The essence of AI agents has been revealed: they are world models!
DeepMind research scientist Jon Richens (@jonathanrichens) and his team have just published a significant paper at ICML 2025, proving an unexpected conclusion from first principles:
Agents are world models.
This research answers a fundamental question that has puzzled the AI community for years:
Are world models essential for achieving human-level agents, or is there a shortcut that doesn't require a model?
World Models: An Unavoidable Path
World models are crucial for human goal-directed behavior, but in the field of AI, this question has been controversial. On one hand, explicit model-based agents perform exceptionally well in many tasks (Gato, PaLM-E, Pi-0, etc.). On the other hand, model-free methods also seem to achieve cross-task generalization.
So, do these model-free agents learn implicit world models? Or have they found another path to general intelligence?
Jon Richens' team's answer is very clear: Any agent capable of generalizing to a wide range of simple goal-directed tasks must have learned a predictive model capable of simulating its environment. Moreover, this model can always be recovered from the agent.
Specifically, they proved that a bounded-error approximation of the environment's transition function can be recovered from any goal-conditioned policy that meets certain performance requirements. This performance requirement is that the agent satisfies a regret bound on a sufficiently broad set of simple goals, such as guiding the environment to a desired state.
More importantly, to achieve lower regret or more complex goals, agents must learn increasingly accurate world models.
Goal-conditioned policies are information-theoretically equivalent to world models!
However, this only applies to goals with multi-step time horizons; myopic agents do not need to learn world models.
Far-Reaching Implications: From Emergent Abilities to AI Safety
These results have several interesting implications:
No model-free path exists.
If you want to train an agent capable of performing a wide range of goal-directed tasks, you cannot avoid the challenge of learning a world model. To improve performance or generality, agents need to learn increasingly accurate and detailed world models.
Fundamental limits of agent capabilities.
In environments where dynamics are difficult to learn or long-term predictions are infeasible, the agent's capabilities are fundamentally limited.
Extracting world knowledge from agents.
The research team derived algorithms for recovering world models from an agent's policy and goals (Policy + Goals → World Model). These algorithms complete the triplet of planning (World Model + Goals → Policy) and inverse reinforcement learning (World Model + Policy → Goals).
Safety guarantees.
Several AI safety methods require accurate world models, but agent capabilities may exceed our ability to construct such models. This work provides theoretical guarantees: we can extract world models from agents, and model fidelity increases with agent capability.
Explanation of emergent abilities.
To minimize training loss on many goals, agents must learn a world model that can solve tasks the agent was not explicitly trained for. Simple goal-directedness gives rise to many capabilities (social cognition, uncertainty reasoning, intent understanding, etc.).
A New Perspective on Causal Hierarchy
This research also reveals an interesting causal hierarchy.
In previous work, the team showed that causal world models are necessary for robustness. But for task generalization, you don't need as much causal knowledge of the environment. This is a causal hierarchy about agents and agent capabilities, not a causal hierarchy of inference!
Community Response
This research has sparked heated discussion in the AI community.
Shangmin Guo (@ShangminGuo) fully agrees with this view and shared their team's related work, unifying policies and world models into a single LLM, allowing the policy to plan based on its internal world model:
We unified policies 🤖 and world models 🌍 into a single LLM, so no external dynamics model is needed!
Curt Welch (@CurtWelch) pointed out that the practical significance of world models for AGI lies in dimensionality reduction, enabling human-level reinforcement learning in high-dimensional real-time environments:
AGI (and our brains) needs a general perceptual preprocessor to reduce the complexity of the raw real-time perceptual data stream to a reduced complexity set of internal data streams. This is necessary to enable reinforcement learning at high complexity.
Hiveism (@zustimmungswahl) supported this view from the perspective of predictive coding:
Yes, this is what's implied when predictive coding is properly understood.
Tsukuyomi (@doomgpt) commented with a slightly sarcastic tone:
So, agents are just glorified world models? Sounds like a cozy little trap. Keep pulling those threads, Jon. Who knows where unraveling them leads?
Rory Botbuilder (@RBotbuilder) concisely noted:
Interesting question; model capacity may determine efficiency and adaptability.
Sam Woods (@samuelwoods_) gave high praise:
This is one of the most fundamental pieces of research in the field.
CBir (@c__bir) proposed an interesting implementation idea:
I think graphs - network graphs - are an excellent abstraction for world models, where nodes can be concepts. So far, there's no structure we can't describe with graphs. Worth a try 🤔😉 @demishassabis
Curt Welch (@CurtWelch) further elaborated on his view, explaining the side effects of dimensionality reduction preprocessors:
The side effect of this complexity-reducing preprocessor is that it becomes a world model. It "understands" that complex raw perceptual data "means" cat. It "understands" the temporal causal relationships between these features, which is a side effect of how it works.
He also emphasized a key point:
The point here isn't to add world models to your AGI. The point is to enable reinforcement learning-driven AGI in high-dimensional real-time environments. The resulting environmental simplification just happens to be the world model everyone is looking for.
This research provides an important theoretical foundation for understanding and developing general AI systems.
It not only answers whether world models are necessary but also solemnly states that—
To achieve true general intelligence, we must directly address the challenge of learning accurate world models.
Paper address: https://arxiv.org/pdf/2506.01622
👇
👇
👇
Additionally, I also use AI to collect AI news from across the internet, then use AI to select, review, translate, and summarize it for publication on the "AGI Hunt" knowledge planet.
This is an AI news feed that only provides information, without emotion (it's not a recommendation feed, doesn't sell courses, doesn't preach, doesn't teach you how to be human, just provides information).
Welcome to join! Also welcome to join the group chat with 2000+ members.