Scientific Approaches to AGI

Continual learning (also lifelong learning) is the capacity of a learning system to acquire new knowledge and skills sequentially over time while retaining what it has already learned. Persistent memory refers to mechanisms, inside or outside a model's parameters, that store information across interactions. Together they address a gap in standard practice, in which most deployed models are trained once and then used with fixed parameters.

A person who works in a job for a year learns the organisation, the colleagues and the tasks, and carries that knowledge forward. A standard large language model deployment does not update its weights from its interactions; any apparent memory comes from text supplied in the context window, from retrieval over external stores, or from periodic retraining by the developer. Continual learning research asks how a system could learn on the job without destroying earlier competence.

The topic matters for general intelligence for two reasons. First, long-horizon work, such as managing a project over weeks, seems to require accumulating information. Second, AGI-REQ-001 lists memory as a candidate requirement, and Hendrycks et al. (2025) report in their abstract that current AI systems have critical deficits in foundational cognitive machinery, particularly long-term memory storage. Whether this deficit is a fundamental barrier or an engineering gap is Contested.

The core difficulty, catastrophic forgetting, is a long-recognised problem in connectionist research: when a network is trained on a new task, learning can overwrite what it learned before. Cognitive neuroscience offers a framework in which the brain uses complementary learning systems, with the hippocampus learning quickly and the neocortex consolidating knowledge slowly, as a solution to this problem (McClelland et al., 1995).

Kirkpatrick et al. (2017) introduced elastic weight consolidation, which slows learning on parameters important to earlier tasks, and demonstrated sequential learning on supervised and reinforcement-learning tasks. Reviews organised the field into families of method (Parisi et al., 2019; De Lange et al., 2021; Wang et al., 2023). With the rise of language models, interest shifted to memory outside the weights: retrieval-augmented generation (Lewis et al., 2020), operating-system-inspired memory management for long contexts (Packer et al., 2023), and memory streams in simulated agents (Park et al., 2023). A separate line studies whether training itself degrades: Dohare et al. (2024) report loss of plasticity in standard deep-learning methods in continual settings. Titans (Behrouz et al., 2024) proposes a neural memory module that learns at test time.

#4.1 The stability-plasticity trade-off

Consider a sequence of tasks $1,\dots,T$ with losses $L_t(\theta)$. After training on task $t$, parameters $\theta_t$ are obtained by minimising $L_t$ starting from $\theta_{t-1}$. Without constraints, $L_s(\theta_t)$ for $s<t$ may rise sharply, which is catastrophic forgetting. A regularisation-based approach adds a penalty. Elastic weight consolidation (Kirkpatrick et al., 2017) uses

where $\theta^{*}$ are the parameters after earlier tasks, $F_i$ is a diagonal Fisher-information estimate of how important parameter $i$ was, and $\lambda$ sets the strength. The trade-off is between stability (retaining old knowledge) and plasticity (learning new knowledge).

#4.2 Families of continual-learning methods

The surveys cited above group methods roughly as follows; the exact taxonomies differ between surveys.

• Regularisation-based: constrain changes to important parameters (for example elastic weight consolidation).
• Replay-based: store or generate examples of earlier data and mix them into new training.
• Architecture-based: allocate separate parameters or modules to new tasks.
• Optimisation-based and other approaches: shape gradient updates or meta-learn to reduce interference.

De Lange et al. (2021) review methods for classification tasks and emphasise careful comparison under controlled settings; Wang et al. (2023) provide a broader survey of theory, method and application.

#4.3 Loss of plasticity

A separate issue from forgetting is that a network trained for a long time on a changing stream may become less able to learn. Dohare et al. (2024) report that standard deep-learning methods lose plasticity in continual settings, and describe a modification (continual backpropagation) that maintains it in their experiments. This result is from a single research group's paper in Nature and is Preliminary until independent replication across settings.

#4.4 Memory outside the weights

Persistent memory for language-model systems is mostly implemented outside the parameters.

• Context window: information supplied in the prompt. It is immediate but bounded and is lost when the session ends unless saved.
• Retrieval: Lewis et al. (2020) combine a parametric language model with a non-parametric memory accessed by a retriever; see TECH-RAG-001.
• Managed memory tiers: MemGPT moves information between fast and slow memory stores using a technique the authors call virtual context management, and evaluates on documents longer than the context limit and on multi-session chat (Packer et al., 2023).
• Agent memory streams: Park et al. (2023) describe simulated agents that store experience records, synthesise them into reflections and retrieve them for planning.
• Skill libraries: Voyager stores executable skills for reuse in an open-ended embodied setting (Wang et al., 2023).
• Test-time memory modules: Behrouz et al. (2024) propose a neural long-term memory that learns to memorise at test time.

Memory outside the weights is not identical to learning: stored text can be retrieved but does not necessarily change how the model generalises. Whether it is functionally sufficient for the purposes of general intelligence is open.

Memory-system terminology borrowed from psychology (episodic, semantic, procedural) is often applied to agent memory by analogy. The analogy is informative but not exact, and the sources cited use it with varying precision.

• Research benchmarks. Sequential image-classification benchmarks are standard for comparing methods (De Lange et al., 2021).
• Long-context assistants. Retrieval and tiered memory are used so that assistants can refer to documents and previous sessions (Lewis et al., 2020; Packer et al., 2023). Specific product features change quickly and are not documented here.
• Simulated agents. Generative agents with memory streams (Park et al., 2023) and Voyager (Wang et al., 2023) are laboratory demonstrations in simulated or game environments, not deployed autonomous workers.
• Personalisation and on-device adaptation. Continual adaptation to a user is a motivating application in the surveys, though the sources reviewed here do not document specific deployments.

Advantages. A system that learns continually could adapt to new information without full retraining, personalise to users, and accumulate skills. External memory is auditable and editable in a way that weights are not.

Limitations.

• Catastrophic forgetting remains a documented problem for sequential training (Kirkpatrick et al., 2017); methods reduce but, in the surveys reviewed, do not remove it under all conditions.
• Evaluation settings in the literature are often simplified compared with real streams of experience (De Lange et al., 2021 emphasise controlled comparison).
• Continual updating raises safety and security questions: a system that learns from interactions can be influenced by adversarial inputs, and its behaviour may drift from the evaluated version. See SAFE-SEC-001 and SAFE-ALIGN-001.
• Stored memory raises privacy issues (see SAFE-PRIV-001).

Open questions.

1. Can context-and-retrieval systems substitute for weight updates in the long run, or is parametric continual learning needed?
2. How should a system decide what to remember, what to forget and what to consolidate?
3. How can continual learners be evaluated for safety when the system changes after evaluation?
4. Does loss of plasticity occur in large-scale pre-training and fine-tuning pipelines in the same way as in the reported settings?

As of 2026-10-15, per the sources reviewed: catastrophic forgetting in sequential neural-network training is an established phenomenon (Kirkpatrick et al., 2017); several families of mitigation are documented in surveys (Parisi et al., 2019; De Lange et al., 2021; Wang et al., 2023); loss of plasticity in long continual settings has been reported (Dohare et al., 2024); and external memory and retrieval systems are widely used with language models (Lewis et al., 2020; Packer et al., 2023). Whether language-model systems can learn persistently and safely from experience at scale is not demonstrated in the sources reviewed. Whether memory is the binding constraint on general capability is Contested. For the continual-learning behaviour of specific current systems, consult the Living Reference collection (MODEL-001); the training procedures of closed systems are generally undisclosed.

This section is descriptive and registers no forecast.

• Near term (0 to 3 years). Expect further development of retrieval, agent-memory scaffolds and test-time memory modules (Behrouz et al., 2024), and evaluations that measure memory use over long tasks.
• Medium term (3 to 10 years). Research directions include integrating parametric and non-parametric memory and updating models safely from deployment experience. The sources reviewed give no evidence about which approach will prevail.
• Long term (10+ years). Systems that learn continuously throughout deployment would change how evaluation, auditing and governance work, because the system under evaluation would not be fixed. This is a conceptual implication, not a prediction.

Dated, testable predictions are collected in the Forecast Register (FCST-REG-001).

• Foundational Concepts:
  • TECH-ML-001 Machine Learning: Learning paradigms and generalisation.
  • AGI-PATH-001 Pathways to AGI: Where memory and learning fit among pathways.
  • FOUND-013 The Brain and Biological Neurons: Biological memory consolidation.
• Related Methods & Technologies:
  • TECH-RAG-001 Retrieval-Augmented Generation: Retrieval as external memory.
  • TECH-AGENT-001 Autonomous AI Agents: Agents that use memory across steps.
  • AGI-WM-001 World Models: Models that must be kept current.
• Advanced Implications & Future Trajectories:
  • AGI-REQ-001 Requirements for General Intelligence: Memory as a candidate requirement.
  • SAFE-ALIGN-001 The AI Alignment Problem: Alignment of systems that change after deployment.
  • ASI-AIRD-001 Automated AI Research: Long-horizon research work needs accumulating knowledge.

Scientific Pathways to AGI represent the competing architectural, algorithmic, and systemic research trajectories hypothesized to lead to Artificial General Intelligence (AGI). As the field lacks an agreed account of the prerequisites for human-equivalent general intelligence, research laboratories pursue divergent paradigms—ranging from scaling autoregressive foundation models and integrating explicit test-time search, to developing neuro-symbolic systems and embodied world models.

The pursuit of AGI is associated with very large reported capital and infrastructure investment and with regulatory attention [VERIFICATION NEEDED: no dated, sourced investment figure is cited; earlier "hundreds of billions of dollars" claim removed]. Identifying the correct scientific pathway is not merely an academic exercise; it determines which architectural constraints must be overcome.

The significance of differentiating these pathways lies in their differing assumptions about the nature of intelligence:

1. The Scaling Hypothesis (Hypothetical): Proponents hold that general capability could emerge from compressing vast amounts of data using deep networks and massive compute.
2. The Algorithmic Hypothesis: Current architectures are fundamentally flawed regarding causal reasoning; genuine intelligence requires explicit, structured symbolic planning or recurrent continuous state variables.
3. The Embodiment Hypothesis: Intelligence cannot arise in a digital vacuum; it requires physical grounding, active intervention in a physics-governed environment, and persistent interaction.

Evaluating these pathways requires distinguishing demonstrated empirical capabilities from theoretical extrapolations and marketing terminology.

Historically, the pathway to general intelligence experienced massive paradigm shifts. A simplified periodization distinguishes a symbolic tradition (1950s onward), associated with the Physical Symbol System Hypothesis (Newell and Simon, 1976) [VERIFICATION NEEDED: not cited in section 11], from connectionist approaches, which date to the 1940s-1960s and resurged in the 1980s. These strands overlapped rather than succeeding one another; see the history entries in Volume II.

Since deep learning advances around 2012 and the Transformer architecture (2017) [VERIFICATION NEEDED: cite Vaswani et al. (2017) and a 2012 milestone], scaled deep learning has been the most heavily resourced approach. Brown et al. (2020) reported that a 175-billion-parameter language model showed few-shot in-context learning across many NLP benchmarks, which proponents of scaling treat as evidence that broad capability can arise from a single pretraining paradigm.

#4.1 Pathway 1: Scaling Foundation Models (The "Bitter Lesson")

This pathway posits that scaling autoregressive multimodal transformers—increasing parameter count ($N$), dataset size ($D$), and compute ($C$)—is sufficient for AGI.

• Core Mechanism: Next-token prediction across text, audio, and visual patches. Proponents argue (a contested hypothesis) that to predict human text well, a model must implicitly learn underlying causal, logical and physical regularities of the world.
• Strongest Evidence: Empirical power-law relationships between test loss and model size, data and compute (Kaplan et al., 2020; Hoffmann et al., 2022), over bounded regimes; the two studies differ on the optimal allocation of parameters and data. Whether abilities such as in-context learning "emerge" abruptly or reflect metric choice is actively debated [Contested].
• Limitations: Single-pass autoregressive decoding does not natively backtrack, though chain-of-thought and sampling-based search can add deliberation (Pathway 2). Current models are reported to use far more text than children do, a comparison that is contested because of evolutionary priors and sensory input volume [VERIFICATION NEEDED: cite a source such as Frank (2023)]. Villalobos et al. (2024) project that the stock of public human-generated text could be fully used between 2026 and 2032 (a forecast with substantial uncertainty; see section 9).

#4.2 Pathway 2: Test-Time Compute & Deliberative Search

Recognizing the limitations of fast, "System 1" autoregressive generation, this pathway incorporates explicit "System 2" search mechanisms.

• Core Mechanism: Augmenting the generator with algorithms like Monte Carlo Tree Search (MCTS), self-reflection, or verification reward models. The model spends significant compute during inference (test-time) to explore multiple reasoning paths, critique them, and select the optimal output.
• Evidence: AlphaCode (Li et al., 2022) generated large numbers of candidate programs and filtered them, and was evaluated on Codeforces competitions. OpenAI reports that its o1 models were trained with reinforcement learning to produce chain-of-thought reasoning; the architecture is undisclosed [VERIFICATION NEEDED: cite the OpenAI o1 system card; capability claims and benchmark scores are not asserted here].
• Limitations: Designing reliable, automated verifiers (reward models) for open-ended, ambiguous real-world domains (where the "correct" answer isn't a verifiable math proof) remains unsolved.

#4.3 Pathway 3: Joint Predictive World Models (JEPA / LeCun)

Championed by researchers like Yann LeCun, this pathway rejects generative autoregression in favor of Joint Embedding Predictive Architectures (JEPA).

• Core Mechanism: The system observes a state, takes an action, and predicts the abstract latent representation of the next state (rather than predicting exact pixels or text tokens). Proponents hypothesize that this encourages the model to learn abstract, invariant structure of the world [Hypothetical].
• Claimed advantages (hypothesized): LeCun (2022) is a position paper proposing this architecture; it reports no empirical results of sample-efficient learning in robotics. [VERIFICATION NEEDED: cite specific empirical JEPA results (e.g. I-JEPA, V-JEPA) with benchmarks if any efficiency claim is to be made.]
• Limitations: Still largely in the fundamental research phase; has not yet demonstrated the broad cognitive generality and language mastery of scaled transformers.

#4.4 Pathway 4: Neuro-Symbolic Integration

This pathway attempts to marry the statistical pattern recognition of deep learning with the logical rigor of classical symbolic AI.

• Core Mechanism: Neural networks handle perception and intuition (mapping raw data to concepts), while a discrete symbolic engine (e.g., a knowledge graph or automated theorem prover like Lean) handles rigorous, multi-step deductive reasoning.
• Evidence: AlphaGeometry (Trinh et al., 2024) solved 25 of 30 olympiad-level plane-geometry problems by combining a language model with a symbolic deduction engine. This is a narrow-domain result (Euclidean plane geometry) and does not by itself indicate a route to AGI; the symbolic engine verifies deductions within a closed formal domain, while the neural component can still propose unproductive constructions. Nature issued an Author Correction for the paper (Nature 627, March 2024).
• Limitations: The "Symbol Grounding Problem"—creating a differentiable, seamless interface between continuous neural vectors and discrete symbolic logic remains a profound engineering bottleneck.

#4.5 Pathway 5: Open-Ended Reinforcement Learning & Self-Play

Inspired by biological evolution and by systems such as AlphaGo Zero and AlphaZero.

• Core Mechanism: Placing agents in complex, procedurally generated environments or multi-agent simulations where they learn through continuous trial-and-error, self-play, and open-ended curriculum learning.
• Evidence: AlphaGo Zero learned Go through self-play reinforcement learning without human game data (Silver et al., 2017); AlphaZero generalized the approach to chess, shogi and Go (Silver et al., 2018). Both concern perfect-information board games. AlphaStar (Vinyals et al., 2019) is a StarCraft II system; it should not be described as trained without human data [VERIFICATION NEEDED: confirm the role of human replay data in AlphaStar training from Vinyals et al., 2019].
• Limitations: Real-world environments (unlike board games) lack a fast, perfectly accurate simulator and a dense, unambiguous reward signal.

Progress toward AGI requires achieving specific sub-milestones, creating hybrid architectures across the above pathways:

• Long-Horizon Planning: The ability to hold a goal over days or weeks, executing sub-tasks and recovering from failures.
• Continual / Lifelong Learning: The ability to continuously update knowledge matrices with new information without suffering catastrophic forgetting of prior knowledge.
• Causal Abstraction: Moving from $P(y \mid x)$ (correlation) to $P(y \mid \text{do}(x))$ (intervention and counterfactual reasoning).

• AlphaGeometry (Trinh et al., 2024): An example of neuro-symbolic integration in one narrow domain, combining a neural language model with a symbolic deduction engine (25 of 30 olympiad-level geometry problems solved).
• OpenAI o1: Described by OpenAI as trained with reinforcement learning to produce chain-of-thought before answering; used here as an example of Pathway 2 as reported by the developer, with the architecture undisclosed [VERIFICATION NEEDED: system card citation].

#Economic and Infrastructure Constraints

The pursuit of these pathways is not entirely a scientific question; it is an infrastructure problem. Scaling is reported to require very large GPU clusters and electrical power [VERIFICATION NEEDED: cite dated disclosures; unsourced figures removed]. If scaling returns diminish, investment in data centers could fall, which some commentators argue would slow progress; this is speculation.

#Unresolved Scientific Questions

1. Is language sufficient for AGI? Can a system achieve general intelligence strictly by reading text, or does general intelligence require multi-modal grounding and active manipulation of physical physics?
2. Can current architectures achieve sample efficiency? Humans are often argued to learn from much less data than current models, though the comparison is contested. Do neural networks lack a fundamental inductive bias that permits this efficiency?

Frontier laboratories currently pursue mainly scaled pretraining combined with reinforcement-learning post-training and test-time compute (Pathways 1 and 2) [VERIFICATION NEEDED: cite sources, e.g. Ouyang et al. (2022) for RLHF, Rafailov et al. (2023) for DPO, if specifics are retained]. This describes engineering prevalence, not scientific consensus: researchers disagree on whether the approach suffices for AGI, and no survey establishing a consensus is cited.

Public human text data may be exhausted for frontier training between roughly 2026 and 2032 (Villalobos et al., 2024); synthetic data is a proposed remedy whose effectiveness at scale is untested. The earlier medium- and long-term bullets (including a neuromorphic route) had no supporting source and are removed. Registered AGI timing forecasts are in AGI-TIME-001.

• Foundational Concepts:
  • AGI-001 Artificial General Intelligence: Theoretical thresholds and evaluation.
• Related Methods & Technologies:
  • TECH-LLM-001 Large Language Models: The primary artifact driving Pathway 1.
  • TECH-RL-001 Reinforcement Learning: The optimization substrate for Pathway 5 and test-time compute.
• Advanced Implications & Future Trajectories:
  • ASI-EXP-001 The Intelligence Explosion Hypothesis: The theoretical consequence of completing the AGI pathway.

A world model is a learned or constructed internal model that predicts how the state of an environment changes over time and in response to an agent's actions, and that can be used to plan, imagine outcomes or interpret observations. In AI research the term covers explicit latent-dynamics models used in reinforcement learning, video- and interaction-generating models, and the question of whether large language models acquire implicit models of the processes that generated their training data.

Planning requires predicting consequences. An agent that can predict what will happen if it acts can evaluate options without executing them, which reduces the number of costly or dangerous real-world trials. Proposals for general machine intelligence frequently place a world model at the centre of the architecture (LeCun, 2022), and AGI-PATH-001 lists world-model approaches as one of several pathways. Whether a world model is necessary for general intelligence is Contested (see AGI-REQ-001); this article describes what is meant by the term and what has been demonstrated.

Three related meanings are often conflated.

1. Control-oriented world models. A model of dynamics learned so that an agent can plan or train a policy in imagination (Ha & Schmidhuber, 2018; Hafner et al., 2023).
2. Generative environment models. Models that produce video or interactive environments conditioned on text, images or actions (Bruce et al., 2024).
3. Implicit world models in sequence predictors. The hypothesis that a model trained only to predict the next token or move develops internal structure that mirrors the underlying world (Li et al., 2023; Gurnee & Tegmark, 2023; Vafa et al., 2024).

The idea of an agent with an internal model of its environment is long-standing in cognitive science and control theory. In the neural-network literature, Ha and Schmidhuber (2018) presented a system in which a generative model of the environment is learned from observations and a small controller is trained within the model's own simulated dynamics. The same period saw model-based reinforcement learning advance in board games: MuZero learned a model sufficient for planning without being given the rules of the game, and reached strong performance on Go, chess, shogi and Atari (Schrittwieser et al., 2020).

Hafner et al. (2023) introduced DreamerV3, which uses a learned world model with a single fixed configuration across many tasks, and which the authors state is the first algorithm to collect diamonds in Minecraft from scratch without human data or curricula. LeCun (2022) proposed an architecture for autonomous agents built around a world model trained with joint-embedding predictive methods rather than pixel-level generation. Generative video and interactive-environment models followed: Genie is described by its authors as an approximately 11-billion-parameter foundation world model trained on unlabelled Internet video that can generate action-controllable virtual worlds from prompts (Bruce et al., 2024). V-JEPA 2 extended joint-embedding methods to video and, with a small amount of robot data, to planning in robot manipulation (Assran et al., 2025).

In parallel, interpretability studies asked whether sequence models already contain world models. Li et al. (2023) trained a GPT-style model on Othello move sequences and found internal board-state representations that could be intervened on. Gurnee and Tegmark (2023) reported linear representations of space and time in Llama-2 models. Vafa et al. (2024) proposed metrics based on the Myhill-Nerode theorem and found that models which do well on standard tests had much less coherent implicit world models than those tests suggested.

#4.1 Formal structure

A world model is commonly formalised as a stochastic transition model over latent states. Given observations $o_t$, actions $a_t$ and latent state $z_t$, a typical model comprises:

• an encoder, $z_t \sim q_\phi(z_t \mid z_{t-1}, a_{t-1}, o_t)$;
• a dynamics (prior) model, $\hat{z}_t \sim p_\theta(\hat{z}_t \mid z_{t-1}, a_{t-1})$;
• a decoder or predictive head, $\hat{o}_t \sim p_\theta(\hat{o}_t \mid z_t)$, and a reward predictor $\hat{r}_t$.

Training maximises a variational objective that balances reconstruction or prediction accuracy against closeness between the posterior and prior over latent states. In this family the learned dynamics can be rolled forward from a starting state to produce imagined trajectories, and a policy is optimised on those trajectories (Ha & Schmidhuber, 2018; Hafner et al., 2023). This describes the general design; the specific objectives, architectures and normalisation tricks differ by system.

#4.2 Planning with a learned model

Given a model, an agent chooses actions by search or optimisation over imagined futures. MuZero combines a learned model of reward, value and policy-relevant quantities with Monte Carlo tree search (Schrittwieser et al., 2020): the learned model need not reconstruct observations, only to predict what matters for planning. This shows that a world model, in the sense of a model useful for decisions, need not be a faithful simulator.

#4.3 Prediction in representation space

LeCun (2022) argues for predicting in an abstract representation space instead of predicting every detail of the input, because many details of the world are not predictable. V-JEPA 2 follows this approach for video: the authors report pre-training on over 1 million hours of internet video and images, and a robot-planning variant post-trained on fewer than 62 hours of unlabelled robot video, deployed zero-shot on robot arms in two laboratories for pick-and-place tasks (Assran et al., 2025). These are results reported by the developing group and are Preliminary.

#4.4 Implicit world models

For language and sequence models, the claim is that internal activations encode latent variables of the data-generating process. Evidence takes the form of probes (decoding the variable from activations) and interventions (changing the representation and observing changes in output). Li et al. (2023) found a nonlinear board representation in the Othello setting; Gurnee and Tegmark (2023) found linear encodings of geographic and temporal variables in Llama-2 models. Vafa et al. (2024) tested coherence on game playing, logic puzzles and navigation, and found that the implicit models were much less consistent than standard tests indicated, which can cause failures on closely related tasks. Probing results show that information is present and decodable; they do not by themselves show that the model uses a coherent model for inference.

A further classification concerns explicit versus implicit models, deterministic versus stochastic dynamics, and observation-space versus latent-space prediction. The choice among these determines the computational cost of rollouts and the kinds of errors the model compounds.

• Game playing and simulated control. MuZero and DreamerV3 are demonstrations in simulated or game environments, not deployed products (Schrittwieser et al., 2020; Hafner et al., 2023).
• Generative environments. Genie generates playable worlds from prompts; the authors also describe using its learned latent action space to train agents to imitate behaviours from unseen videos (Bruce et al., 2024). This is a laboratory demonstration.
• Robot planning. V-JEPA 2-AC, per its developers, performed zero-shot pick-and-place on Franka arms in two laboratories without data collected from those robots (Assran et al., 2025). This is a laboratory result reported by the developers; it has not been shown here to extend to open-ended tasks. For the broader field see TECH-ROBOT-001.
• Interpretability studies. The Othello and Llama-2 results (Li et al., 2023; Gurnee & Tegmark, 2023) are used as evidence in debates about what language models represent. See FOUND-014 for the related philosophical question.

Advantages. Planning in a model can make learning more sample-efficient than purely model-free learning; a model supports counterfactual evaluation of actions; and a single model can serve multiple tasks. (The claim of sample efficiency is an expected benefit of the approach; the sources reviewed here demonstrate performance but do not compare data efficiency in every case.)

Limitations.

• Compounding error. Small prediction errors accumulate over long rollouts, limiting the planning horizon.
• Stochasticity and unpredictability. Many aspects of the world cannot be predicted in detail, which motivates prediction in latent space (LeCun, 2022) but leaves open how to choose what to predict.
• Evaluation difficulty. A model may look plausible while being physically or logically inconsistent. Vafa et al. (2024) show standard task accuracy can overstate the coherence of implicit models.
• Generative realism is not understanding. A model that produces visually convincing video need not have an accurate model of physics. The sources reviewed here describe generation, not physics-accurate prediction.
• Scale and cost. Training large video models is computationally expensive; see ECON-COMP-001.

Open questions.

1. Is a learned world model necessary for general intelligence, or can language-based reasoning over text and tools substitute for it? Contested.
2. Do language models contain usable world models, or only decodable correlates? Contested.
3. How should world-model quality be measured so that tests catch inconsistency rather than reward surface fluency?
4. Can world models learned from passive video support planning in the real world at scale?

As of 2026-10-15 the sources reviewed support the following. Learned world models enable strong performance in simulated and game environments (Schrittwieser et al., 2020; Hafner et al., 2023), on the authors' reports. Generative interactive models and video-based predictive models exist as research systems (Bruce et al., 2024; Assran et al., 2025). Probing studies find structured internal representations in language and sequence models (Li et al., 2023; Gurnee & Tegmark, 2023), while a metric-based study finds substantial incoherence in implicit models (Vafa et al., 2024). The overall picture of what implicit world models LLMs have is Contested. Figures such as parameter counts, training-data sizes and benchmark scores in the papers above are given at the dates of those papers; for current systems consult the Living Reference collections (MODEL-001, FACT-001). The architectures and training data of closed commercial video and world-model systems are, in general, undisclosed, and this article makes no claims about them.

This section is descriptive and registers no forecast.

• Near term (0 to 3 years). Expect continued work on video-based predictive models, interactive-environment generators and benchmarks for physical and causal consistency.
• Medium term (3 to 10 years). Whether world-model approaches provide a path to general capability, or are one component alongside language-based reasoning and tools, will depend on evidence that does not yet exist in the sources reviewed.
• Long term (10+ years). Speculative; connections to embodied systems and scientific simulation are discussed in TECH-ROBOT-001 and APP-SCI-001.

• Foundational Concepts:
  • AGI-PATH-001 Pathways to AGI: World-model approaches among proposed routes to AGI.
  • TECH-RL-001 Reinforcement Learning: Model-based reinforcement learning.
  • AGI-REQ-001 Requirements for General Intelligence: World models as a candidate requirement.
• Related Methods & Technologies:
  • TECH-DIFF-001 Diffusion Models: A common basis of generative video models.
  • TECH-ROBOT-001 Robotics and Embodied AI: Use of world models for robot planning.
  • AGI-CL-001 Continual Learning and Persistent Memory: Keeping a world model up to date.
• Advanced Implications & Future Trajectories:
  • FOUND-014 Consciousness and Machine Understanding: Whether internal models amount to understanding.
  • AGI-TEST-001 Testing Claims of General Intelligence: Testing coherence of world models.