World Models + Planning

📋Page Status

Page Type:ContentStyle Guide →Standard knowledge base article

Quality:54 (Adequate)⚠️

Importance:62 (Useful)

Last edited:2026-01-28 (3 weeks ago)

Words:2.2k

Structure:

📊 21📈 1🔗 5📚 38•7%Score: 14/15

LLM Summary:Comprehensive analysis of world models + planning architectures showing 10-500x sample efficiency gains over model-free RL (EfficientZero: 194% human performance with 100k vs 50M steps), but estimating only 5-15% probability of TAI dominance due to LLM superiority on general tasks. Key systems include MuZero (superhuman on 57 Atari games without rules), DreamerV3 (first to collect Minecraft diamonds from scratch), with unique safety advantages (inspectable beliefs, explicit goals) but risks from reward misgeneralization and mesa-optimization.

Issues (2):

QualityRated 54 but structure suggests 93 (underrated by 39 points)
Links1 link could use <R> components

Quick Assessment

Dimension	Assessment	Evidence
Current Dominance	Low (games/robotics only)	MuZero, DreamerV3 superhuman in games; limited general task success
TAI Probability	5-15%	Strong for structured domains; LLMs dominate general tasks
Sample Efficiency	10-500x better than model-free	EfficientZero: 194% human performance with 100k steps (DQN needs 50M)
Interpretability	Partial	World model predictions inspectable; learned representations opaque
Compute at Inference	High	AlphaZero: 80k positions/sec vs Stockfish’s 70M; relies on MCTS search
Scalability	Uncertain	Games proven; real-world complexity unproven at scale
Key Advocate	Yann LeCun (Meta)	Argues LLMs are “dead end”; JEPA is path to AGI

Key Links

Source	Link
Official Website	worldmodels.github.io**
Wikipedia	en.wikipedia.org

Overview

World models + planning represents an AI architecture paradigm fundamentally different from large language models. Instead of learning to directly produce outputs from inputs, these systems learn an explicit model of how the world works and use search/planning algorithms to find good actions.

This is the paradigm behind AlphaGo, MuZero, and the approach Yann LeCun advocates with JEPA (Joint-Embedding Predictive Architectures). The key idea: separate world understanding from decision making.

Estimated probability of being dominant at transformative AI: 5-15%. Powerful for structured domains but not yet competitive for general tasks. MuZero (Nature, 2020) achieved superhuman performance across Go, chess, shogi, and 57 Atari games without knowing game rules. DreamerV3 (2023) became the first algorithm to collect diamonds in Minecraft from scratch, demonstrating generalization across 150+ diverse tasks with fixed hyperparameters.

Architecture

Loading diagram...

Key Components

Component	Function	Learnable
State Encoder	Compress observations to latent state	Yes
Dynamics Model	Predict how state changes with actions	Yes
Reward Model	Predict rewards/values	Yes
Policy Network	Propose likely good actions	Yes
Value Network	Estimate long-term value	Yes
Search/Planning	Find best action via lookahead	Algorithm (not learned)

Key Properties

Property	Rating	Assessment
White-box Access	PARTIAL	World model inspectable but learned representations are opaque
Trainability	HIGH	Model-based RL, self-play, gradient descent
Predictability	MEDIUM	Explicit planning, but world model errors compound
Modularity	MEDIUM	Clear separation of world model, policy, value
Formal Verifiability	PARTIAL	Planning algorithm verifiable; world model less so

Safety Implications

Advantages

Advantage	Explanation
Explicit goals	Planning objectives are visible
Inspectable beliefs	Can examine what the world model predicts
Separable components	Can analyze dynamics, rewards, policy separately
Bounded search	Planning depth is controllable
Model-based interpretability	Can ask “what if” questions of world model

Risks

Risk	Severity	Explanation
Goal misgeneralization	HIGH	Reward model may capture wrong objective
World model errors	MEDIUM	Errors compound over planning horizon
Mesa-optimization	HIGH	Planning is explicitly optimization
Deceptive world models	UNKNOWN	Could world model learn to deceive planner?
Instrumental convergence	MEDIUM	Planning may discover dangerous strategies

Major Implementations

Game/Simulation Benchmarks

System	Developer	Year	Domain	Key Achievement	Performance
AlphaGo Zero	DeepMind	2017	Go	Tabula rasa superhuman	100-0 vs AlphaGo Lee
AlphaZero	DeepMind	2018	Chess, Shogi, Go	24hr to superhuman	155 wins, 6 losses vs Stockfish (1000 games)
MuZero	DeepMind	2020	Games (no rules given)	Learned dynamics model	SOTA on 57 Atari games
DreamerV3	DeepMind/Toronto	2023	150+ diverse tasks	First Minecraft diamonds from scratch	SOTA on 4 benchmarks
EfficientZero	Tsinghua	2021	Atari 100k	500x more sample-efficient than DQN	194% mean human performance

Robotics and Real-World Applications (2024)

System	Application	Key Innovation	Status
MoDem-V2	Robot manipulation	Recurrent State-Space Model for collision-free planning	ICRA 2024
DreMa	Robotic imitation	Learnable digital twins with Gaussian Splatting	Research
4D Latent World Model	3D robot planning	Sparse voxel latent space for geometric reasoning	Research
UniSim	Interactive simulation	Real-world simulators from video	ICLR 2024 Outstanding

Broader Architectural Proposals

System	Developer	Scope	Status	Key Claim
JEPA	Meta (LeCun)	General intelligence	I-JEPA, V-JEPA released	Predicts embeddings, not pixels
V-JEPA	Meta	Video understanding	Released 2024	Physical world model
Genie	DeepMind	World generation	Research	Interactive worlds from video
Sora	OpenAI	Video prediction	Limited release	Implicit world model debate

Research Landscape

Foundational Papers

Paper	Year	Venue	Contribution	Citations (2025)
World Models (Ha & Schmidhuber)	2018	NeurIPS	VAE + RNN architecture; training in “dreams”	2,500+
AlphaGo Zero	2017	Nature	Tabula rasa Go mastery	8,000+
AlphaZero	2018	Science	General game mastery	5,000+
MuZero	2020	Nature	Learned dynamics without rules	3,000+
DreamerV3	2023	JMLR	Fixed hyperparameters across domains	500+
EfficientZero	2021	NeurIPS	500x sample efficiency improvement	400+
A Path Towards Autonomous Intelligence (LeCun)	2022	Position	JEPA framework proposal	1,000+

Key Labs and Researchers

Entity	Focus	Notable Contributions
DeepMind	Game AI, planning	AlphaGo/Zero, MuZero, DreamerV3, Genie
Meta FAIR	JEPA, self-supervised	I-JEPA, V-JEPA, VL-JEPA
Yann LeCun	Architectural advocacy	Argues LLMs lack world understanding
Danijar Hafner	Dreamer series	DreamerV1/V2/V3 lead author
Tsinghua	Sample efficiency	EfficientZero, EfficientZero V2
Berkeley	Robotics	Model-based control, MPC

The LeCun Critique of LLMs

Yann LeCun, Meta’s VP and Chief AI Scientist, has been the most vocal advocate for world models as the path to AGI. His 2022 position paper “A Path Towards Autonomous Machine Intelligence” argues:

LLMs are “dead end” - Autoregressive token prediction does not produce genuine world understanding
Prediction in embedding space - JEPA predicts high-level representations, not raw pixels/tokens
Six-module architecture - Perception, world model, cost, memory, action, configurator
Hierarchical planning - Multiple abstraction levels needed for complex tasks

The JEPA Framework

JEPA (Joint Embedding Predictive Architecture) differs fundamentally from generative models:

Aspect	Generative (LLMs, Diffusion)	JEPA
Prediction target	Raw tokens/pixels	Abstract embeddings
Uncertainty handling	Must model all details	Ignores irrelevant variation
Training signal	Reconstruction loss	Contrastive/predictive loss
Information focus	Surface-level patterns	Semantic structure

Meta has released three JEPA implementations: I-JEPA (images, 2023), V-JEPA (video, 2024), and VL-JEPA (vision-language, 2024).

Counter-Arguments

LeCun Claim	Counter	Assessment
LLMs don’t understand	They demonstrate understanding on diverse benchmarks	Partially valid: benchmark performance vs. true understanding debated
Autoregressive is limited	GPT-4/o1 shows complex reasoning	Contested: reasoning may still be pattern matching
Need explicit world model	Implicit world model may emerge in LLMs	Open question: Sora debate suggests possible emergence
JEPA is superior	No JEPA system matches GPT-4 capability	Currently true: JEPA hasn’t demonstrated general task success

Comparison with LLM Approaches

Aspect	World Models	LLMs	Winner (2025)
Planning mechanism	Explicit MCTS search (80k pos/sec)	Implicit chain-of-thought	Context-dependent
World knowledge	Learned dynamics model	Compressed in weights	LLMs (broader)
Sample efficiency	10-500x better (EfficientZero)	Requires billions of tokens	World Models
Generalization	Compositional planning	In-context learning	LLMs (more flexible)
Task diversity	Requires domain-specific training	Single model, many tasks	LLMs
Current SOTA	Games, robotics control	Language, reasoning, code	Domain-dependent
Compute at inference	High (search required)	Lower (single forward pass)	LLMs
Interpretability	Partial (can query world model)	Low (weights opaque)	World Models

Sample Efficiency Comparison

Algorithm	Steps to Human-Level (Atari)	Relative Efficiency
DQN (2015)	50,000,000	1x (baseline)
Rainbow (2017)	10,000,000	5x
SimPLe (2019)	100,000	500x
EfficientZero (2021)	100,000	500x (194% human)
DreamerV3 (2023)	Variable	Fixed hyperparameters

Model-based approaches achieve comparable or superior performance with 2 orders of magnitude less data, critical for robotics and real-world applications where data collection is expensive or dangerous.

Safety Research Implications

Research That Applies Well

Research Area	Why It Applies
Reward modeling	Explicit reward models are central
Goal specification	Planning objectives are visible
Corrigibility	Can potentially modify goals/world model
Interpretability of beliefs	Can query world model predictions

Unique Safety Challenges

Challenge	Description
Reward hacking	Planning will find unexpected ways to maximize reward
World model exploitation	Agent may exploit inaccuracies in world model
Power-seeking	Planning may naturally discover instrumental strategies
Deceptive planning	Could agent learn to simulate safe behavior while planning harm?

Trajectory Assessment

Arguments For Growth (40-60% probability of increased importance)

Factor	Evidence	Strength
Sample efficiency	500x improvement demonstrated (EfficientZero)	Strong
Robotics demand	Physical tasks need prediction; 2024 survey shows growing adoption	Strong
LeCun’s advocacy	Meta investing heavily in JEPA research	Moderate
Compositionality	Planning naturally combines learned skills	Moderate
Scaling evidence	DreamerV3 shows favorable scaling with model size	Moderate

Arguments Against (40-60% probability of continued niche status)

Factor	Evidence	Strength
LLMs dominating	GPT-4, Claude perform well on most general tasks	Strong
World model accuracy	Errors compound over planning horizon	Strong
Computational cost	MCTS is expensive; AlphaZero searches 80k vs Stockfish’s 70M positions/sec	Moderate
Limited generalization	No world model system matches LLM task diversity	Strong
Hybrid approaches emerging	LLM + world model combinations may dominate both	Moderate

Probability Assessment: Paradigm Dominance at TAI

Scenario	Probability	Reasoning
World models dominant	5-10%	Would require breakthrough in scalable world modeling
Hybrid (LLM + world model) dominant	25-40%	Combines strengths; active research area
LLMs dominant (current trajectory)	40-55%	Empirically winning; massive investment
Novel paradigm	10-20%	Unknown unknowns

Key Uncertainties

Uncertainty	Current Evidence	Resolution Timeline	Impact on Safety
Can world models scale to real-world complexity?	DreamerV3 handles 150+ tasks; robotics limited	2-5 years	High: determines applicability
Will hybrid approaches dominate?	Active research (LLM + world model); no clear winner	3-7 years	Moderate: affects which safety research applies
Is explicit planning necessary for AGI?	o1/o3 suggests implicit reasoning works; debate ongoing	2-5 years	High: determines alignment approaches
Can world model accuracy be verified?	No robust methods exist; critical safety gap	5-10 years	Very High: core alignment concern
Will JEPA fulfill LeCun’s vision?	I-JEPA/V-JEPA released but limited impact	2-4 years	Moderate: alternative paradigm
Can planning be made safe?	Power-seeking, deception emerge naturally from optimization	5-15 years	Critical: core alignment problem

Sources and Further Reading

Primary Research Papers

Schrittwieser, J. et al. (2020). “Mastering Atari, Go, chess and shogi by planning with a learned model.” Nature.
Hafner, D. et al. (2023). “Mastering Diverse Domains through World Models.” JMLR.
Ha, D. & Schmidhuber, J. (2018). “World Models.” NeurIPS.
Ye, W. et al. (2021). “Mastering Atari Games with Limited Data.” NeurIPS.
Silver, D. et al. (2018). “A general reinforcement learning algorithm that masters chess, shogi, and Go.” Science.

Position Papers and Surveys

LeCun, Y. (2022). “A Path Towards Autonomous Machine Intelligence.” Position Paper.
“A review of learning-based dynamics models for robotic manipulation.” Science Robotics (2024).
“A Comprehensive Survey on World Models for Embodied AI.” (2024).

Implementations

DreamerV3 GitHub - Official implementation
MuZero Blog Post - DeepMind
V-JEPA - Meta AI

Dense Transformers - Alternative LLM approach
Neuro-Symbolic - Related hybrid approach
Provable Safe - World model verification is central