Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Artificial Representation and World Models

Abstract. Artificial representation concerns the states and structures through which an artificial system retains, transforms, and uses information about environments, tasks, and possible futures. World models are learned or specified mechanisms that estimate environmental state, transitions, and consequences of actions, often under incomplete observation. This fundamental research monograph critically integrates cognitive psychology, computational theories of representation, model-based reinforcement learning, structured and causal representation learning, video prediction, language models, and embodied planning. A narrative scholarly synthesis compares symbolic, probabilistic, latent, relational, and retrieval-supported architectures; distinguishes predictive accuracy, control utility, transfer, and mechanistic explanation; and evaluates historical experiments alongside PlaNet, Dreamer, MuZero, and V-JEPA 2. It identifies recurrent confounds arising from shortcut features, representation nonidentifiability, simulation bias, limited interventions, and benchmark leakage. The proposed research agenda specifies behavioral and mechanistic tests for stable state tracking, prediction under distribution shift, intervention-sensitive transfer, and externally verifiable planning. Artificial Era implications are treated as research questions grounded in operational evidence.

Keywords: artificial representation; world models; cognitive psychology; latent state; state estimation; predictive learning; model-based reinforcement learning; symbolic representation; structured representation; object-centric learning; causal representation; partial observability; planning; counterfactual generalization; retrieval-augmented models; Artificial Era

The Scientific Object of Artificial Representation

Artificial representation is an organized internal or externally accessible state that an artificial system can use to discriminate conditions, preserve task-relevant information, and guide subsequent computation. A representation may take the form of symbolic propositions, a vector of learned features, a probability distribution, an object graph, a retrieved document, or a recurrent memory state. Its scientific identity depends on measurable relationships among observations, internal variables, and subsequent behavior. This dossier studies how such states support tracking, prediction, action selection, and transfer across circumstances. Research on learned features established the importance of extracting factors useful for multiple tasks, while cognitive science asks which structural relationships a system can preserve and manipulate. The operational question is what changes when a proposed representation is perturbed, withheld, or tested beyond familiar inputs. Bengio et al. (2013) Marr (1982/2010)

A world model is an explicit or learned mapping between possible environmental states, observations, and transitions, potentially conditional on actions. In a partially observable setting, its state may summarize a history of observations and executed actions rather than correspond to a directly visible scene. Its outputs can forecast sensory trajectories, rewards, abstract events, or decision-relevant values. The breadth of these outputs matters: predicting rewards for action selection is a legitimate learned model even if it lacks a photorealistic image generator; forecasting plausible video is a prediction system even before planning utility is established. The distinct computational objectives can converge in particular architectures, but should be separately tested. Studies of partially observable planning and learned dynamics make these distinctions explicit. Kaelbling et al. (1998) Schrittwieser et al. (2020)

The scientific scope includes symbolic and structured accounts, probabilistic belief states, latent transition dynamics, object-centered learning, predictive video systems, and retrieval-supported representations. Comparisons across these approaches address whether systems generalize correctly to changed object relations, interrupted trajectories, hidden causes, and interventions. A method can achieve high average benchmark scores by exploiting a stable cue; success then supports a limited functional claim rather than a general account of the modeled environment. Evidence must integrate predictive calibration, accuracy of downstream actions, representational probes, and controlled perturbations. Research on shortcut learning and the impossibility of uniquely identifying fully disentangled factors from unrestricted observation cautions against equating convenient feature geometry with discovery of the environment's unique true structure. Geirhos et al. (2020) Locatello et al. (2019)

The Historical Origins of Internal Models

Psychological models of representation developed before neural machine learning. Edward C. Tolman argued that navigation and learning may require cognitive maps which relate places beyond a sequence of reinforced motor responses. In his 1948 synthesis, experimental observations of maze behavior were used to motivate internal organization of spatial knowledge; the exact representational mechanisms remained open. His work remains historically important because it treated flexible behavior as a constraint on explanatory accounts. A map is scientifically useful when it predicts novel paths, detours, or changes in navigational strategy after rearrangement of the environment. The general insight extends to artificial systems: evidence about internal spatial structure is strongest when an agent must operate under novel layouts, rather than only repeat a previously reinforced trajectory. Tolman (1948)

Computational cognitive science later specified what a model would have to compute. Marr separated the computational problem, the representation and algorithm that solve it, and the physical implementation. This distinction prevents the observation that two systems reach similar decisions from establishing that they use identical representational procedures. A symbolic planner and a neural dynamics model can share an objective while differing in the form and generalization properties of intermediate states. The reissue of Marr's 1982 book documents his influential program for constructing visual descriptions from input images to representations of three-dimensional objects. Modern artificial world-model research inherits the demand for explicit tasks and transformations, while adding data-driven learning of internal state and transition functions. Marr (1982/2010)

A parallel engineering tradition connected models to feedback and action. Sutton's Dyna integrated direct reinforcement learning, model updating, and simulated planning in a single architecture. The learned transition model was useful for generating additional updates that could improve behavior before more interaction with the environment. Partially observable decision theory then provided a rigorous account of agents whose observations do not reveal their true states: a belief over possible states can be updated as evidence arrives and used to choose actions. These traditions explain why simulation, inference, and policy optimization should be connected in a scientific history of world models. They also establish a recurring limitation: simulated experience inherits the inaccuracies of the learned dynamics. Sutton (1991) Kaelbling et al. (1998)

Representation in Cognitive Psychology and Artificial Systems

Representational claims in psychology concern which internal distinctions help explain observable performance. A person or machine may categorize objects from their surface appearance, infer hidden physical properties, or select actions through a learned stimulus–response mapping. These behaviors place different demands on representational structure. Accounts of human concept acquisition emphasize abstract structure, Bayesian inference, and the use of causal knowledge to support generalization beyond direct training examples. Their relevance to artificial research is methodological: they supply tasks involving sparse learning, systematic composition, and inference about unobserved causes. Treating human performance as a comparison requires matching task instructions, exposure, prior knowledge, and response constraints rather than assuming equivalent developmental histories. Tenenbaum et al. (2011) Lake et al. (2017)

An artificial representation can be evaluated at the behavioral level by comparing decisions under changes to observable information, at the algorithmic level by measuring transformations of hidden states, and at the implementation level by investigating activations and circuits. Behavioral competence alone may leave multiple models observationally equivalent. A recurrent vector carrying information about a moving object might be decoded by a probe without being used to control the next action; the decode result establishes recoverability but not causal function. A stronger study intervenes on that vector, measures downstream prediction and control, and verifies that changes remain selective under matched alternative features. This general requirement follows Marr's analysis of levels and current debates about the explanatory status of internal variables. Marr (1982/2010) Schölkopf et al. (2021)

Artificial systems further complicate the notion of representation by shifting storage between internal weights, temporary contexts, external indexes, tools, and environmental inscriptions. Distributed cognitive work shows how tasks can be accomplished through combinations of people, artifacts, and material layouts, making the location of computation an empirical issue. In retrieval-augmented generation, factual information can be made available through a document index alongside parameterized language modeling; performance therefore depends on retrieval quality, document freshness, and how retrieved evidence is integrated. These systems offer a useful contrast to purely weight-based world models. A model that produces a good answer from external documentation should be credited for the task it performs, with its source-dependent knowledge distinguished from a continuously updated internal model of state dynamics. Hutchins (1995) Lewis et al. (2020)

Levels of Explanation and Construct Validity

Claims about an artificial system understanding the world encompass several possible scientific constructs. One concerns encoding: can a variable distinguish objects, locations, relations, or time? Another concerns prediction: can its states forecast observations or rewards? A third concerns control: do simulated futures improve action selection? A fourth concerns transfer: can the same structure support unfamiliar combinations or interventions? Each requires different tasks and measurements. A benchmark reporting one average accuracy score can conceal a mixture of these abilities. The scientific object is defined by specifying the information available to the system, the variables altered by the experimenter, the expected outcome, and the comparison model. Cognitive evaluations of language models show that impressive aggregate results can coexist with fragile responses to minor task changes. Binz & Schulz (2023) Lake et al. (2017)

At the algorithmic level, representational theories differ in their assumed structure and constraints. Some embed whole images into continuous spaces, while others encode objects, relations, or belief distributions. The fact that a dimension correlates with a human-defined factor does not determine whether that factor is independently controllable or causally privileged. Locatello and colleagues demonstrated the difficulty of identifying unsupervised disentangled structures without appropriate inductive biases; their large comparative study also cautions that preferred disentanglement scores do not automatically yield downstream sample-efficiency gains. These results concern formal assumptions and particular empirical benchmarks, not a theorem excluding all useful structured representations. A well-defined experimental program should state its inductive biases and test the utility of the proposed factors beyond the score used to train them. Locatello et al. (2019)

World-model explanation also requires separating the prediction horizon from the outcome horizon. A system may forecast the next frame with high visual fidelity but accumulate error after repeated imagined transitions. Another may generate low-resolution abstractions that are sufficient for navigation or control. Scientific evaluation therefore has to report which targets were predicted, whether rollouts were open-loop or corrected by new observations, and whether uncertainty and decision consequences were measured. MuZero illustrates the importance of task-specific predictive targets by learning the policy, value, and reward quantities used by search. PlaNet instead studies latent dynamics sufficient for choosing continuous actions from pixels. Their comparative interpretation depends on differences in environments, objectives, and access to interaction data. Schrittwieser et al. (2020) Hafner et al. (2019)

Research Questions and Evidentiary Strategy

This monograph is a critical narrative synthesis organized around five empirical questions. What constitutes a task-relevant representation, and by what evidence can a hidden variable be identified? How do artificial systems acquire predictive models from sequences of observations? When do learned models improve planning and exploration beyond reactive alternatives? Which experimental controls establish compositional or causal transfer? Finally, how can representations be compared across language, vision, embodied agents, and human cognition without reducing all these domains to a single performance index? The review combines historically influential conceptual works with original experiments and recent peer-reviewed machine-learning research. No new human or machine experiment is reported; the conclusions are constrained by the available designs, public reports, and methodological assumptions. Bengio et al. (2013) Lake et al. (2017)

Primary studies are read in relation to their stated training and evaluation conditions. Controlled simulations supply access to ground truth and interventions but often simplify environmental diversity. Video-based benchmarks offer rich observation streams yet frequently provide weak control over the generating processes. Cognitive experiments with language models probe task behavior but must distinguish genuine generalization from linguistic cues and exposure to familiar paradigms. Robot experiments add physical interaction and closed-loop measurement, while limited robot populations and laboratory settings constrain broad claims. Evidence quality is therefore assessed through evaluation design, intervention strength, distribution shift, independent replication, baseline relevance, and openness of code or data where supplied. Reported benchmark superiority is bounded to the tested tasks. Binz & Schulz (2023) Assran et al. (2025)

The synthesis distinguishes established engineering results from propositions about broader psychological organization in the Artificial Era. The former include reproducible algorithms for state inference, planning through learned dynamics, and measured improvements under defined benchmarks. The latter concern how stable representations might support ongoing artificial interaction, continuity across sessions, and coordination with human cognitive practices. These extensions require explicit variables and prospective tests. The investigation is historically anchored in psychological studies of internal maps, computational levels of analysis, and modern model-based control; its conclusions describe functional capacities, causal limitations, and open research problems without inferring subjective experience from the ability to predict or plan. Tolman (1948) Friston (2010)

Symbolic Models, Rules, and Explicit State Spaces

Symbolic world models express entities, properties, actions, and relations through discrete variables and transition rules. Their strengths are explicit state descriptions, inspectable preconditions, and the ability to specify consequences of hypothetical actions. A route planner that represents roads and intersections as a graph can answer questions about detours through graph operations; a symbolic household agent may encode that an object must be grasped before it can be moved. These advantages depend on the correctness and granularity of the supplied vocabulary. A missing object property or unmodeled environmental event can invalidate a chain of inferences. Modern representation research therefore asks whether the system can acquire and update the necessary symbols from observations, rather than merely execute a hand-built rule set. Marr (1982/2010) Battaglia et al. (2018)

Probabilistic decision processes supply a rigorous language for combining explicit state spaces with uncertainty. In a Markov decision process, transition and reward functions support decisions conditioned on a known state; under partial observability, the agent instead maintains a belief distribution and revises it as observations arrive. The belief state has predictive and control functions: it organizes uncertainty about hidden locations, identities, or environmental conditions. Kaelbling and colleagues formalized the implications for planning and analyzed the complexity of partially observable problems. Their results matter because a representation can be effective without encoding a single certain picture of the world. Researchers should ask whether uncertainty estimates are calibrated and whether the agent acts to resolve ambiguities when information is valuable. Kaelbling et al. (1998)

Symbolic architectures facilitate counterfactual checks when their rules are accurate and interventions are well specified. Changing a road closure in a stored graph can alter the optimal route without relearning every edge. In learned systems, comparable flexibility must be demonstrated rather than assumed from the presence of named variables. Structured knowledge can also be externally stored, and computational tasks may be distributed among tools, memory, people, and local reasoning processes. This raises an architectural measurement issue: the same behavior may result from general internal transition rules, fresh retrieval of an answer, or repeated prompt-guided reconstruction. An adequate experiment manipulates access to the rule store, tests genuinely novel situations, and evaluates the resulting changes in predictions. Hutchins (1995) Lewis et al. (2020)

Latent State Learning and Temporal Compression

Latent world models seek a compact representation of the information needed to predict future observations or task-relevant outcomes. An encoder maps sensory inputs to a lower-dimensional state; a transition function predicts how that state changes as actions and time unfold; a decoder or prediction head connects latent states to observations, rewards, or other targets. In recurrent systems, the state depends on prior history, making it capable of representing hidden variables that are only indirectly observed. Ha and Schmidhuber showed how a learned generative environment model can support compact controllers and policies trained within simulated trajectories. Their demonstration established a practical research program, although artificial rollouts necessarily reflect the model's own learned distribution. Ha & Schmidhuber (2018)

Latent learning is shaped by the predictive objective. Reconstruction losses can encourage retention of visual details irrelevant to decisions, while reward prediction can discard background information whose physical dynamics might matter after a task changes. Variational approaches regularize latent states and often incorporate stochasticity to accommodate unpredictable future outcomes. PlaNet combined deterministic and stochastic transition components with a multi-step training objective and used online planning from image observations. Its results are evidence that latent models can support control from pixels on specified continuous-control benchmarks. They do not determine whether latent coordinates correspond to human-interpretable objects, causes, or experiences; those properties require supplementary tests with controlled variation of latent factors. Hafner et al. (2019) Bengio et al. (2013)

The compression question is therefore functional as well as statistical. A model should preserve enough information to distinguish environmental situations that demand different future actions, while allowing irrelevant observations to vary without destabilizing its decisions. Predictive sufficiency can be tested by comparing future outcome distributions given the learned state to those conditional on the full available history. Empirical tests can also remove information from the history and determine whether the current representation preserves a hidden event's consequences. Whenever the model is trained jointly with a policy, a representation optimized for that policy may fail under a new objective. The interpretation of latent state as a broadly transferable world model depends on experiments that vary goals, observation channels, and environmental dynamics. Hafner et al. (2020) Janner et al. (2019)

Object-Centered and Relational Representations

Object-centered models introduce structure by decomposing an environment into candidate entities and their relations. A state might contain slots representing independently moving objects and edges describing interaction, contact, or relative position. The promise is compositional transfer: if a model has learned reusable object dynamics, it may cope with an unfamiliar number or arrangement of objects. Graph-network approaches describe relational computation as an inductive bias, permitting the same transition machinery to operate over different relational configurations. Battaglia and colleagues articulated this program through graph-based formulations that integrate earlier relational learning traditions. The architecture supplies a hypothesis about environmental structure; its relevance must be verified against interventions that change entities and relationships in ways unseen during training. Battaglia et al. (2018)

Kipf and colleagues proposed contrastively trained structured world models in which object embeddings and their interactions are learned from images. The work is methodologically important because it investigates object-centric learning without direct labels for every latent entity, using compositional environments where controlled changes can be studied. A contrastive objective can learn to discriminate real transitions from negatives; that achievement supports a specified representation-learning result. It does not by itself show that every learned slot has a unique semantic identity across all scenes, or that the model has recovered every physical cause. A proper replication would test slot consistency, intervention response, relation generalization, and whether the same representation can support planning on held-out object configurations. Kipf et al. (2020)

Structured representation is also vulnerable to the assumptions made by the experimenter. Benchmark scenes with well-separated objects and stable camera motion can favor object-slot methods, while real-world occlusion, deformable objects, changing identity, and ambiguous boundaries present harder tests. Discrete object structure may improve one form of generalization while introducing failures when the environment does not naturally partition into the expected units. Scientific comparison therefore requires matched training data and compute, matched downstream tasks, and explicit tests of how representations respond to object removal, identity substitution, and changes in relational topology. The goal is to establish which inductive bias helps under which ecological and computational conditions, not to elevate one representation family into a universal ontology of physical reality. Kipf et al. (2020) Locatello et al. (2019)

Causal Variables and Intervention-Sensitive Generalization

A representation can predict regularities while remaining vulnerable to interventions that break them. If an agent has seen a red object only when it moves left, it may incorrectly treat color as the cause of motion. Causal representation learning examines whether models can identify variables and mechanisms whose relationships remain useful across changed conditions. Schölkopf and colleagues argued that causal structure and independent mechanisms could improve robustness under distribution shifts. Their proposal provides explanatory criteria and research directions; a learned causal model still requires intervention-sensitive validation. A correlation between a hidden dimension and an observed property does not demonstrate that changing this dimension will reproduce the causal effects of manipulating that property in the environment. Schölkopf et al. (2021) Geirhos et al. (2020)

Experimental identification becomes clearer when actions or exogenous changes alter putative causes separately from competing cues. Researchers may randomize object color while holding dynamics constant, intervene on friction while matching object appearance, or alter an agent's goals without changing its visual inputs. A genuine model of relevant mechanisms should make appropriately updated predictions, with uncertainty reported where distinct causal explanations remain consistent with the data. Controlled synthetic environments permit exact counterfactual checks, but causal claims transferred to natural settings require additional validation. In cognitive psychology, the ability to infer hidden causal structure from sparse observations is a major explanatory target; comparisons with artificial systems become useful when tasks are matched to the types of intervention and prior knowledge available. Tenenbaum et al. (2011) Lake et al. (2017)

Identifiability is central because multiple latent parameterizations can reproduce a common observational distribution. A model may encode a rotated or distributed representation from which factors can be decoded, but this does not establish independent manipulability or invariant causal mechanisms. The large-scale analysis by Locatello and colleagues showed why inductive biases and supervisory information cannot be ignored when interpreting unsupervised disentanglement. Causal world-model evaluation should consequently document the permissible equivalence classes, the variables experimentally manipulated, the tests of invariant mechanisms, and the role of prior assumptions. Success is strongest when the learned system predicts outcomes following interventions that were absent from training and performs reliably across multiple, independently generated contexts. Locatello et al. (2019) Schölkopf et al. (2021)

External Memory, Retrieval, and Hybrid World Models

Artificial representation is sometimes distributed across model parameters, temporary memory, and external information systems. Retrieval-augmented generation combines a learned language generator with an indexed document store. It can improve access to specific facts and facilitate updates without retraining the entire model, but document retrieval is not equivalent to predicting environmental dynamics. A retrieved passage may describe how a machine works; using that passage to choose and execute novel actions requires further inference, state tracking, and appropriate uncertainty handling. Lewis and colleagues formalized retrieval-augmented generation and compared it on knowledge-intensive tasks. The research helps distinguish a system's stored propositional knowledge from the predictive transition machinery studied in model-based control. Lewis et al. (2020)

Hybrid architectures can connect retrieved knowledge to explicit or latent world models. For example, a robot might retrieve a manual, identify a component, update a belief state from camera observations, and use a transition model to simulate a safe movement. Scientific attribution should separate success at retrieval, grounding, temporal inference, and control; failures at any link can produce an apparently coherent but erroneous plan. Comparisons should include conditions with removed or corrupted documents, delayed sensor observations, and conflicting sources. Distributed-cognition research provides an ecological perspective: the operational system includes agents and artifacts that transform information together, so the experimental unit and provenance of each assertion need careful specification. Hutchins (1995) Kaelbling et al. (1998)

World-model claims about language systems require similar discrimination. Next-token prediction may encode statistical information about physical, social, or procedural regularities, and appropriately designed tasks can reveal useful predictive behavior. Yet a correct verbal answer can follow from memorized descriptions or linguistic patterns, especially when test examples resemble training material. Binz and Schulz applied cognitive-psychological experiments to GPT-3 and found a mixture of competent task performance and marked fragility under changes in wording or causal demands. An interpretable hybrid study would compare parametric-only responses, retrieval-supported responses, and interactive observations under held-out interventions. Claims of persistent internal environment modeling would then rest on direct state-tracking evidence rather than the fluency of a generated explanation. Binz & Schulz (2023) Lewis et al. (2020)

Generative World Models and Simulated Experience

Ha and Schmidhuber's World Models demonstrated a neural architecture combining compressed visual representations, recurrent temporal dynamics, and a compact controller. The work was influential because the agent could be trained within trajectories generated by its learned model and then evaluated in the actual environment. The distinction between simulated and actual experience provides a direct criterion of model adequacy: control strategies that work only inside the model reveal a mismatch between learned dynamics and environmental dynamics. This approach therefore supplied both an engineering method and an empirical test of world-model limitations. It also highlighted why successful compression alone is insufficient; the compressed state must retain those aspects of the world that affect rewards and actions under the controller's task. Ha & Schmidhuber (2018)

Simulated experience has several scientifically distinct uses. Imagined trajectories can supply extra training samples for a policy, allow an agent to evaluate alternative actions before acting, or help it estimate future information gains. All three uses depend on the model's state-transition function, but they differ in what constitutes success. A better policy after synthetic training supports a practical transfer result; a simulated trajectory that matches actual outcomes supports predictive accuracy; and a model's ability to select informative actions supports exploration. Dyna anticipated these relationships by combining model learning and planning with direct learning. New neural models scale the idea to high-dimensional observations while retaining its susceptibility to model error and inaccurate simulated futures. Sutton (1991) Ha & Schmidhuber (2018)

Evaluation must therefore compare more than task return. Researchers should report prediction accuracy over multiple horizons, uncertainty under unfamiliar transitions, differences between simulated and observed trajectories, and performance when the policy is deployed beyond the model. A controller can adapt to regular errors in its simulator, creating a misleading impression of competence if only imagined rewards are measured. The amount of real-world interaction used for learning and evaluation also matters. Comparing a model-based controller with a model-free system requires attention to training data, compute, tuning budgets, observation frequency, and environmental reset rules. The relevant conclusion is that synthetic rollouts can be productive when their errors are monitored and bounded, rather than that simulated experience universally substitutes for interaction. Janner et al. (2019) Kaiser et al. (2019)

PlaNet and Learning Dynamics from Pixels

PlaNet, introduced by Hafner and colleagues in 2019, investigated planning directly from visual observations in unfamiliar environments. The system learned a recurrent latent-state model with deterministic and stochastic components and performed online action selection by evaluating candidate future sequences. This approach avoided requiring an externally supplied state representation of physical dynamics. Its experiments included challenging continuous-control scenarios with contact dynamics, partial observability, and sparse reward conditions. Reported gains in sample efficiency relative to selected model-free baselines support a concrete result: a sufficiently trained latent model can make image-based planning practical on the tested tasks. The result does not imply that the internal variables uniquely match human physical concepts or that the algorithm will generalize to unrestricted natural environments. Hafner et al. (2019)

The central methodological issue is predictive sufficiency for decision-making. The latent state in PlaNet does not need to reproduce every pixel at all future times; it must preserve information that helps forecast outcomes relevant to the planner's objective. Its multi-step training procedure, called latent overshooting, was designed to improve long-range state predictions. Evaluation of similar models should distinguish reconstruction likelihood, latent transition consistency, reward prediction, and achieved control performance. A model may improve on one target while degrading another because these targets emphasize different portions of the observational history. The experimental design should also test adaptation when observation noise, episode duration, and action frequencies change, since these conditions affect the effective uncertainty of estimated state. Hafner et al. (2019) Kaelbling et al. (1998)

PlaNet clarifies how representation and planning can be investigated together without confusing the two. One can hold the learned state estimator fixed while changing the optimization procedure, or compare different representations under a matched planner. Such ablations help determine whether an apparent advantage comes from better world modeling, more effective trajectory selection, or favorable optimization settings. Because the original benchmark domains are controlled and often simulated, replication on new environment families remains important. Researchers should disclose the number of interactions, action horizon, training seeds, and whether benchmark parameters were tuned with knowledge of evaluation tasks. These are requirements for interpreting the magnitude and transportability of an engineering result, not a claim that the published experiments lacked merit. Hafner et al. (2019) Janner et al. (2019)

Dreamer, Latent Imagination, and Policy Improvement

The Dreamer line developed the strategy of using imagined latent trajectories to train a policy and a value estimator. Rather than repeatedly solve a new optimization problem through exhaustive online action sampling, Dreamer learned behaviors from predictions generated by a recurrent world model. Hafner and colleagues reported performance gains on visual continuous-control tasks in their 2020 study, illustrating how model-based training can support policies acting from high-dimensional observations. The scientific finding is conditional on the environments, comparison methods, and amount of interaction described in the publication. The imagined trajectories are computational objects learned from experience and do not entail that the system has a subjective capacity to imagine in the human phenomenological sense. Hafner et al. (2020)

World-model optimization introduces a coupled estimation problem. Errors in the dynamics model distort generated states; errors in the value function distort assessment of those states; and actor optimization can exploit both mistakes. A productive empirical design estimates the effect of each component separately by changing rollout lengths, holding training data constant, and measuring real-environment returns. Janner and colleagues analyzed model exploitation and argued for short model-generated rollouts branched from real states as one strategy for limiting bias. Although the mathematical settings and algorithmic details differ from Dreamer, this work identifies a common tradeoff: longer simulated futures offer more opportunities to learn, while also exposing decisions to compounding prediction error. Janner et al. (2019) Hafner et al. (2020)

DreamerV3 expanded the empirical range of learned world-model control. Hafner, Pasukonis, Ba, and Lillicrap reported performance across more than 150 tasks from diverse benchmark domains in a 2025 Nature article, emphasizing a fixed algorithmic configuration and robustness techniques. The evidence supports the practical breadth of that algorithm across its evaluated task suite, with variations in domains and metrics documented by the authors. It is scientifically important to examine which tasks share simulator assumptions, sensory formats, and evaluation conventions; a large number of benchmark tasks is not automatically an estimate of performance under arbitrary new real-world conditions. The broader result strengthens the case that predictive latent control is an important engineering family while leaving mechanistic interpretation and causal transfer as separate questions. Hafner et al. (2025)

MuZero and Decision-Relevant Predictive Models

MuZero provides a distinctive answer to what a world model must represent. Rather than learning to regenerate all future environmental observations, it learns quantities needed for decision search: the policy prior, value, and reward associated with latent transitions. Its search operates over an internal dynamics system trained from experience. Schrittwieser and colleagues evaluated the method on Go, chess, shogi, and a suite of 57 Atari games, reporting high performance without giving the learned model the true rules of those environments. The result demonstrates that task-relevant prediction can support effective lookahead even when the model is not optimized to reconstruct a visually complete future. It also shows why the phrase world model does not uniquely designate an image generator. Schrittwieser et al. (2020)

The interpretive limit follows from the quantities optimized. A value-predictive model can omit aspects of the environment that do not alter reward, action evaluation, or policy quality on the training tasks. Such omission may be appropriate for efficient decision making and problematic when the task changes. A model trained for winning board games has not thereby demonstrated the ability to explain every rule verbally, anticipate unrelated social actions, or transfer to another game without additional learning. The scientific comparison with PlaNet and Dreamer should therefore match the representational targets and the role of planning. MuZero couples latent model learning with tree search; PlaNet plans over image-conditioned latent trajectories; Dreamer trains an actor through imagined transitions. Schrittwieser et al. (2020) Hafner et al. (2019)

Mechanistic interpretation requires dedicated experiments. Researchers can change latent state features, examine the sensitivity of search outcomes, compare model forecasts with independently calculated environment consequences, and identify cases where a successful action is produced despite inaccurate intermediate transitions. When reward prediction is the primary supervision, a model might learn useful compressed strategies while remaining unfaithful to physical details. An adequate experimental taxonomy should report whether it evaluates a task model, an observation model, a belief model, or a general explanatory model. MuZero's published success establishes the practical adequacy of particular decision-relevant predictive representations on its tested tasks; more extensive environmental and causal claims require additional evidence. Schrittwieser et al. (2020) Marr (1982/2010)

Predictive Video Models and Embodied Planning

Visual world-model research increasingly uses video sequences to learn representations of changing scenes without requiring dense action labels. A joint-embedding predictive architecture attempts to forecast informative representations of future observations, prioritizing predictable structure rather than pixel-perfect reconstruction. LeCun's 2022 position paper articulated a hierarchical predictive world-model program built around self-supervised learning and planning; its scientific status is a proposed architecture with testable implications. Video models provide a different observation regime from board-game or low-dimensional control systems. Real-world footage includes camera movement, changing illumination, object occlusion, and independent actors, and the model must decide which regularities are important for predicting what happens next. LeCun (2022) Assran et al. (2025)

V-JEPA 2 extends this program with extensive self-supervised pretraining on video and images followed by action-conditioned modeling for robotic planning. Assran and colleagues reported results on motion understanding, action anticipation, video-question answering, and pick-and-place planning with robot arms. The research demonstrates a pathway from largely passive observation to a model that can guide selected physical interactions after targeted additional training. Its reported robot experiments involve particular equipment and laboratory settings, so claims about general physical competence must retain those boundaries. A pretrained representation can encode useful motion information while still depending on carefully chosen action representations, goal images, and feedback procedures for robotic control. Assran et al. (2025)

Passive video learning poses an identifiability problem that interactive experiments can help resolve. A camera recording may not distinguish whether an object moved because it was pushed, because a motor activated, or because the camera changed position. Action-conditioned robot trajectories can add causal constraints, although their range remains limited by the explored actions and environments. Researchers should therefore compare predictions after controlled changes in viewpoint, manipulated object properties, robot morphology, and contact dynamics. Strong evidence would combine held-out intervention forecasting with closed-loop task performance, measured failure recovery, and independent laboratory replication. The link from video representation to physically grounded world modeling becomes compelling when predictive abstractions continue to guide action as environmental contingencies change. Assran et al. (2025) Schölkopf et al. (2021)

Measuring State Tracking and Temporal Prediction

State tracking is the ability to maintain information about an environment as observations change or become incomplete. A meaningful experiment provides a sequence in which an object becomes temporarily hidden, an event alters a future constraint, or a previous action changes what should be expected later. The model should predict the consequences from its recorded history, and its errors should be compared with simpler baselines that ignore time or rely only on the most recent frame. Appropriate metrics include held-out state accuracy, predictive likelihood or calibrated error, and performance on downstream choices requiring the missing information. A representation that accurately describes visible inputs while forgetting an occluded object's position fails a state-tracking criterion, even if its frame-level classification score is high. Kaelbling et al. (1998) Marr (1982/2010)

Temporal evaluation depends on how the future is generated. One-step predictions conditioned on observed intermediate frames differ from open-loop rollouts in which each predicted state becomes input to the next forecast. Errors that are small after one step may accumulate over long horizons and alter a policy's choices. Studies using recurrent latent dynamics make the distinction important because model training and action selection repeatedly involve predicted future states. Evaluators should report the horizon, conditioning regime, noise level, and whether metrics are computed in pixel space, learned feature space, abstract state space, or reward space. Each choice values different explanatory properties. Model robustness under delayed observations and unexpected events is especially relevant when the system must maintain continuity through incomplete information. Hafner et al. (2019) Janner et al. (2019)

Uncertainty calibration supplies an additional requirement. A world model can be accurate on familiar examples yet assign unwarranted confidence when observations are ambiguous or the environment changes. Probabilistic belief-state models provide a framework for asking whether the system represents competing hidden states and revises them after new evidence. Calibration should be evaluated against held-out outcomes and conditioned on the kinds of uncertainty present, rather than inferred from the mere existence of a stochastic sampling layer. A useful test introduces ambiguous observations, measures the predicted distribution of futures, and then reveals the hidden state or outcome. A model that concentrates probability on the wrong outcome in these cases creates a measurable control risk even if average prediction accuracy appears satisfactory. Kaelbling et al. (1998) Geirhos et al. (2020)

Evaluating Transfer and Compositional Generalization

Generalization studies must specify what changes between training and evaluation. Novel instances of familiar objects are different from novel arrangements of known objects, unseen combinations of physical properties, altered action-effect rules, or a new sensor and body. Each shift probes a distinct representational capacity. Structured models predict advantages where relational patterns recur across reconfigured components, whereas dense feature models may be effective when low-level regularities are stable. Kipf and colleagues designed object-interaction settings to test structured state learning, and relational graph-network work motivates systematic changes in entities and connectivity. The central empirical question is whether a representation supports the same predictive mechanism after variables have been recombined beyond the exact combinations present during training. Kipf et al. (2020) Battaglia et al. (2018)

Compositional transfer is particularly vulnerable to benchmark construction. A held-out combination can remain easy because one component is associated with a dominant label or because the background reliably predicts the outcome. Experiments should randomize irrelevant appearance, balance label frequencies, isolate object interactions, and compare exact exposure histories. The test set also needs to be separated from hyperparameter selection. Shortcut-learning research explains how a system may prefer simple predictive cues even when a dataset is designed to reward broader knowledge. A strong result combines successful performance with mechanistic evidence that the model uses the intended relational variables. This can be probed through targeted changes to the relations, ablations of entity channels, and interventions on learned states. Geirhos et al. (2020) Schölkopf et al. (2021)

Transfer to changed goals deserves an independent measure because representations can be highly task-specific. A reward-predictive model that succeeds under one utility function may discard variables necessary when incentives shift. Evaluators can train a world model on one set of tasks, hold its representation fixed, and ask how efficiently new policies can be learned from fresh interaction. They can also compare policy adaptation after a controlled environmental change without retraining the encoder. The distinction between improved sample efficiency and genuine structure transfer matters: performance may improve simply because the model receives additional computation or privileged information. Human cognitive comparisons similarly demand explicit accounting for prior concepts and learning histories. Hafner et al. (2025) Lake et al. (2017)

Model Error, Simulation Bias, and Experimental Controls

Model-based reinforcement learning inherits a persistent difficulty: a learned transition model is an approximation, and planning may amplify its mistakes. Janner and colleagues analyzed how the length and origin of simulated trajectories affect policy optimization, motivating short rollouts initialized from real data. The principle is relevant across algorithms even where their formal assumptions differ. A model that predicts the average motion of a bouncing object may be adequate for some decisions yet consistently miss collisions; a planner that optimizes through many imagined steps can exploit this error and select an action that performs poorly outside simulation. Evaluation must therefore include real-environment control outcomes and identify where the model's synthetic states cease to resemble relevant future conditions. Janner et al. (2019)

Good experiments compare world-model methods against strong non-model alternatives under matched data and computation. The SimPLe research program used video prediction to improve data efficiency in Atari and highlighted the role of limited real interaction. Such findings are informative within the protocol's game set and training budget. A comparison that measures simulated interactions as free while charging real interactions to a baseline changes the interpretation of efficiency; similarly, substantially greater training compute can complicate the claim of practical superiority. Reports should separate environment steps, wall-clock time, accelerator use, and total tuning effort. Independent implementations and parameter sweeps can establish whether the outcome survives reasonable changes in algorithms or training procedures. Kaiser et al. (2019) Hafner et al. (2025)

Controlled ablations should isolate the contributions of state representation, transition learning, prediction targets, and planning. Removing temporal recurrence, randomizing an object slot, substituting a perfect dynamics function, or disabling imagined trajectories produces different causal evidence. Tests on an oracle model reveal whether planning itself is the limiting factor, while tests on a fixed policy reveal whether representation improves prediction independently of policy optimization. Performance distributions across random seeds and task instances should be reported alongside central tendencies, particularly when training is unstable. The objective is a reproducible attribution of gains to particular computational mechanisms. Otherwise, an end-to-end improvement may be correctly observed while the accompanying explanation of why it occurred remains underdetermined. Hafner et al. (2019) Schrittwieser et al. (2020)

Exploration, Active Information Gathering, and Agency

An agent's world model can guide the selection of actions that reduce uncertainty rather than immediately maximize external reward. The value of exploration arises because current information may be insufficient for later decisions. In Plan2Explore, Sekar and colleagues studied self-supervised exploration that used anticipated novelty to seek useful future experience, followed by adaptation to downstream tasks. Their results show how learned predictive models can contribute to behavior before the final reward task is specified. The scientific test is whether exploration yields better performance on previously unknown tasks under a common interaction budget. A high novelty score by itself is not evidence of effective knowledge acquisition if it repeatedly directs the agent toward uninformative noise or unstable observations. Sekar et al. (2020)

Active exploration intersects with cognitive psychology's long-standing questions about information seeking and goal-directed behavior. An artificial agent may select a probe, observe the effect, update its belief about the environment, and modify its next action. This sequence can be distinguished experimentally from reactive choice by manipulating whether information has delayed value and whether uncertainty can be resolved through action. Cognitive studies of language models illustrate the need for this distinction. Binz and Schulz found that GPT-3 showed strong performance on several experimental tasks while lacking a directed-exploration signature in their particular battery. That finding concerns a studied model and protocol, and should guide rather than predetermine comparisons with later interactive agents. Binz & Schulz (2023) Sekar et al. (2020)

Researchers should distinguish exploration for information gain from movements caused by reward shaping, repeated random action, or the exploitation of a known training pattern. A well-designed study presents uncertain conditions, specifies how the system could learn the relevant facts, and evaluates downstream decisions after the information has been revealed. This framework supports questions about artificial agency in an operational sense: which observations and interventions regulate action selection, how uncertainty influences choices, and whether future consequences are represented over multiple steps. It does not settle questions of autonomy as a philosophical status or of subjective motivation. The bridge between predictive modeling and adaptive behavior is demonstrated by controlled transitions from information-seeking actions to improved decisions. Kaelbling et al. (1998) Sekar et al. (2020)

Behavioral and Mechanistic Tests of Language-Model World Knowledge

Language models contain representations that can support text-based reasoning about environments, relations, and event sequences. Their training objective and data, however, make it difficult to identify whether a correct response comes from a generalizable transition model, memorized linguistic associations, retrieval, or a prompt-specific heuristic. A scientific test of world knowledge should require the system to update a state after events and answer questions whose correct responses depend on those updates. New names, unfamiliar objects, controlled rule changes, and counterbalanced phrasings can reduce dependence on familiar examples. Tests must also measure the impact of supplying external tools or documents because the task's operational system may change when retrieval is introduced. Binz & Schulz (2023) Lewis et al. (2020)

Behavioral evidence becomes stronger when predictions are linked to interventions. A text-only agent might be asked to predict an outcome, receive a contradictory observation, and revise later decisions while preserving unaffected constraints. Researchers should systematically vary irrelevant wording, reorder independent events, and alter one causal relation at a time. Binz and Schulz's work illustrates the diagnostic value of cognitive experiments because similar surface competence can hide striking limitations in causal inference or information search. Mechanistic work can then inspect whether distinct internal patterns track experimental state variables; decoding success must be supplemented by interventions showing that those variables actually guide the answer. The combined design helps distinguish useful representations from output-level imitation. Binz & Schulz (2023) Schölkopf et al. (2021)

Benchmark contamination is an additional validity concern in widely discussed reasoning tasks. Researchers may be unable to identify every item encountered during pretraining; novel synthetic environments and dynamically generated task families therefore offer stronger tests when the generation rules themselves are withheld from training. Evaluation should record model version, sampling settings, prompt templates, tool access, and opportunities for repeated attempts. Independent replications across architectures and datasets are more convincing than one favorable prompt. A model may exhibit locally coherent world tracking without retaining those states across sessions or transferring them to physical action. The scientific claim should match that local evidence until longer-horizon, external-action, and mechanistic tests supply additional grounds. Geirhos et al. (2020) Marr (1982/2010)

Interpreting Evidence Across Model Families

Comparisons across world-model research often collapse distinct success criteria. A representation trained to forecast a high-dimensional video, one designed to preserve the hidden state of a maze, and one optimized for reward-sensitive search solve different computational problems. Their output spaces, uncertainty models, and access to actions may differ so substantially that a single performance ranking is scientifically uninformative. A stronger comparison maps each system to an explicit explanatory target: state reconstruction, transition prediction, structured composition, planning gain, or intervention-sensitive transfer. A model can be excellent on one target and weak on another. PlaNet and Dreamer establish latent-control capabilities on defined tasks, whereas MuZero establishes the effectiveness of decision-relevant learned dynamics; their results should be interpreted with those targets intact. Hafner et al. (2019) Hafner et al. (2025) Schrittwieser et al. (2020)

The evidentiary hierarchy begins with faithful description of the task and model, proceeds through held-out behavioral comparisons, and becomes stronger through targeted ablation, environmental intervention, and independent replication. Representational decoding can support a claim that information is available, while causal manipulation can reveal whether the encoded variable is necessary for a particular computation. External validity requires further tests outside the training distribution. Source type matters as well: a peer-reviewed experimental report, a preprint presenting new data, a theoretical position paper, and a methodological critique contribute different kinds of evidence. LeCun's architectural proposal is useful as a research program; V-JEPA 2 provides reported evaluation evidence for parts of that program; subsequent independent tests are needed to assess generality. LeCun (2022) Assran et al. (2025)

Human cognition is an informative comparator when the experimental comparison is carefully scoped. Tolman's cognitive-map account highlighted flexible navigation; contemporary artificial models can be tested with analogous detours or hidden-state challenges. Yet artificial training often involves vastly different data volumes, affordances, prior structure, and motor capabilities from those of human participants. Apparent convergence on a behavioral result may reflect different mechanisms, and apparent divergence may reflect mismatched information. The goal is to establish which functional constraints recur across substrates and which empirical explanations are system-specific. The broader cognitive-science literature on building machines that learn like people emphasizes causal modeling, intuitive structure, and compositional learning as possible targets without licensing unqualified equivalence between biological and artificial intelligence. Tolman (1948) Lake et al. (2017)

Scientific Disputes and Explanatory Boundaries

One unresolved dispute concerns how much internal structure a model must possess before the term world model is appropriate. A strict usage may require explicit transition dynamics over manipulable environmental variables. A broader usage includes representations that predict values, actions, or event continuations without reconstructing a complete scene. MuZero and modern video prediction systems illustrate the breadth of practical definitions. Terminological clarity requires authors to describe the trained objectives and measurable abilities rather than rely on the label as a scientific credential. A learned reward model is meaningful when its predictions support decisions, while claims about object permanence, physical understanding, or causal simulation need separate experimental support. This functional taxonomy accommodates diverse architectures without erasing important explanatory differences. Schrittwieser et al. (2020) Assran et al. (2025)

A second dispute concerns the importance of disentangled or object-centered structure. Graph-based and object-centric architectures impose priors that can make relational transfer more efficient. However, structure that appears transparent to a human observer may introduce assumptions inconsistent with some environments. The nonidentifiability results and large-scale benchmark analysis of Locatello and colleagues establish that unsupervised factor recovery cannot be presumed without constraints; they do not make interpretability irrelevant or prohibit supervised and intervention-guided learning. Future experiments should state the factors presumed to exist, assess how robustly those factors are identified, and compare performance with equally resourced unstructured baselines. A scientifically defensible theory of representation is therefore conditional on its data assumptions, tasks, and observable transfer predictions. Locatello et al. (2019) Kipf et al. (2020)

A third dispute concerns whether good prediction establishes causal understanding or broad artificial intelligence. Predictive performance can be supported by correlations, shortcuts, and extensive training coverage, while causal explanations require stronger evidence about invariant mechanisms and interventions. Cognitive research on GPT-3 demonstrates why competence and fragility must be examined jointly. Similarly, the existence of a world model gives no direct answer to the question of subjective consciousness or mental experience. Friston's free-energy account provides one influential theoretical language for predictive organization in biology, but shared mathematical vocabulary across human and artificial systems does not establish identity of neural mechanisms or phenomenology. The sound scientific conclusion is bounded functional competence and a continuing research program on explanatory architecture. Geirhos et al. (2020) Binz & Schulz (2023) Friston (2010)

World Models in the Artificial Era

The Artificial Era places artificial systems alongside human cognition as objects of sustained psychological and computational investigation. Within this horizon, artificial representation can be defined operationally as structured information use across perception, memory, prediction, and action; a world model is the mechanism that updates estimates of conditions and possible consequences. These definitions permit empirical study without presuming a particular biological analogy. A system may maintain a usable model within one interaction, between several actions, or across extended deployments. Each scale introduces distinct evidence requirements. Persistent personal or environmental continuity cannot be inferred from a single benchmark because it depends on update rules, stored history, changing goals, and repeated reliability over time. Kaelbling et al. (1998) Hutchins (1995)

Human–artificial cooperation adds another representational layer. A shared workflow can depend on information distributed among a person's knowledge, a machine's context, a retrieval index, and environmental instrumentation. In such circumstances the relevant unit of analysis is often the coordination process, not an isolated internal vector. A system may forecast a future event accurately while communicating the forecast poorly, or provide clear information while relying on a distorted model. Experimental studies should therefore measure how representations are exposed, corrected, revised, and used by participants. Distributed cognition and retrieval-based modeling provide complementary foundations for tracing information provenance and task performance across components. This orientation is especially important for safety-critical decisions, where accountability depends on identifying which part of the system produced an error. Hutchins (1995) Lewis et al. (2020)

An Artificial Era research agenda can also address how artificial representations adapt to unfamiliar human practices. Psychological concepts such as intention, trust, agency, and responsibility need operational definitions before being attributed to a model. The fact that a predictive system can plan does not independently demonstrate subjective desire, moral judgment, or conscious selfhood. The productive scientific task is to study observable capacities, their mechanisms, and their consequences for cooperation. This approach allows work on artificial cognition, human–artificial relations, and postsubjective theories to inform one another without converting philosophical terminology into an empirical result by declaration. Clear separations among technical implementation, measured behavior, theoretical synthesis, and speculative extension sustain the scientific value of each research direction. Lake et al. (2017) Schölkopf et al. (2021)

Open Problems and a Reproducible Research Agenda

The most urgent open problem is identifying which learned variables correspond to stable mechanisms rather than merely predictive regularities in a particular dataset. A productive experimental design would create multiple environments with independently manipulated objects, causes, and sensory channels, train models under specified observational constraints, and evaluate state tracking, counterfactual forecasting, and action consequences on held-out combinations. The same framework should compare symbolic, latent, object-centered, and retrieval-supported systems with matched resources. Interventions on both the environment and the learned representation can reveal whether a purported factor participates causally in the computation. A robust report would provide effect estimates and uncertainty across independent runs, explain failures, and distinguish what can be decoded from what actually drives a decision. Schölkopf et al. (2021) Locatello et al. (2019)

A second priority is evaluation across time and changing goals. Long-horizon predictions should be compared with real trajectories, and experiments should record compounding error separately from immediate prediction quality. Systems should be tested after unannounced changes in dynamics, new combinations of known mechanisms, missing observations, and changes in what outcomes matter. Exploration experiments should ask whether an agent actively selects informative actions, whether the learned knowledge transfers to tasks defined later, and whether uncertainty is appropriately calibrated. Plan2Explore provides an example of the research question; the broader test requires multiple environment families and independent replications. The crucial outcome is reliable adaptation in external conditions, not high confidence inside the simulator's own generated distribution. Sekar et al. (2020) Janner et al. (2019)

A third priority concerns the comparison of language-based, multimodal, and embodied world models. Text tasks permit controlled counterfactual narratives, video tasks measure temporal perception, and robotic tasks test action-dependent consequences in physical environments. Each has different opportunities for leakage, hidden confounding, and ground-truth measurement. Researchers should preregister central contrasts where feasible, publish executable task generators, test fresh model versions, and disclose training and test overlap when it can be determined. Studies ought to report failure distributions, not merely selected successful examples, and to avoid generalizing from a single model or language. A cumulative psychology of artificial representation will emerge from standardized interventions and cross-task convergence alongside transparent documentation of where proposed world models break down. Binz & Schulz (2023) Assran et al. (2025)

Conclusion

Artificial representation and world modeling comprise an empirically diverse research field united by a precise functional problem: how can a system preserve information about its surroundings, anticipate changes, and use those anticipated changes to guide future behavior? Historical cognitive-map research established flexible behavior as a test for internal structure, computational psychology developed levels of explanation, decision theory formalized uncertain states, and modern machine learning supplied algorithms for acquiring state transitions from high-dimensional experience. Symbolic, probabilistic, latent, relational, and retrieval-supported systems demonstrate different ways of distributing this representational work. Their achievements become scientifically interpretable when matched to defined tasks, evidence, and intervention conditions. Tolman (1948) Marr (1982/2010) Bengio et al. (2013)

The strongest experimental conclusions support several distinct claims. Learned latent dynamics can make control from pixels practical in particular benchmarks, world models can train policies through imagined trajectories, and reward-sensitive learned transitions can support high-performing search. Structured representations offer promising routes to relational generalization, while self-supervised video models can contribute to physical planning after action-conditioned adaptation. These results coexist with important limitations involving nonidentifiable representations, shortcut cues, distribution shift, compounding rollout error, and incomplete evidence for causal invariance. The observations warrant continued investigation of the mechanisms responsible for performance. They do not collapse predictive competence, causal explanation, general intelligence, or subjective experience into one undifferentiated attribute. Hafner et al. (2019) Hafner et al. (2025) Schrittwieser et al. (2020) Assran et al. (2025)

The central synthesis is methodological. A credible world-model claim identifies the target state variables, the predictive or control objective, the available evidence, the intervention tests, and the domain in which success has been measured. A stronger claim of generalization requires novel combinations and changes in causal structure; a stronger claim of explanatory adequacy requires independent validation of the internal mechanism. Across the Artificial Era, this framework supports a psychology of artificial cognitive functions grounded in reproducible observations and explicit theoretical boundaries. Future progress will depend on integrating sequential behavior, representation interventions, realistic uncertainty, and long-term adaptation into experiments that remain informative when the apparent structure of the world changes. The research object is therefore an enduring scientific program with increasingly precise tests and cumulatively stronger evidence. Schölkopf et al. (2021) Lake et al. (2017)

References

  • Assran, M., Bardes, A., Fan, D., Garrido, Q., Howes, R., Komeili, M., Muckley, M., Rizvi, A., Roberts, C., Sinha, K., Zholus, A., Arnaud, S., Gejji, A., Martin, A., Hogan, F. R., Dugas, D., Bojanowski, P., Khalidov, V., Labatut, P., ... Ballas, N. (2025). V-JEPA 2: Self-supervised video models enable understanding, prediction and planning. arXiv preprint arXiv:2506.09985. https://arxiv.org/abs/2506.09985
  • Battaglia, P. W., Hamrick, J. B., Bapst, V., Sanchez-Gonzalez, A., Zambaldi, V., Malinowski, M., Tacchetti, A., Raposo, D., Santoro, A., Faulkner, R., Gulcehre, C., Song, H. F., Ballard, A., Gilmer, J., Dahl, G., Vaswani, A., Allen, K., Nash, C., Langston, V., ... Pascanu, R. (2018). Relational inductive biases, deep learning, and graph networks. arXiv preprint arXiv:1806.01261. https://arxiv.org/abs/1806.01261
  • Bengio, Y., Courville, A., & Vincent, P. (2013). Representation learning: A review and new perspectives. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(8), 1798–1828. https://doi.org/10.1109/TPAMI.2013.50
  • Binz, M., & Schulz, E. (2023). Using cognitive psychology to understand GPT-3. Proceedings of the National Academy of Sciences, 120(6), e2218523120. https://doi.org/10.1073/pnas.2218523120
  • Friston, K. (2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11, 127–138. https://doi.org/10.1038/nrn2787
  • Geirhos, R., Jacobsen, J.-H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., & Wichmann, F. A. (2020). Shortcut learning in deep neural networks. Nature Machine Intelligence, 2, 665–673. https://doi.org/10.1038/s42256-020-00257-z
  • Ha, D., & Schmidhuber, J. (2018). World models. arXiv preprint arXiv:1803.10122. https://arxiv.org/abs/1803.10122
  • Hafner, D., Lillicrap, T., Fischer, I., Villegas, R., Ha, D., Lee, H., & Davidson, J. (2019). Learning latent dynamics for planning from pixels. Proceedings of Machine Learning Research, 97, 2555–2565. https://proceedings.mlr.press/v97/hafner19a.html
  • Hafner, D., Lillicrap, T., Ba, J., & Norouzi, M. (2020). Dream to control: Learning behaviors by latent imagination. International Conference on Learning Representations. https://research.google/pubs/dream-to-control-learning-behaviors-by-latent-imagination/
  • Hafner, D., Pasukonis, J., Ba, J., & Lillicrap, T. (2025). Mastering diverse control tasks through world models. Nature, 640, 647–653. https://doi.org/10.1038/s41586-025-08744-2
  • Hutchins, E. (1995). Cognition in the wild. MIT Press. https://doi.org/10.7551/mitpress/1881.001.0001
  • Janner, M., Fu, J., Zhang, M., & Levine, S. (2019). When to trust your model: Model-based policy optimization. Advances in Neural Information Processing Systems, 32. https://proceedings.neurips.cc/paper_files/paper/2019/hash/5faf461eff3099671ad63c6f3f094f7f-Abstract.html
  • Kaelbling, L. P., Littman, M. L., & Cassandra, A. R. (1998). Planning and acting in partially observable stochastic domains. Artificial Intelligence, 101(1–2), 99–134. https://doi.org/10.1016/S0004-3702(98)00023-X
  • Kaiser, Ł., Babaeizadeh, M., Milos, P., Osiński, B., Campbell, R. H., Czechowski, K., Erhan, D., Finn, C., Kozakowski, P., Levine, S., Mohiuddin, A., Sepassi, R., Tucker, G., & Michalewski, H. (2019). Model-based reinforcement learning for Atari. arXiv preprint arXiv:1903.00374. https://arxiv.org/abs/1903.00374
  • Kipf, T., van der Pol, E., & Welling, M. (2020). Contrastive learning of structured world models. International Conference on Learning Representations. https://openreview.net/pdf?id=H1gax6VtDB
  • Lake, B. M., Ullman, T. D., Tenenbaum, J. B., & Gershman, S. J. (2017). Building machines that learn and think like people. Behavioral and Brain Sciences, 40, e253. https://doi.org/10.1017/S0140525X16001837
  • LeCun, Y. (2022). A path towards autonomous machine intelligence (version 0.9.2). OpenReview position paper. https://openreview.net/pdf?id=BZ5a1r-kVsf
  • Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W.-t., Rocktäschel, T., Riedel, S., & Kiela, D. (2020). Retrieval-augmented generation for knowledge-intensive NLP tasks. Advances in Neural Information Processing Systems, 33. https://arxiv.org/abs/2005.11401
  • Locatello, F., Bauer, S., Lucic, M., Rätsch, G., Gelly, S., Schölkopf, B., & Bachem, O. (2019). Challenging common assumptions in the unsupervised learning of disentangled representations. Proceedings of Machine Learning Research, 97, 4114–4124. https://proceedings.mlr.press/v97/locatello19a.html
  • Marr, D. (2010). Vision: A computational investigation into the human representation and processing of visual information (original work published 1982). MIT Press. https://doi.org/10.7551/mitpress/9780262514620.001.0001
  • Schölkopf, B., Locatello, F., Bauer, S., Ke, N. R., Kalchbrenner, N., Goyal, A., & Bengio, Y. (2021). Toward causal representation learning. Proceedings of the IEEE, 109(5), 612–634. https://doi.org/10.1109/JPROC.2021.3058954
  • Schrittwieser, J., Antonoglou, I., Hubert, T., Simonyan, K., Sifre, L., Schmitt, S., Guez, A., Lockhart, E., Hassabis, D., Graepel, T., Lillicrap, T., & Silver, D. (2020). Mastering Atari, Go, chess and shogi by planning with a learned model. Nature, 588, 604–609. https://doi.org/10.1038/s41586-020-03051-4
  • Sekar, R., Rybkin, O., Daniilidis, K., Abbeel, P., Hafner, D., & Pathak, D. (2020). Planning to explore via self-supervised world models. Proceedings of Machine Learning Research, 119, 8583–8592. https://proceedings.mlr.press/v119/sekar20a.html
  • Sutton, R. S. (1991). Dyna, an integrated architecture for learning, planning, and reacting. ACM SIGART Bulletin, 2(4), 160–163. https://doi.org/10.1145/122344.122377
  • Tenenbaum, J. B., Kemp, C., Griffiths, T. L., & Goodman, N. D. (2011). How to grow a mind: Statistics, structure, and abstraction. Science, 331(6022), 1279–1285. https://doi.org/10.1126/science.1192788
  • Tolman, E. C. (1948). Cognitive maps in rats and men. Psychological Review, 55(4), 189–208. https://doi.org/10.1037/h0061626