggest change this year really was.

It’s not that any single model suddenly became stronger, nor that any single benchmark has been broken again—it’s that I feel an increasing number of studies now treat agents as an integral part of the research process, rather than merely the subject of study.

In the past, much of the work focused on designing better models: proposing new architectures, improved training methods, and higher scores.

This year, many discussions have gradually shifted to another level—what if models can proactively plan, invoke tools, search for information, write code, and run experiments? How would that change the nature of research itself?

This impressed me.

Previously, we viewed "training models" and "conducting research" as two separate activities, but now the boundary between them is becoming increasingly blurred.

An increasing number of workflows are integrating experimental design, data processing, hypothesis testing, result analysis, and even code generation into a single cycle.

Models are no longer just providing a final answer but are now actively participating in the entire research process.

I believe this shift is more noteworthy than the model's capabilities themselves, as it means that in the future, when discussing a research achievement, we may no longer focus solely on the model, but rather on how the entire research process was carried out:

Which steps can be automated, which decisions still require human input, which环节 are most prone to bias, and how should new collaborative processes be designed?

Another interesting observation is that, this year, when chatting with many researchers, the most discussed topic wasn't any specific model parameter or training technique, but rather workflow.

How to design experiments, how to quickly validate an idea, and how to connect different tools have become the focus of many discussions.

The agent is more about redefining the relationship between researchers and tools, rather than simply adding a new AI capability.

If I were to summarize the biggest takeaway from ICML 2026, I wouldn't say that agents will replace researchers, but rather that they are redefining how research is organized.

What truly matters in the future may not be who owns the most powerful model, but who can build a more efficient human-AI collaboration process, enabling better ideas to be validated, iterated upon, and ultimately transformed into truly impactful research.

What truly matters is not just the paper, but the next consensus.

Over the past few years, each ICML has produced several classic papers.

What truly drives AI forward is often not a single paper, but a group of researchers beginning to focus on the same problem together.

These directions may not yet be consensus views, but they are likely to be the most important new benchmarks for Frontier AI in the coming years.

Global AI Bridge aims to continuously document the latest insights from frontier AI researchers worldwide, foster ongoing open discussions around these topics, and connect researchers, builders, founders, and investors to enable meaningful cross-institutional and cross-border collaboration.

Over the next year, we will continue to launch a series of open discussions (Open Topics) focused on these research areas, including Agent, World Model, Data-Centric AI, AI Evaluation, and AI4Science.

If you are also researching these topics, wish to participate in topic discussions, or want to co-initiate new discussions, please contact the Global AI Bridge coordinator.

Perhaps the next "ICML Top Takeaways" will come from the next discussion.