World Model
A World Model refers to a computational model or cognitive framework capable of forming internal representations and simulations of the external environment. Originating in cognitive science, the concept has been widely adopted in artificial intelligence. Core capabilities include perception and modeling of environmental states, reasoning and prediction of causal relationships and dynamic changes, and supporting agents in planning and decision-making without direct interaction with the real environment. World models are regarded as an important component of artificial general intelligence.
Core Facts
The concept of world models originates from cognitive science and reinforcement learning, strongly promoted by Yann LeCun and others, referring to an agent's internal representation of the external environment.
A World Model refers to an AI system's internal construction of representations of environmental states and their evolutionary rules, enabling it to predict the consequences of actions.
The concept of the world model originates from the autonomous machine intelligence architecture proposed by Yann LeCun; its core idea is to enable AI to internally construct a simulator of the operating rules of the physical world in order to predict the consequences of actions.
Timeline (last 90 days)
WOVEN's argument essentially attributes part of MLLMs' deficiencies to a lack of world model capability—that is, although the model can describe image content, it cannot simulate how it evolves with actions.
The paper's core conclusion: closing the gap between agent modeling and object modeling will depend on how the model is trained, not just on how much data it has seen.
Leapmotor's intelligent driving route chooses a world model, not relying on BEV perception, not depending on rule-based code, with lidar only as safety redundancy; there is a foundation world model in the cloud and a real-time world model on the vehicle, which can be deployed at different computing power levels such as 200 TOPS, 640 TOPS and 1280 TOPS, with code volume reduced by 90%.
World models can generate training signals closer to the real distribution, thereby narrowing the gap between simulation and reality and helping with sim-to-real transfer.
A world model predicts what will happen next in the physical world, similar to how a language model predicts the next token.
Currently, the main applications of world models are concentrated in scenarios such as game generation, robot training, and autonomous driving simulation, and their evaluation criteria are usually visual realism, physical consistency, and interaction smoothness.
World models natively support offline learning, allowing agents to learn environment dynamics on collected historical data and plan repeatedly without continuous interaction with the real environment.
By allowing agents to accumulate experience in internal simulations, world models can reduce the need for real-world interactions by one to two orders of magnitude.
The latency and cost of real-time neural rendering are the biggest engineering bottlenecks for bringing such AI-generated world products to market
Neural network-based world models can automatically learn physical rules and visual styles from data, unlike traditional game engines that rely on preset rules and art assets
40 more timeline events
All Facts (20)
The concept of world models originates from cognitive science and reinforcement learning, strongly promoted by Yann LeCun and others, referring to an agent's internal representation of the external environment.
80%VerifiedA World Model refers to an AI system's internal construction of representations of environmental states and their evolutionary rules, enabling it to predict the consequences of actions.
75%VerifiedThe concept of the world model originates from the autonomous machine intelligence architecture proposed by Yann LeCun; its core idea is to enable AI to internally construct a simulator of the operating rules of the physical world in order to predict the consequences of actions.
75%VerifiedA world model refers to a robot's internal predictive capability of the physical environment—given the current state and an action, it can predict what will happen next.
75%VerifiedA world model refers to an internal cognitive representation that enables an agent to predict the consequences of actions, perform causal reasoning, and mentally simulate situations that have not yet occurred.
65%VerifiedWorld models typically need to integrate multimodal data such as video, sensor data, and physical simulations to learn patterns including object motion, causal relationships, and spatial topology.
65%VerifiedThe requirement for temporal coherence in video generation models compels the models to learn dynamics representations closer to the real world during training.
65%VerifiedYann LeCun is Meta's Chief AI Scientist and regards world models as a key component on the path to artificial general intelligence.
65%VerifiedWorld models are essentially learned dynamical equations of an environment—given the current state and action, they predict the next state.
65%UnverifiedBeing able to simulate reality does not equate to truly understanding reality; simulating reality measures more of a surface-level fitting capability rather than an underlying understanding of the world.
70%UnverifiedAtlas is a world model project oriented toward Spatial Intelligence.
70%UnverifiedThe concept of mental models was first proposed by cognitive scientist Kenneth Craik in his 1943 work The Nature of Explanation.
70%UnverifiedDeepMind developed an interactive world model system called Genie.
70%UnverifiedProject Genie is a Foundation World Model project launched by Google DeepMind.
70%UnverifiedYann LeCun believes that current large language models based on the autoregressive Transformer architecture cannot lead to AGI, and that truly general intelligence requires an entirely new world model architecture.
60%UnverifiedThe concept of world models was systematically articulated as early as 1943 by Kenneth Craik in The Nature of Explanation, and was later integrated into the field of reinforcement learning.
60%UnverifiedWOVEN's argument essentially attributes part of MLLMs' deficiencies to a lack of world model capability—that is, although the model can describe image content, it cannot simulate how it evolves with actions.
50%UnverifiedThe paper's core conclusion: closing the gap between agent modeling and object modeling will depend on how the model is trained, not just on how much data it has seen.
50%UnverifiedLeapmotor's intelligent driving route chooses a world model, not relying on BEV perception, not depending on rule-based code, with lidar only as safety redundancy; there is a foundation world model in the cloud and a real-time world model on the vehicle, which can be deployed at different computing power levels such as 200 TOPS, 640 TOPS and 1280 TOPS, with code volume reduced by 90%.
50%UnverifiedWorld models can generate training signals closer to the real distribution, thereby narrowing the gap between simulation and reality and helping with sim-to-real transfer.
50%