---
kind: concept
slug: world-model
title: World model
summary: >-
  A learned model that predicts how a world changes after an action, so an agent can try options first. Variants: latent planners, hierarchical controllers, learned rewards, search over imagined futures, goal-conditioned models, driving forecasts, text simulators, models of other minds and typed shared state for multiplayer worlds. Physical consistency and long-rollout fidelity remain open.
tags: [model-based-rl, video-generation, embodied-ai, simulation, llm-agents, autonomous-driving, planning, multiplayer]
papers: ["1803.10122", "2311.17918", "2405.18418", "2408.14197", "2411.02446", "2411.14499", "2510.16729", "2512.02417", "2512.03556", "2512.18477", "2512.18832", "2512.18850", "2601.02378", "2608.06257"]
related: [model-based-reinforcement-learning, learned-simulator, model-predictive-control, diffusion-models, latent-diffusion, video-generation, llm-agent, neural-scaling-laws, bev-representation, imitation-learning, variational-autoencoder, neuroevolution, hierarchical-reinforcement-learning, monte-carlo-tree-search, intrinsic-motivation, vision-language-action-model, theory-of-mind, reward-design, goal-conditioned-reinforcement-learning, farthest-point-sampling]
status: draft
updated_at: 2026-11-01
---

## Overview

A world model predicts how a situation changes when an action is applied
[arxiv:2411.14499 §3.1]. The survey separates two jobs. The first is an
implicit representation of how the world works. The second is predicting
future states, so the model can simulate outcomes for decisions or for data
generation [arxiv:2411.14499 §1] [arxiv:2411.14499 §2]. The survey says the
definition is still contested [arxiv:2411.14499 §1].

The invariant the variants share is the one-step transition: a model that maps
a state and an action to the next state, and can be rolled forward to look
further ahead [arxiv:2411.14499 §2, Categorization]. A text-environment variant
keeps this form but outputs the next observation as text and adds a binary
success signal [arxiv:2512.18832 §3.1] [eq. 2]. Some robot variants also
predict a done signal, and use it as a reward or as a termination test
[arxiv:2512.03556 eq. 1] [arxiv:2405.18418 eq. 1].

This page draws on fourteen cards. The survey supplies the family taxonomy and
the video and embodied variants. The World Models paper supplies the first
end-to-end example, a latent model with an evolved controller. The
text-environment study supplies the one empirical result on text environments.
The two occupancy papers and a multiview video generator supply the driving
forecast variants, with author-reported results on nuScenes. Six newer cards add variants in
simulation: a hierarchical humanoid controller, a learned reward inside a
video model, search over imagined futures for manipulation, a decoupled
driving model for changed vehicles, reward-free pretraining for driving, and a
goal-conditioned model trained on episodes between subgoals. A multiplayer
study adds a variant that keeps one typed world state shared by all camera views.
A survey card covers models of other agents' minds. The survey traces a
historical line from Ha and Schmidhuber (2018) and the Dreamer series to video
generators [arxiv:2411.14499 §2] [arxiv:2411.14499 §3.1]. The World Models
paper has a card and is covered below. The Dreamer series has no card that
lists this slug, so it is only named here.

## Variants

### Latent dynamics model for planning

The model is trained on real transitions and then used to choose actions. The
survey casts learning as supervised fitting, either with a squared error or a
probabilistic loss [arxiv:2411.14499 eq. 1] [arxiv:2411.14499 eq. 2]. Planning
uses the model in three ways: model predictive control over sampled action
sequences, Monte Carlo tree search over learned values, and trajectory
optimisation in latent space [arxiv:2411.14499 §3.1]. Card:
[Understanding World or Predicting Future? A Comprehensive Survey of World Models](/papers/2411.14499).

### Latent model with an evolved controller

The World Models agent has three parts. A VAE encodes each frame into a latent
z. A recurrent mixture-density model predicts the next z from the action and
its hidden state. A single linear controller maps both to an action
[arxiv:1803.10122 §2] [arxiv:1803.10122 eq. 1]. Only the controller sees the
reward. The two world-model parts are trained without it [arxiv:1803.10122 §3.1],
and the controller is evolved with CMA-ES, which needs one final score per
rollout [arxiv:1803.10122 §2.3] [arxiv:1803.10122 §A.4]. The Car Racing
controller has 867 weights [arxiv:1803.10122 tab. 1]. The authors report
906 ± 21 over 100 trials, against 838 ± 11 for a Gym leaderboard entry
[arxiv:1803.10122 tab. 2]. Card: [World Models](/papers/1803.10122).

The same model can also serve as a dream. For VizDoom Take Cover, the
controller is trained inside the recurrent model, which also predicts whether
the agent dies, and the policy is then run on the real game
[arxiv:1803.10122 §4.1] [arxiv:1803.10122 §4.4]. The dream temperature τ
changes the outcome. At τ = 0.10 the virtual score is 2,086 ± 140, but the real
score is 193 ± 58. The authors say the controller exploits the model at low τ.
At τ = 1.15 the real score is 1,092 ± 556 [arxiv:1803.10122 tab. 4]
[arxiv:1803.10122 §4.5]. The paper does not define precisely how τ acts
[arxiv:1803.10122 §2.2] [arxiv:1803.10122 §4.3]. The authors name two limits of
the approach: the VAE keeps detail the task does not need, and the recurrent
memory has limited capacity [arxiv:1803.10122 §7].

### Hierarchical latent model for humanoid control

Two TD-MPC2 world models are trained separately [arxiv:2405.18418 §3]. A
low-level tracker is trained once on noisy rollouts of expert policies for 836
motion-capture clips, then frozen [arxiv:2405.18418 §3.1] [arxiv:2405.18418 §4.1].
A visual high-level planner takes the proprioceptive state and a 64x64 camera
image, and outputs a 15-value command for the frozen tracker. It is trained
from scratch for each task on the task reward [arxiv:2405.18418 §3.2]
[arxiv:2405.18418 §4.1]. On eight simulated humanoid tasks the authors report
return comparable to TD-MPC2, except on stairs, where TD-MPC2 scores higher
[arxiv:2405.18418 §4.2] [arxiv:2405.18418 fig. 5]. In a 51-person study, 97.8%
of the 510 judgements preferred the hierarchical agent's motion
[arxiv:2405.18418 tab. 6]. The tracker takes about 12 days on one RTX 3090
[arxiv:2405.18418 §4.1]. Card:
[Hierarchical World Models as Visual Whole-Body Humanoid Controllers](/papers/2405.18418).

### Model as a reinforcement-learning environment with a learned reward

Here the world model is the training environment for a policy, and it also
supplies the reward. The action model takes the policy's 7-DoF end-effector
action and returns the next frame and a done signal. A text model writes a
goal frame from the instruction, at the step where its own done signal first
passes a threshold [arxiv:2512.03556 §3.1] [arxiv:2512.03556 eq. 6]. The reward
adds a sparse term, one when the action model's done signal passes a threshold,
and a dense term, the LPIPS distance between the current frame and the goal
frame [arxiv:2512.03556 eq. 5] [arxiv:2512.03556 eq. 7] [arxiv:2512.03556 eq. 8].
The policy is trained with PPO inside the model [arxiv:2512.03556 §4.3]. With
an OpenVLA policy, the authors report 84.1% on out-of-domain pick and place,
against 54.5% with the simulator's reward [arxiv:2512.03556 tab. 1]. In-domain
differences are small [arxiv:2512.03556 tab. 1]. World-model training took
about 96 hours on an H20 cluster [arxiv:2512.03556 §4.2]. All evaluation is in
ManiSkill, with no real robot [arxiv:2512.03556 §5]. Card:
[RoboScape-R: Unified Reward-Observation World Models for Generalizable Robotics Training via RL](/papers/2512.03556).

### Search over imagined futures at decision time

The model is used inside a search loop, not a fixed rollout. A diffusion-based
vision-language-action policy proposes candidate actions at each expanded node.
A video model then imagines the outcome of one candidate, and Monte Carlo tree
search with a PUCT rule selects the root action with the most visits
[arxiv:2512.18477 §3.3] [arxiv:2512.18477 eq. 5]. The policy is CogACT used
without further training. The video model is iVideoGPT-medium, fine-tuned on
Bridge with a reward head [arxiv:2512.18477 §4.1]. Each decision uses 8
candidates and 8 simulations [arxiv:2512.18477 §3.1] [arxiv:2512.18477 §4.1].
On four SimplerEnv WidowX tasks the average success is 51.0%, against 47.9%
for the base policy [arxiv:2512.18477 tab. 1]. The authors call the margin
modest [arxiv:2512.18477 §4.3], and the paper reports no seeds or variance.
Card:
[STORM: Search-Guided Generative World Models for Robotic Manipulation](/papers/2512.18477).

### Decoupled ego-dynamics model for changed vehicles

This variant splits the model. One part is a parameterised vehicle dynamics
model, and the other is an environment model that sees only the ego vehicle's
state [arxiv:2512.02417 §4.1] [arxiv:2512.02417 eq. 9]. A learned policy outputs
a goal, and a predefined PID controller turns that goal into steering and
throttle, so the car's own response is handled by feedback
[arxiv:2512.02417 §4.2.1]. Two steps target changed mass or steering limits.
At deployment, the original policy's action is replaced by the action that
brings a new car to the state the old car would reach [arxiv:2512.02417 §4.3]
[arxiv:2512.02417 eq. 20]. During training, imagined rollouts are run under
both nominal and shifted parameters [arxiv:2512.02417 §4.4]. On MetaDrive
roundabouts the authors report 0.70 success, against 0.53 for DreamerV3. On
intersections the paper's model is slightly lower, at 0.88 against 0.91
[arxiv:2512.02417 tab. 1]. The mass and steering robustness study is shown
only as a figure [arxiv:2512.02417 fig. 9]. Card:
[Vehicle Dynamics Embedded World Models for Autonomous Driving](/papers/2512.02417).

### Reward-free pretraining for driving

A DreamerV3-style world model and an explorer are pretrained in CARLA with no
task reward. The explorer is rewarded by the disagreement between an ensemble of
next-state predictors [arxiv:2512.18850 §3.2.1] [arxiv:2512.18850 §3.4]. The
frozen explorer is then run zero-shot on an unseen town. Success averages
73.20% on Town02, against 48.61% for ICM and 55.83% for RND
[arxiv:2512.18850 tab. 6]. After 10k fine-tuning steps on task reward,
collision avoidance reaches 94.78% on Town02 [arxiv:2512.18850 tab. 6]. The
results are means over five seeds, with no spread shown [arxiv:2512.18850 §5].
The scope is one simulator, one pair of towns, discrete controls and a
semantic-segmentation input [arxiv:2512.18850 §6]. Card:
[InDRiVE](/papers/2512.18850).

### Goal-conditioned model trained on subgoal episodes

The world model is goal-conditioned and is trained on data picked to cover
the transitions that matter for reaching goals. The paper's method, MUN, adds a
second kind of real episode to a goal-conditioned Dreamer-style agent. The
agent visits two subgoals drawn from a set of key states, so the model sees
backward and cross-trajectory transitions. The policy is then trained only on
imagined rollouts, with reward equal to the negative learned temporal distance
[arxiv:2411.02446 §3.1] [arxiv:2411.02446 alg. 1] [arxiv:2411.02446 eq. 7]. The
subgoals come from farthest-point sampling over the actions in a batch of
episodes. The authors use it to pick states where the action changes most, such
as grasping and releasing an object [arxiv:2411.02446 §3.2] [arxiv:2411.02446
alg. 2]. The one-step prediction error is the lowest of the five methods the
paper compares on all six tasks [arxiv:2411.02446 tab. 6]. Over 500 simulated
trajectories, the paper's model has the lowest compounding error on five tasks. On
Fetch Slide, the variant without subgoal selection is slightly lower
[arxiv:2411.02446 tab. 7]. On 3-Block stacking the authors report more than 95%
success, against about 60% for three baselines within 2.5M steps
[arxiv:2411.02446 §4.3]. Each run takes 36 to 70 GPU hours on an A100
[arxiv:2411.02446 tab. 1]. Card:
[Learning World Models for Unconstrained Goal Navigation](/papers/2411.02446).

### Language-model world knowledge

Language models can carry knowledge about space, time, geography, cities and
human behaviour. The survey reports linear representations of space and time,
coarse and sometimes inaccurate urban knowledge, and good results on several
Theory of Mind tasks with exceptions such as detecting faux pas
[arxiv:2411.14499 §3.2]. A second pattern pairs a language model with an
explicit model and a planner or verifier; the survey's examples include PDDL
domains written by a language model and solved by an off-the-shelf planner
[arxiv:2411.14499 §3.1]. Card:
[Understanding World or Predicting Future? A Comprehensive Survey of World Models](/papers/2411.14499).

### Mental world model of other agents

The survey's review of Theory of Mind takes the physical world model and
extends its state with the agent's own memory and with the observed actions of
other agents [arxiv:2601.02378 §2.1]. Two reasoning paradigms are compared.
Prompting a language model is easy to apply but works as a black box. Explicit
model-based inference is interpretable but costly [arxiv:2601.02378 §4]
[arxiv:2601.02378 tab. 5]. The numbers quoted for these methods come from the
cited works and were not re-run. One example is 76.6% accuracy for LIMP on
MuMA-ToM, against 50.6% for GPT-4o [arxiv:2601.02378 §4.2]. The review argues
that most benchmarks are text-only and offline [arxiv:2601.02378 §5]. It runs
no experiments of its own. Card:
[Modeling the Mental World for Embodied AI: A Comprehensive Review](/papers/2601.02378).

### Fine-tuned language model as text-environment simulator

Here the language model is itself the transition model. It is fine-tuned with
supervised learning on logged agent-environment dialogues, and given the
dialogue so far and the agent's latest action, it writes the next environment
message and a success flag [arxiv:2512.18832 §3.1] [arxiv:2512.18832 §4]. The
paper covers ALFWorld, SciWorld, TextWorld, WebShop and StableToolBench. A
fine-tuned Qwen2.5-7B reaches 99.87% exact match on ALFWorld next states and
98.60% on SciWorld. On TextWorld and WebShop it reaches 70.60% and 79.05%, and
on StableToolBench 48.90% exact match, against prompt-only baselines of 77.04%,
73.08%, 49.12% and 66.09% on the first four [arxiv:2512.18832 tab. 1]. Exact
match is strict, and the paper calls its TextWorld figure a lower bound
[arxiv:2512.18832 §4].

Consistency is the harder test. The paper runs an agent in the real environment
and in the simulator, then replays the simulated actions in the real
environment. The ratio of those success rates is around 0.91 to 0.96 on the
three structured environments and 0.67 on WebShop for the Qwen model, averaged
over seven API agents [arxiv:2512.18832 §5.2] [arxiv:2512.18832 tab. 2]. The
simulator is less reliable for agents it was not trained on. The weakest agent
falls below 70% consistency, and mixing data from several agents raises one
agent's ratio from 0.49 to 0.81 [arxiv:2512.18832 §5.2] [arxiv:2512.18832 tab. 3].

The paper also reports three agent uses. Pre-execution checks in WebShop reach
36.70% success at 50 checks, against 29.36% with no checks for GPT-4o
[arxiv:2512.18832 tab. 4]. Synthetic trajectories are competitive with real
ones on SciWorld and WebShop, and a warm-start stage helps final success on
ALFWorld and SciWorld. Both gains appear only as plots, with no numbers in the
text [arxiv:2512.18832 §7] [arxiv:2512.18832 fig. 6] [arxiv:2512.18832 fig. 7].
Simulated success is the model's own prediction, not a check against the real
environment [arxiv:2512.18832 app. A.2]. Card:
[From Word to World: Can Large Language Models be Implicit Text-based World Models?](/papers/2512.18832).

### Video generator as simulator

Video generators such as Sora are treated as the main current route to
world simulation, with the caveat that it is debated whether they count as
world models [arxiv:2411.14499 §4.1]. The desirable capabilities the survey
lists are long-term prediction, multimodal input, interactivity and coverage
of varied environments. Its named limits are weak causal reasoning and
failures on physical laws [arxiv:2411.14499 §4.1]. Card:
[Understanding World or Predicting Future? A Comprehensive Survey of World Models](/papers/2411.14499).

### Embodied and action-conditioned environments

Simulators and generated environments for robots and agents are grouped as
indoor, outdoor, and dynamic action-conditioned environments. The survey notes
that most work so far builds static scenes, and that the trend is toward
dynamic, first-person, action-conditioned worlds that respect physical
constraints [arxiv:2411.14499 §4.2]. Card:
[Understanding World or Predicting Future? A Comprehensive Survey of World Models](/papers/2411.14499).

### Occupancy world model for driving

A camera-only model turns past images into a bird's-eye-view (BEV) state,
forecasts how that state changes, and reads occupancy and a trajectory from it.
Drive-OccWorld uses a BEVFormer-style encoder and an autoregressive decoder
that predicts 3D occupancy and flow over four steps of 0.5 s, which is two
seconds on nuScenes [arxiv:2408.14197 app. C.1] [arxiv:2408.14197 app. C.3].
Its memory is normalised by ego motion [arxiv:2408.14197 eq. 5]. A planner
samples candidate trajectories and keeps the one with the lowest occupancy cost
[arxiv:2408.14197 §3.4]. With the predicted trajectory as the action, the
authors report an average open-loop L2 of 0.85 m, against 1.03 m for UniAD
under the NoAvg protocol [arxiv:2408.14197 tab. 5]. Card:
[Driving in the Occupancy World: Vision-Centric 4D Occupancy Forecasting and Planning via World Models for Autonomous Driving](/papers/2408.14197).

IR-WM changes what the decoder predicts. It outputs the change to the BEV state,
which is added to the previous state [arxiv:2510.16729 §3.2.2] [eq. 6]. An
alignment step then modulates the features with predicted semantics and ego
motion, to limit error build-up over time [arxiv:2510.16729 §3.2.2]. The
authors report that alignment adds 0.2 to 0.4 points of per-class IoU on the
classes they list [arxiv:2510.16729 tab. 8], and an average L2 of 0.53 m for
planning [arxiv:2510.16729 tab. 4]. The forecast can reach the planner in three
ways: tightly coupled, semi-coupled (the default) and fully decoupled
[arxiv:2510.16729 §3.2.4]. Card:
[Vision-Centric 4D Occupancy Forecasting and Planning via Implicit Residual World Models](/papers/2510.16729).

The two occupancy papers use nuScenes with occupancy labels [arxiv:2510.16729 §4.1.1]
and train on 8 A100 GPUs [arxiv:2408.14197 app. C.3] [arxiv:2510.16729 §4.1.3].
Their cards record gaps between the paper and the released code. Drive-OccWorld's
released configs cannot switch the planner on, and IR-WM's default training path
appears to fail as read. Check the cards' Paper vs code sections before you
reproduce a planning number.

### Multiview video diffusion model with image-based planning rewards

Drive-WM generates the six surround cameras of nuScenes as one video. It uses a
latent video diffusion model, conditioned on layout tokens, a text sentence and
ego-motion tokens [arxiv:2311.17918 §3.3] [arxiv:2311.17918 eq. 5]. It starts
from a single-view image model, adds temporal layers and then multiview layers,
and freezes the spatial weights while it trains the rest [arxiv:2311.17918 §3.1].
To keep overlapping views consistent, it generates three reference views first
and then three stitched views, each conditioned on its neighbours
[arxiv:2311.17918 §3.2]. For planning, a pre-trained planner (VAD) proposes
candidates. The model renders each candidate's future, and a reward multiplies a
map term and an object term, both measured by perception models run on the
generated images [arxiv:2311.17918 §4.1] [arxiv:2311.17918 §4.2]. The reward
weights and thresholds are not stated [arxiv:2311.17918 §4.2].

The authors report FID 15.8 and FVD 122.7 for multiview video, against 16.4 FID
for their factorized variant, which raises the matched-point ratio (KPM) from 45.8% to 94.4%
[arxiv:2311.17918 tab. 1] [arxiv:2311.17918 tab. 2]. In open-loop planning, the
average L2 is 0.80 m and the average collision rate 0.26%, against 1.02 m and
0.93% for VAD with a random command [arxiv:2311.17918 tab. 3]. Most of the gain
comes from the object term alone, which gives the same average L2 at a slightly
higher collision rate [arxiv:2311.17918 tab. 4]. Training used an A40 with 48 GB
and 40k iterations at batch 32 [arxiv:2311.17918 app. B.1]. The paper reports no
runtime per candidate. Card:
[Driving into the Future: Multiview Visual Forecasting and Planning with World Model for Autonomous Driving](/papers/2311.17918).

### Typed shared state for multiplayer worlds

Several players share one world, and each camera is drawn from one state.
A learned Logic Engine advances that state by one tick from all players'
actions. It is a decoder-only Transformer that predicts each typed record
under a schema mask [arxiv:2608.06257 §3.2] [arxiv:2608.06257 eq. 5]. A
separate residual U-Net, the Rendering Engine, draws each requested camera
from the same predicted state [arxiv:2608.06257 §3.3] [arxiv:2608.06257 eq. 7].
Because every view decodes from that one state, cross-view disagreement is zero
by construction, and the authors say the metric therefore says little about
the renderer [arxiv:2608.06257 §4.5]. In the 1,024-snake Snake setting the
state has 5,121 records per tick [arxiv:2608.06257 §3.1].

On a matched two-player, two-view Snake benchmark, the authors report parser
state recovery of 0.764. The best video baseline, B-PV, scores 0.128, and
MultiWorld scores 0.067 [arxiv:2608.06257 tab. 2]. The same table gives an
invalid rate of 0.177, against 0.052 for the two dense baselines
[arxiv:2608.06257 tab. 2]. Full-state exact match on Snake is 41.7% at one
tick and 0.0% at 32 and 128 ticks [arxiv:2608.06257 app. B.2]. A typed step
takes 45.55 ms, against 3.07 to 3.94 ms for the dense baselines
[arxiv:2608.06257 tab. B.4]. Rendering is reported on eight games from
held-out reference states, with PSNR from 23.72 to 40.24 dB
[arxiv:2608.06257 tab. 3]. The paper's statement of 10,000 recurrent ticks for
1,024 entities has no table in the appendix that the card could find. Its
stress runs go up to 2, 4 and 8 snakes [arxiv:2608.06257 app. C.3]. The
MultiWorld baseline has 5.60B trainable parameters and MASS's predictor has
5.66M, so the comparison is not parameter-matched [arxiv:2608.06257 app. B.1].
Card:
[MASS: Multiplayer World Models with Authoritative Shared State](/papers/2608.06257).

## When to use which

The survey gives no measured comparison between its families, and it states
no data or compute requirements for them. The text-environment study reports
numbers for one model family on four text environments. The driving papers
report numbers on one dataset, and their latency figures use different devices.
The multiplayer study's headline comparison uses one game with two players.
The simulation papers use different simulators (ManiSkill, CARLA, MetaDrive,
SimplerEnv, a humanoid simulator and the six goal-conditioned tasks of the
MUN paper), different metrics and different
baselines, so their gains cannot be ranked against each other. The table
reflects only what each source says.

| Situation | Use | Why |
|---|---|---|
| You can learn the environment's transitions from data and need to choose actions by lookahead | Latent dynamics model with planning | The model scores candidate action sequences by predicted return over a horizon [arxiv:2411.14499 eq. 3] [arxiv:2411.14499 §3.1] |
| Observations are pixels, a very small policy is acceptable, and you can afford many rollouts for a gradient-free search | World Models controller evolved with CMA-ES | The controller has 867 weights, and the search runs about 1,800 generations of 64 candidates, each scored on 16 rollouts [arxiv:1803.10122 tab. 1] [arxiv:1803.10122 §A.4] |
| Real rollouts are costly and you want to practise in the learned model before transfer | Dream training with a tuned temperature, checked on the real environment | Real scores ranged from 193 to 1,092 across the τ values reported, so the choice of τ decides the outcome [arxiv:1803.10122 tab. 4] |
| A humanoid must move from pixels with human-like motion, and you can pretrain a general tracker once and reuse it | Hierarchical latent model with a frozen tracker | Return is comparable to TD-MPC2 on eight tasks except stairs, and motion was preferred in the authors' study. The tracker alone takes about 12 days on one RTX 3090 [arxiv:2405.18418 §4.1] [arxiv:2405.18418 §4.2] [arxiv:2405.18418 tab. 6] |
| You have a robot policy and a way to score goal-like frames, and you want to avoid writing a reward per task | Model as RL environment with a learned goal-frame reward | Out-of-domain pick and place reaches 84.1% against 54.5% with the simulator reward, but the in-domain gain is small and the world model takes about 96 hours to train [arxiv:2512.03556 tab. 1] [arxiv:2512.03556 §4.2] |
| A policy proposes several distinct actions and you can afford many model calls per decision | Search over imagined futures with MCTS | Average success rises from 47.9% to 51.0% on four tasks, a small margin with no variance reported [arxiv:2512.18477 tab. 1] [arxiv:2512.18477 §3.3] |
| The vehicle's mass or steering limit will change after training, and a parameterised vehicle model is available | Decoupled ego-dynamics model with deployment-time adjustment and shifted training | Roundabout success is 0.70 against 0.53 for DreamerV3, but intersections are slightly lower. Deployment-time adjustment cannot help when the new car cannot reach the needed states, such as a smaller steering limit [arxiv:2512.02417 tab. 1] [arxiv:2512.02417 §4.3] |
| You need a driving policy from pure exploration, with no task reward until a short fine-tune, and the target is a new town | Reward-free ensemble-disagreement pretraining | Zero-shot success on Town02 is 73.20%, against 55.83% for RND and 48.61% for ICM [arxiv:2512.18850 tab. 6] |
| You must reach many goals, some never given in training, using only a sparse goal reward and your own practice data | Goal-conditioned model trained on subgoal episodes | 3-Block navigation success on 225 manually chosen start-goal pairs is 95%, against 56% for GC-Dreamer. The method is slower to run, at 36 to 70 GPU hours per run [arxiv:2411.02446 tab. 3] [arxiv:2411.02446 tab. 1] |
| The environment is text, you have logged agent trajectories, and the environment has a fixed state | Fine-tuned language model as simulator | Near-exact next states on ALFWorld and SciWorld, and consistency of 0.91 to 0.96 on three structured environments [arxiv:2512.18832 tab. 1] [arxiv:2512.18832 tab. 2] |
| The environment is open-ended, such as shopping or tool calls | Fine-tuned language model as simulator, validated against the real environment | WebShop consistency falls to 0.67 and StableToolBench exact match is 48.90% [arxiv:2512.18832 tab. 1] [arxiv:2512.18832 tab. 2] |
| Your agent differs from the agents that produced the training logs | Mix trajectories from several agents before training the simulator | Mixing raised one agent's consistency from 0.49 to 0.81 [arxiv:2512.18832 tab. 3] |
| You need common-sense, spatial or social facts and no dynamics training | Language-model world knowledge | Linear space and time representations are reported, but urban knowledge is coarse and Theory of Mind results have gaps [arxiv:2411.14499 §3.2] |
| You must reason about what other people believe, intend or know, and an embodied agent needs that | Mental world model, with prompting or explicit inference | The review gives the trade-off between black-box prompting and interpretable but costly inference. Its numbers are quoted from other papers, and most benchmarks are text-only [arxiv:2601.02378 §4] [arxiv:2601.02378 §5] |
| You need many generated scenarios for training or policy evaluation | Video generator or dynamic embodied environment, checked against real behaviour | The survey lists data engine, RL environment and policy evaluator as uses, but also reports physical-law failures [arxiv:2411.14499 §5.5] [arxiv:2411.14499 §4.1] |
| The model must run on an edge device | Compress states into a latent space for planning | The survey names this edge-side role explicitly [arxiv:2411.14499 §5.5] |
| Physically accurate counterfactual outcomes are required | Treat as open | The survey lists physical rules and counterfactual simulation as unsolved [arxiv:2411.14499 §6] |
| Camera-only driving, and a forecast the planner reads at each step | Drive-OccWorld-style forecast with a cost-based planner | Its encoder is the latency bottleneck at 140 to 406 ms and its decoder takes 188 to 220 ms on an A6000 [arxiv:2408.14197 tab. 9] |
| Camera-only planning, and no future features available at run time | IR-WM fully decoupled planner | It reads only the current BEV state. Average L2 is 0.87 m, against 0.53 m for the default, with 558 ms against 706 ms latency on a device not stated [arxiv:2510.16729 tab. 5] [arxiv:2510.16729 §3.2.4] |
| The forecast should shape the learned features, and the planner needs no forecast output | IR-WM semi-coupled default | Occupancy is only an auxiliary supervision signal and optional output [arxiv:2510.16729 §3.2.4] |
| You want to see or score full surround-camera futures for each planner candidate, and can pay for a video diffusion model per candidate | Multiview video diffusion with image-based rewards | Average L2 is 0.80 m against 1.02 m for VAD with a random command, and collision is 0.26% against 0.93% [arxiv:2311.17918 tab. 3]. The reward weights are not stated, no runtime is reported, and the released repository has no method code [arxiv:2311.17918 §4.2] [arxiv:2311.17918 tab. 3] |
| Several players share one world, and their camera views must agree with each other | Typed shared state with a learned Logic Engine and a separate renderer | Parser state recovery on Snake is 0.764, against 0.128 for the best video baseline, and cross-view disagreement is zero by construction. The invalid rate is higher, at 0.177 against 0.052, and a step takes about 45.6 ms against 3 to 4 ms for the dense baselines [arxiv:2608.06257 tab. 2] [arxiv:2608.06257 tab. B.4] |

The World Models pipeline starts from 10,000 random rollouts for its
world-model training [arxiv:1803.10122 §3.2]. Data cost matters for the text
variant. The training trajectories in the study
come from GPT-4o rollouts, at 40K per environment and 70K for WebShop
[arxiv:2512.18832 §4]. Data scaling saturates near 20K trajectories on the
structured environments, while WebShop keeps improving to about 70K
[arxiv:2512.18832 §6.1] [arxiv:2512.18832 fig. 2]. The study's released
repository has no license file, so check its terms before reusing its code.

## Implementation sketch

The survey gives no algorithm. The loop below combines its model-learning
objective and its planning objective. It is our reading of those two parts,
not code from the paper [arxiv:2411.14499 §3.1].

```pseudocode
# Fit a one-step transition model to real data  [arxiv:2411.14499 eq. 1]
def fit_world_model(transitions, model):
    for state, action, next_state in transitions:
        predicted = model(state, action)
        update(model, squared_error(next_state, predicted))
    return model

# Score sampled action sequences by predicted return  [arxiv:2411.14499 eq. 3]
def plan(model, reward, state, horizon, samples):
    best_sequence, best_return = None, float("-inf")
    for sequence in sample_sequences(samples, horizon):
        predicted_return = rollout_return(model, reward, state, sequence)
        if predicted_return > best_return:
            best_sequence, best_return = sequence, predicted_return
    # Execution and re-planning are not stated in the survey
    return best_sequence
```

For the text-environment variant, `model` is a fine-tuned language model. It
takes the dialogue history and the action, and returns the next observation and
a success flag [arxiv:2512.18832 §3.1] [arxiv:2512.18832 eq. 2]. The paper's
rollout keeps the agent's own history alongside the simulator's history. The
paper states no step limit for these rollouts [arxiv:2512.18832 §3.1].

For the driving variants, the same loop runs over a BEV state rather than a
text or symbolic state. IR-WM adds each predicted change to the previous state
[arxiv:2510.16729 eq. 6]. Drive-OccWorld scores sampled trajectories against
the predicted occupancy at each step, then feeds the chosen trajectory back into
the next forecast [arxiv:2408.14197 §3.4]. Drive-WM runs the same loop over
generated camera images, and scores each candidate with perception outputs
[arxiv:2311.17918 §4.1] [arxiv:2311.17918 §4.2].

The robot variants that search or train inside the model follow the same
shape. STORM expands a tree of policy candidates and gives one unvisited child
a model rollout per simulation, then picks the most-visited root action
[arxiv:2512.18477 alg. 1]. RoboScape-R uses the model as the environment: each
step returns a frame and a done signal, and the reward is computed from them
[arxiv:2512.03556 alg. 1]. MASS runs the same loop over a list of typed
records rather than a flat state. It decodes each record under a schema mask,
assembles the next state, and renders every camera from that one state
[arxiv:2608.06257 alg. 1].

The World Models controller does not plan with the model. It is scored by
rollouts in the real environment or in the dream, and CMA-ES updates its
weights from those scores [arxiv:1803.10122 §3.2] [arxiv:1803.10122 §A.4].
Use the loop above for planners, and this search for a fixed small policy.

## Contradictions

- **Physical realism of video generators.** The introduction says Sora adheres
  to physical laws, and a related claim says it does so "perfectly"
  [arxiv:2411.14499 §1]. Later sections report weak causal reasoning and
  failures on physical laws [arxiv:2411.14499 §4.1], and the open-problems
  section lists the same gap [arxiv:2411.14499 §6]. The intro's statements are
  the authors' descriptions, not measured results.
- **Benchmark size.** A table lists UrbanVideo-Bench as 5.2k urban egocentric
  clips [arxiv:2411.14499 tab. 8]. The open-problems text says 5.2k drone clips
  [arxiv:2411.14499 §6]. Check the original benchmark before citing either.
- **Consistency of the strongest agents.** The text says GPT-4.1, GPT-5 and
  Claude reliably exceed 90% consistency [arxiv:2512.18832 §5.2]. The table
  gives GPT-5 and Claude 0.89 and 0.88 on SciWorld, and 0.58 to 0.82 on WebShop
  across the three [arxiv:2512.18832 tab. 2].
- **Text F1 on StableToolBench.** The fidelity paragraph reports 49% F1
  [arxiv:2512.18832 §5.1]. The table's exact-match column holds 48.90 and 49.25,
  and its F1 column holds 79.15 and 78.97 [arxiv:2512.18832 tab. 1]. Use the
  table.
- **Planning protocols are not comparable.** Drive-OccWorld reports an average L2
  of 0.85 m under NoAvg and 0.47 m under TemAvg [arxiv:2408.14197 tab. 5]. IR-WM
  reports 0.53 m next to the NoAvg baseline rows, and does not name its own
  protocol [arxiv:2510.16729 tab. 4]. Drive-WM averages L2 over 1, 2 and 3 s
  against a VAD baseline on nuScenes [arxiv:2311.17918 tab. 3]. Do not set
  these figures side by side.
- **Baseline provenance.** IR-WM lists Drive-OccWorld numbers in its tables. The
  paper does not say whether they were re-run or copied
  [arxiv:2510.16729 tab. 1]. Treat the baseline gaps as the authors' claims until
  they are checked.
- **Driving action signal.** Drive-WM's main text defines the action as the ego
  displacement per step [arxiv:2311.17918 §3.3] [arxiv:2311.17918 §4.1]. Its
  appendix says the action-based data uses speed and steering angle
  [arxiv:2311.17918 app. C.1]. The paper does not say which one its model used.
- **Dreamer and TD-MPC2 baselines across papers.** The hierarchical humanoid
  paper, the driving papers and the reward-free driving paper each report
  Dreamer-style or TD-MPC2-style baselines, on different tasks and simulators
  [arxiv:2405.18418 §4.1] [arxiv:2512.02417 tab. 1] [arxiv:2512.18850 tab. 6].
  The numbers do not conflict, but they are not comparable.

## Reproduction caveats

The new cards record the same kind of gap between paper and code as the
occupancy cards. The humanoid paper's text suggests the high-level command is
an absolute end-effector position, while its code adds a scaled offset to the
current positions [arxiv:2405.18418 §1] [arxiv:2405.18418 §3]. The reward-free
driving paper's equation takes the variance of the ensemble, while its code
takes the standard deviation before averaging [arxiv:2512.18850 §3.2.1]
[arxiv:2512.18850 eq. 4]. RoboScape-R adds a distance term with no stated sign
convention [arxiv:2512.03556 eq. 5] [arxiv:2512.03556 alg. 1], so a literal
implementation can reward moving away from the goal. Check each card's Paper
vs code section before you reproduce a number from these papers. The Drive-WM
card finds no model, planner or reward code in the linked repository, so its
planning numbers cannot be checked against code
([Drive-WM card](/papers/2311.17918)).

The MUN card records the same kind of gap. The released configs for three of
the six tasks sample subgoals at random, not with the farthest-point sampling
the paper describes [arxiv:2411.02446 §3.2]. The code also decides when a
subgoal is reached from a learned distance, so the success signal is an
estimate [arxiv:2411.02446 alg. 1]. The paper's training steps and Walker
subgoal count differ from its own tables [arxiv:2411.02446 tab. 1] [arxiv:2411.02446 tab. 2].
The repository has no LICENSE file, so check its terms before you reuse its code.

The MASS card finds no official code. Its project page has no code link, and
the lab repository found there is a static site. The renderer loss weights and
the checkpoint-selection weights are not stated, so the typed results cannot be
checked against source [arxiv:2608.06257 app. A.2] [arxiv:2608.06257 app. B.1].
([MASS card](/papers/2608.06257)).

## Papers

- [World Models](/papers/1803.10122): a VAE, a recurrent mixture-density model and a linear controller evolved with CMA-ES. It reports Car Racing scores and a Take Cover policy trained inside the learned model.
- [Driving into the Future: Multiview Visual Forecasting and Planning with World Model for Autonomous Driving](/papers/2311.17918): a latent video diffusion model that generates six nuScenes camera views from layout, text and ego-motion conditions, and ranks VAD-planned candidates with image-based map and object rewards.
- [Hierarchical World Models as Visual Whole-Body Humanoid Controllers](/papers/2405.18418): two TD-MPC2 world models, a frozen tracker for 836 motion clips and a visual planner that sets its commands, evaluated on eight simulated humanoid tasks.
- [Learning World Models for Unconstrained Goal Navigation](/papers/2411.02446): a goal-conditioned Dreamer-style model trained on episodes between subgoals chosen by farthest-point sampling over actions, evaluated on six simulated goal-conditioned tasks.
- [Understanding World or Predicting Future? A Comprehensive Survey of World Models](/papers/2411.14499): a survey that frames world models as implicit representation and future prediction, covering model-based RL, language-model knowledge, video generators, embodied environments and applications. It reports no experiments of its own.
- [From Word to World: Can Large Language Models be Implicit Text-based World Models?](/papers/2512.18832): fine-tunes an LLM to predict next observations in text environments, and tests its fidelity, consistency and use for agent training and checks.
- [Driving in the Occupancy World: Vision-Centric 4D Occupancy Forecasting and Planning via World Models for Autonomous Driving](/papers/2408.14197): a camera-only driving world model that forecasts 4D occupancy and flow, with action conditioning and a cost-based planner.
- [Vision-Centric 4D Occupancy Forecasting and Planning via Implicit Residual World Models](/papers/2510.16729): predicts the residual change of a BEV state instead of the full future scene, adds feature alignment, and plans from the forecast.
- [Vehicle Dynamics Embedded World Models for Autonomous Driving](/papers/2512.02417): splits a Dreamer-style model into ego-vehicle and environment parts, with a learned goal policy, a PID controller, and adjustment and augmentation for changed vehicle parameters.
- [RoboScape-R: Unified Reward-Observation World Models for Generalizable Robotics Training via RL](/papers/2512.03556): trains robot policies with PPO inside an action-conditioned video model, using a reward from a goal frame and a done signal.
- [STORM: Search-Guided Generative World Models for Robotic Manipulation](/papers/2512.18477): wraps a diffusion VLA policy in Monte Carlo tree search, with a fine-tuned video model scoring imagined outcomes.
- [InDRiVE](/papers/2512.18850): reward-free driving pretraining with ensemble disagreement as the intrinsic reward, tested zero-shot and few-shot in CARLA towns.
- [Modeling the Mental World for Embodied AI: A Comprehensive Review](/papers/2601.02378): a survey separating physical from mental world models, comparing Theory of Mind reasoning paradigms and cataloguing benchmarks. It runs no experiments.
- [MASS: Multiplayer World Models with Authoritative Shared State](/papers/2608.06257): a learned Logic Engine advances one typed world state for all players, and a separate Rendering Engine draws each camera from it, evaluated on a two-player Snake benchmark and on eight games for rendering.