TLDR: This paper critiques current approaches to “World Models” in AI, which aim to simulate real-world environments for intelligent agents. It argues that many existing models overemphasize video generation, rely too heavily on continuous representations, and use objectives that can lead to issues. The authors propose a new “Physical, Agentic, and Nested (PAN)” architecture that leverages multimodal data, mixed continuous and discrete representations, hierarchical generative modeling with an LLM backbone, and observation-grounded training, enabling more robust and general-purpose AI reasoning and planning, exemplified by a mountaineering task.
In the quest to build truly intelligent artificial agents, the concept of a ‘World Model’ has emerged as a central idea. A World Model is essentially an algorithmic stand-in for the real world, allowing AI agents to simulate possibilities, reason, and act purposefully without direct interaction with the environment. This capability, akin to the ‘hypothetical thinking’ seen in humans, is crucial for developing AI that can perform a wide range of complex tasks, from navigating physical spaces to strategizing in social situations.
The paper, ‘Critiques of World Models’, delves into the ongoing debates surrounding what a world model truly is, how it should be built, used, and evaluated. It draws inspiration from science fiction, like Frank Herbert’s ‘Dune’, where a character can simulate all possible futures to make optimal decisions, highlighting the aspirational goal for AI.
Current approaches to building world models often emphasize video and image generation, focusing heavily on visual quality. Systems like Genie 2 for gaming, World Labs for 3D scenes, Wayve GAIA-2 for physical control, and general video generation models like Sora, all prioritize visual output. However, the authors argue that this focus might be a misunderstanding. They contend that a world model’s primary purpose is not to generate visually stunning videos, but to serve as a ‘sandbox for reasoning and thought-experiment’.
The paper critically examines several prevailing ideas in world modeling across five key dimensions: data, representation, architecture, objective, and usage. For instance, some argue that sensory inputs (like video) are superior to text due to sheer data volume. The authors counter that while sensory data is vast, much of it is redundant, whereas natural language is a highly compressed and semantically rich form of human experience, encoding abstract concepts like justice or motivation that are difficult to derive from raw perception alone. They advocate for leveraging all modalities of experience, recognizing that each offers different layers of information.
Another point of contention is whether world states should be represented by continuous embeddings (for gradient-based optimization) or discrete tokens. The paper suggests a ‘mixed representations’ approach, combining discrete tokens for robust, interpretable, and symbolic reasoning (like human language) with continuous embeddings for capturing fine-grained sensory nuances. They even provide theoretical arguments showing that discrete representations can capture arbitrarily fine distinctions if scaled appropriately, particularly by increasing sequence length rather than vocabulary size.
Regarding architecture, some propose abandoning autoregressive, generative models in favor of joint embedding predictive architectures (JEPA) to avoid error accumulation. However, the paper argues that JEPA is still functionally autoregressive and generative. Instead of abandoning generative models, they propose a ‘Generative Latent Prediction (GLP)’ architecture. This hierarchical approach uses different layers for different granularities: latent diffusion models for continuous perceptual data and an enhanced Large Language Model (LLM) backbone for discrete, long-horizon structures. This design aims to mitigate error compounding and enable more expressive reasoning.
The choice of training objective is also scrutinized. While some prefer latent-reconstruction objectives (predicting future latent embeddings), the paper warns that these are prone to ‘collapse’, where the model learns nothing meaningful. They advocate for ‘generative reconstruction loss’, which grounds the learning objective in observable data by reconstructing the actual next observation. This provides a more stable and interpretable training signal, ensuring that learned representations are meaningfully tied to the real world.
Finally, the paper discusses how world models should be used for decision-making. Some favor Model-Predictive Control (MPC) for its sample efficiency and safety. However, MPC can be computationally expensive during inference and typically plans only a few steps ahead. The authors propose using world models within a Reinforcement Learning (RL) framework, where the agent learns from an infinite space of imagined worlds simulated by the world model. This shifts computational cost to the training phase, allowing for more strategic, long-term reasoning and flexible decision-making, akin to how humans plan and adapt.
Building on these critiques, the paper previews a new architecture called ‘PAN’ – a Physical, Agentic, and Nested world model. PAN is designed to be a truly general-purpose world model, integrating data from all modalities, employing mixed continuous and discrete representations, utilizing a hierarchical generative model with an enhanced LLM backbone and a diffusion-based predictor, training with generative loss grounded in observation data, and applying the model to simulate experiences for training agents via reinforcement learning. The authors illustrate PAN’s potential with a complex mountaineering expedition use case, demonstrating its ability to handle multimodal inputs, multi-scale reasoning, and long-term planning.
Also Read:
- Beyond Language and Vision: Why AI Needs to Master the Art of Movement
- LTLCrit: Enhancing LLM Agents with Formal Logic for Safe and Efficient Embodied Planning
PAN processes sensory inputs through a ‘Sensory Encoder’ that maps signals into both discrete tokens (for abstract concepts) and continuous embeddings (for nuanced perceptual details). A ‘World Model Backbone’ (combining LLM and diffusion models) predicts future world states based on proposed actions. A ‘Multimodal Decoder’ then reconstructs the next observation, ensuring the model’s predictions are grounded in sensory reality. This comprehensive framework aims to provide a rich internal sandbox for simulation, experimentation, and foresight, enabling AI agents to plan ahead, navigate uncertainty, and act with flexibility and intent. For more details, you can refer to the full research paper here.
