Paper deep dive
A Unifying Perspective on Causal World Models: From Observations to Representations to Structure
Avinash Kori, Fabrizio Russo
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 8/16/2026, 2:30:46 AM
Summary
This paper proposes a unifying framework for Causal World Models (CWMs), defining them formally as Markov decision processes that integrate causal representation learning, causal discovery, and model-based decision-making. The authors argue that useful world models must capture entity properties, interactions, and environment dynamics to support causal reasoning, intervention, and counterfactual analysis. The paper formalizes the transition from raw observations to structured relational states and discusses identifiability conditions, distinguishing between strong identifiability and identifiability up to semantic equivalence.
Entities (8)
Relation Signals (6)
Causal World Models → isformalizedas → Markov Decision Process
confidence 95% · we formalise a CWM as a Markov decision process that links observations, latent states, actions, transition distributions, and utility
Causal World Models → relatesto → Identifiability
confidence 94% · Finally, we relate CWMs to the literature on identifiability
Relational Variables → constitutes → Structured Relational State
confidence 93% · The structured relational state is r_t... where r_t is the structured relational state from Definition 1
Causal World Models → integrates → Causal Representation Learning
confidence 92% · connecting world modelling with existing work in causal representation learning
Causal World Models → integrates → Causal Discovery
confidence 92% · connecting world modelling with existing work in ... causal discovery
Causal World Models → supports → Causal Reasoning
confidence 90% · ground WMs in representations and structures that support causal reasoning and informed decision-making
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:World Models (WM) are increasingly seen as a foundation for intelligent agents that can predict, plan, and act beyond their training distribution. In this paper, we study WMs from a causal perspective across multiple levels of abstraction, ranging from perceptual observations to building a conceptual representation of the structure governing the environment dynamics. We argue that useful WMs must go beyond generative capabilities alone: they should also capture entity properties, entity-to-entity interactions, and entity-to-environment interactions that determine and explain the dynamics of a system. We provide a formal definition of Causal WMs (CWMs) grounded in the tasks they are intended to support, connecting world modelling with existing work in causal representation learning, object-centric learning, causal discovery, structural causal models, and model-based decision-making. Finally, we relate CWMs to the literature on identifiability, clarifying when the components of a WM can be recovered from data and up to which equivalence. With this, we ground WMs in representations and structures that support causal reasoning and informed decision-making.
Tags
Links
- Source: https://arxiv.org/abs/2608.13456v1
- Canonical: https://arxiv.org/abs/2608.13456v1
Trouble viewing inline? Open PDF directly →
Full Text
36,101 characters extracted from source content.
Expand or collapse full text
A Unifying Perspective on Causal World Models: From Observations to Representations to Structure Avinash Kori Affiliation: Department of Computing, Imperial College London, London, UK a.kori21,fabrizio@imperial.ac.uk Fabrizio Russo Affiliation: Department of Computing, Imperial College London, London, UK a.kori21,fabrizio@imperial.ac.uk Abstract World Models (WM) are increasingly seen as a foundation for intelligent agents that can predict, plan, and act beyond their training distribution. In this paper, we study WMs from a causal perspective across multiple levels of abstraction, ranging from perceptual observations to building a conceptual representation of the structure governing the environment dynamics. We argue that useful WMs must go beyond generative capabilities alone: they should also capture entity properties, entity-to-entity interactions, and entity-to-environment interactions that determine and explain the dynamics of a system. We provide a formal definition of Causal WMs (CWMs) grounded in the tasks they are intended to support, connecting world modelling with existing work in causal representation learning, object-centric learning, causal discovery, structural causal models, and model-based decision-making. Finally, we relate CWMs to the literature on identifiability, clarifying when the components of a WM can be recovered from data and up to which equivalence. With this, we ground WMs in representations and structures that support causal reasoning and informed decision-making. 1 Introduction World models have become one of the central aspirations of modern Artificial Intelligence. They are invoked to describe agents that imagine, plan, and learn to act safely outside their training distribution (8; 16; 19; 26). The challenge is to build such models from observations, trajectories, interventions, and domain knowledge, while avoiding two unhelpful extremes: treating WMs as an unstructured black box with no explicit variables or mechanisms, or assuming in advance that the right variables and mechanisms have already been specified, so that they can be directly used for causal intervention. Recent work occupies different points along this spectrum. Predictive world models such as Dreamer learn compact latent states from sensory streams and use imagined rollouts for control across diverse tasks, but the learned states need not have an explicit causal or semantic interpretation (9). Hierarchical variants such as ResDreamer move closer to task-level reasoning by emphasising decision-relevant signals over photorealistic reconstruction, while still treating the internal representation largely as a learned latent substrate (31). Joint-embedding approaches such as LeJEPA address a different part of the problem by connecting predictive representation learning with formal identifiability guarantees, but they do not by themselves specify the causal variables, mechanisms, or intervention targets needed for acting in the world (13). Causal dynamics methods illustrate the complementary structural route: once state variables are given, sparse causal dependencies can improve generalisation and yield reusable state abstractions for downstream tasks (30; 29). Recent work on robust and general agents makes the motivation especially direct: agents need WMs that support causal reasoning about actions, mediation, and distribution shifts, not merely next-state prediction (23). However, even in this literature, WMs are often still operationalised primarily as transition models (22; 3). Figure 1: Conceptual overview of a Causal Ladder of CWM: illustrated as a four-level progression from perception to intervention and imagination. Rungs 1, 2 and 3 follow Pearl’s causal hierarchy (20; 1). At the base, Rung 1 represents raw sensory perception, where raw observations tx_t are processed by an inference model to infer entity-level features tv_t. In Rung 1–2, these features are lifted into the structured relational state tr_t, whose diagonal blocks retain entity attributes and whose off-diagonal blocks encode interactions. The graph rG_r represents causal mechanisms over this state; xG_x is used only when observations have themselves been decomposed into explicit variables. At Rung 2, agents can manipulate entities to maximise their utility. At the top, Rung 3 represents counterfactual and generative reasoning, enabling agents to analyse different possible scenarios. Yet the phrase World Model often names an ambition more than a well-specified object. Recurrent latent dynamics models, next-observation predictors, simulators, joint-embedding predictors, learned state abstractions, as well as the implicit knowledge stored in a large generative model have all been referred to as WMs (7; 4). The difficulty is that this common label conflates distinct modelling commitments. A predictor specifies what is likely to happen next; a representation compresses observations by mapping unstructured observations into usable structured variables representing relevant aspects of the world; and a causal model specifies mechanisms enabling interventional analysis. What remains underspecified is how to connect these pieces towards building robust WMs with known properties and guarantees: which variables should be represented, whether they are identifiable from data, which relations among them are causal, when an action-conditioned prediction can be ascribed an interventional-effect interpretation, and when additional information or domain knowledge is required. The problem we focus on is therefore how to turn observations into a CWM by combining three strands of work: causal representation learning (24; 2), addressing when latent variables can be recovered from unstructured observations; causal discovery (32), studying when cause-effect relationships over the structured variables (represented by rG_r; see Figure 1) can be recovered; and model-based decision-making (6), covering when actions and utilities can be used for controlled behaviour. When observations are explicitly decomposed into variables, xG_x denotes a graph over those observed or encoded variables. We do not assume x=rG_x=G_r: their correspondence requires an explicit conditional-independence-preserving representation map. Faithfulness relates each distribution to its own graph and, together with the Markov and sufficiency assumptions, yields the usual Markov-equivalence limit for observational discovery over rG_r (25; 27). Figure 1 summarises this view: a CWM is not a single monolithic predictor, but a sequence of modelling commitments that moves from perception to representation (Rung 1), from representation to causal structure (Rung 1-2), and from structure to intervention (Rung 2), utility, and counterfactual reasoning (Rung 3). We give particular attention to the state-abstraction pipeline at time t t→ϕt=(ti)i∈Ot→ψOtt,x_t φv_t=(v_t^i)_i∈ O_t _O_tr_t,\\ (1) where ϕφ maps the observed random variable tx_t to entity-indexed latent random variables tv_t, with tiv_t^i for each entity i in the current entity set OtO_t, and ψOt _O_t assembles the complete structured relational state tr_t. Its diagonal blocks retain entity attributes and its off-diagonal blocks encode interactions. The first map ϕφ raises a representation-learning question: under what assumptions can raw observations identify the entities and attributes that should enter the state representation? The second map ψOt _O_t raises a structure-learning question: which attributes and inter-entity relations should be represented as variables, and how can samples from those variables support causal discovery? These steps are coupled to, but distinct from, transition learning, policy or action modelling, utility specification, and the structural assumptions needed to use action-conditioned transitions to predict the effects of interventions (1; 30). Additionally, an action model endows CWMs with the capability to evaluate goal oriented actions ta_t and predict their effects on the structured relational state t+1r_t+1, as well as a decoder f^:ℛOt+1→ f:R_O_t+1→X translating t+1r_t+1 to the next observation t+1x_t+1. Concretely, this work makes two contributions. First, we formalise a CWM as a Markov decision process that links observations, latent states, actions, transition distributions, and utility (Section 2). Second, we argue that guarantees for CWMs should be understood component-wise, each imposing different assumptions and admitting different admissible equivalences, so that the right notion of identifiability for CWMs is the one that preserves the downstream reasoning task and the corresponding notion of control (Section 3). 2 Causal World Models Causal mechanisms govern the dynamics of the world, motivating a formal framework that can capture such mechanisms for a given environment. Our view of WMs integrates this property by introducing a structured formalisation that progresses from observations to latent representations (capturing entity-level abstractions), extends it to their interactions, modelling this as a causal discovery task. To ground the discussion, we use a classic example, introducing additional context as needed. Example 1. Let’s consider a robot learning to act in a tabletop environment containing a red block, a blue block, and a goal region. The robot receives images t∈x_t∈X and executes actions t∈a_t∈A, such as pushing the red block left or moving the gripper towards the goal. This is a single modelling problem, but it can be analysed at different levels of abstraction that link the raw pixels to choosing the optimal policy. Here, tx_t denotes the raw observation available to the robot, such as an image of the scene. The tuple tv_t captures entities such as the red block, blue block, gripper, and goal region; each tiv^i_t encodes attributes such as position, colour, shape, velocity, and contact state. The diagonal block ti,ir_t^i,i is exactly tiv_t^i, while an off-diagonal block such as ti,jr_t^i,j can encode the red block leaning on the blue block. We now formalise the notion of relational variables: Definition 1 (Relational Variables). Let O=∏i∈OiV_O= _i∈ OV^i be the entity-indexed latent state space for an entity set O, and let t=(ti)i∈O∈Ov_t=(v_t^i)_i∈ O∈V_O. For each ordered pair i,j∈Oi,j∈ O, let MijM_ij be a possibly empty component index set; for i=ji=j, MiiM_i indexes all coordinates or attribute blocks of tiv_t^i. Define m,ti,j:=m,ti,i=j,gmi,j(ti,tj),i≠j.r_m,t^i,j:= casesv_m,t^i,&i=j,\\ g_m^i,j(v_t^i,v_t^j),&i≠ j. cases The corresponding pair-specific node set is ri,j=mi,j:m∈Mij.V_r^i,j=\r_m^i,j:m∈ M_ij\. Thus ti,i≡tir_t^i,i≡v_t^i. With ℐr(O)=a=(i,j,m):i,j∈O,m∈Mij,I_r(O)=\a=(i,j,m):i,j∈ O,\ m∈ M_ij\, write a,t≡m,ti,jr_a,t≡r_m,t^i,j. The structured relational state is t=(a,t)a∈ℐr(O)∈ℛOr_t=(r_a,t)_a _r(O)∈R_O, and ψO _O in Eq. 1 denotes the assembly of these diagonal and off-diagonal blocks. Directed relations may occupy distinct (i,j)(i,j) and (j,i)(j,i) blocks; symmetry is imposed only for relation types declared symmetric. Definition 2 (Formal state description). At time t, the state is t=(t,t)∈×ℛOts_t=(x_t,r_t)∈X×R_O_t, where OtO_t is the current entity set and t=(a,t)a∈ℐr(Ot)r_t=(r_a,t)_a _r(O_t) is the structured relational state from Definition 1. Relational variables provide the state description on which a CWM operates. They turn entity-level latent representations into variables whose mechanisms can be learned, intervened on, and evaluated against task goals. We can therefore define a CWM as the decision model that couples this relational variable to observations, actions, transitions, and utility. Definition 3 (Causal World Models (CWM)). A CWM is a tuple W=(,,ℛOO,ℙ,),W= (X,A,\R_O\_O,P,U ), where ⊆⋃iℝdiX _iR^d_i is the observation space, A is the action space, ℛOR_O is the relational latent-state space for entity set O, and :×→ℝU:A×S is a utility function over actions and states. For time t, let t,t+1∈x_t,x_t+1∈X, t∈a_t∈A, t∈ℛOtr_t∈R_O_t, and t+1∈ℛOt+1r_t+1∈R_O_t+1. Under the Markovian state assumption, the observed one-step decision distribution is obtained by marginalising the latent relational variables, ℙ(t,t+1,t)=∫t,t+1ℙ(t,t+1,t,t+1,t),P(x_t,x_t+1,a_t)= _r_t,r_t+1P(x_t,x_t+1,r_t,r_t+1,a_t),\\ The action-conditioned observation-level transition then factorises as ℙxx′(a) _x (a) =ℙ(t+1=x′∣t=x,t=a) =P(x_t+1=x x_t=x,a_t=a) =∫t,t+1ℙ(t+1∣t+1)ℙ(t+1∣t,t=a)ℙ(t∣t). =\!\! _r_t,r_t+1\!\!\!\!\!\!\!\!\!\!\!\!\!P(x_t+1\! \!r_t+1)P(r_t+1\! \!r_t,a_t=a)P(r_t\! \!x_t). Remark 1. A CWM can be read as a Markov decision process over the formal state ts_t in Definition 2. The utility U evaluates a considered action and state with respect to the agent’s goal; in particular, one selects actions satisfying (a,t)≥(a′,t)U(a,s_t)≥U(a ,s_t) for alternatives a′a . The factorisation in Definition 3 has the following components: 1. ℙ(t+1∣t+1)→P(x_t+1 r_t+1)→ Prediction model, generates next observation given t+1r_t+1, following the structure in xG_x, 2. ℙ(t+1∣t)→P(r_t+1 r_t)→ Transition model, with ℙ(t∣t)P(a_t r_t) as the action model and ℙ(t+1∣t,t)P(r_t+1 r_t,a_t) as the action-conditioned transition. For discrete actions, ℙ(t+1∣t)=∑a∈ℙ(t+1∣t,a)ℙ(a∣t)P(r_t+1 r_t)= _a∈AP(r_t+1 r_t,a)P(a r_t). Under consistency and positivity, when actions are randomised or, more generally, there is no unmeasured action–outcome confounding given tr_t (10), this conditional identifies the controlled transition ℙ(t+1∣t;do(t=a))P(r_t+1 r_t;do(a_t=a)), where do(t=a)do(a_t=a) denotes externally setting the action to a (21); otherwise it is only a predictive conditional, 3. ℙ(t∣t)→P(r_t x_t)→ Inference model, first infers the entity-centric representation tv_t and then assembles the structured relational state tr_t, where ℙ(t∣t)=∫tℙ(t∣t)ℙ(t∣t)P(r_t x_t)= _v_tP(r_t v_t)P(v_t x_t), while following the structure rG_r. The definition is deliberately abstract: it does not prescribe a particular estimator, but identifies the probabilistic components that will later be associated with the literature on representation learning, causal discovery, and causal decision-making. The remark separates a CWM into components that must be learned or specified: an inference model from observations to relational variables, a transition or intervention model over relational variables, an action model, a prediction model back to observations, and a utility function. tx_ttr_tta_tt+1r_t+1t+1x_t+1Uϵx _x Figure 2: Causal influence diagram for the CWM factorisation. Circles denote random variables: grey circles are observations and blue circles are relational variables. The orange rectangle is the decision node, and the green diamond is the utility node. Solid arrows denote the dependencies in Definition 3: t↦tx_t r_t is inference, (t,t)↦t+1(r_t,a_t) r_t+1 is transition, t+1↦t+1r_t+1 x_t+1 is prediction. Dashed arrows indicate possible latent confounders. Figure 2 instantiates Definition 3 as a causal influence diagram, a graphical model that combines causal structure with decision and utility nodes (5). In our setting, the observation tx_t is encoded into relational variables tr_t, the action ta_t selects or modifies the transition mechanism leading to t+1r_t+1, and the next observation t+1x_t+1 is decoded from the updated relational state. The utility U evaluates the outcome relative to the agent’s goal, such as moving the red block into the goal region without undesired collisions. The diagram makes explicit that a WM should not only predict the next observation, but also support goal-oriented reasoning about which actions are useful for a task. With this interpretation in place, we can state the causal sufficiency assumption behind Definition 3, as shown in Figure 2. A WM can support causal reasoning only if the variables that jointly drive prediction and utility are available to the model, either as recorded components of the observation or as latent factors inferred from it. If a relevant common cause is missing from both levels, it may act as an unobserved confounder, possibly distorting causal inferences. This is the standard causal sufficiency assumption in causal discovery (25), which we make here to focus on the identifiability of the latent structure rather than the additional challenges of learning with unobserved confounders. Under this assumption, the CWM definition organises the modelling problem component-wise. If tx_t has already been decomposed into the causal variables of interest, the task may be causal discovery over xG_x together with action-conditioned transition learning. For unstructured observations such as raw pixels, we do not posit a pixel-level causal DAG; instead, causal representation learning recovers tv_t and the structured state tr_t, and causal discovery targets the causal graph rG_r. 3 CWM Identifiability Identifiability asks which aspects of a CWM are uniquely determined by a data regime. Exact recovery is one possible target, but latent-variable models more commonly admit transformations that leave the observable distribution unchanged (11; 12; 28). Causal graphs and interventional quantities introduce different observational ambiguities (27; 21; 1); hence identifiability must be stated for a particular component and downstream use. Intuitively, the question is whether two candidate CWMs compatible with the same evidence can still disagree on a subset of their components; the definitions below distinguish exact agreement from disagreements declared admissible. Definition 4 (Strong identifiability). Let W be a class of candidate WMs, and let ℙ(W)P_ D(W) denote the observable distribution induced by W∈W∈ W under a data regime D. The class W is strongly identifiable from D if, for any W1,W2∈W_1,W_2∈ W, ℙ(W1)=ℙ(W2)⇒W1=W2.P_ D(W_1)=P_ D(W_2) W_1=W_2. Strong identifiability is rarely the right target for CWMs. Entity permutations and invertible latent reparameterisations, including rescaling, can leave the learnt mechanisms unchanged. For example, two CWMs may place the red and blue blocks in opposite slots; after applying the same permutation to both relational indices and the entities affected by each action, they may induce identical predictions and decisions despite having unequal parameterisations. Requiring W1=W2W_1=W_2 would distinguish such semantically equivalent models: equivalence expresses this admissible variability. Definition 5 (Identifiability up to equivalence). Let ∼ be an equivalence relation over W, where W1∼W2W_1 W_2 means that the two models differ only by transformations that preserve the relevant world-model semantics. The class W is identifiable up to ∼ from D if ℙ(W1)=ℙ(W2)⇒W1∼W2.P_ D(W_1)=P_ D(W_2) W_1 W_2. For the CWM (Def. 3), Table 1 makes the equivalence ∼ of Definition 5 concrete across its representation, distributions, structure and utility, combining known recovery results with bespoke interface conditions. It distinguishes representation equivalences for tv_t, ti\v_t^i\, B,tiv_B,t^i, and ti,jr_t^i,j from compatibility conditions on the CWM interfaces. These equivalences jointly induce a representation alignment T, with ~=T() r=T(r) and, on formal states (Def. 2), T(,)=(,T())T(x,r)=(x,T(r)). We write = d= for equality in distribution. CWM Component Admissible Equivalence or Compatibility Condition tv_t: Latent state Invertible affine equivalence ~t=t+ v_t= Hv_t+c; under stronger mixture assumptions this reduces to permutation, coordinate-wise scaling, and translation (12). tii∈Ot\v_t^i\_i∈ O_t: Entity blocks Permutation of entity/slot indices: ~ti=tπ(i) v_t^i=v_t^π(i) (18). B,tiv_B,t^i: Entity-attribute block Within-entity affine identifiability (15); for a declared attribute partition, the block-preserving specialisation is ~B,ti=BB,tπ(i)+Bi v_B,t^i= H_Bv_B,t^π(i)+c_B^i. ti,jr_t^i,j, i≠ji≠ j: Inter-entity relational block Under our alignment of compositional blocks, following (14), require entity permutation with element-wise affine compatibility: ~ti,j=ijtπ(i),π(j)+ij r_t^i,j= D_ijr_t^π(i),π(j)+c_ij. Diagonal blocks inherit the equivalence of tiv_t^i. ψO _O: Relational assembly map Equivariance under a common permutation of entity labels in both indices; order invariance is required only for relation types declared symmetric. pf^♯(t+1∣t+1)p_ f_ (x_t+1 r_t+1): Prediction model Weak injectivity of the decoder push-forward (12) and local inverse consistency. With EO:=ψO∘ϕE_O:= _O φ, require EO≈f^−1E_O≈ f^-1 on the relevant support: EO(f^())≈E_O( f(r))≈r and f^(EO())≈ f(E_O(x))≈x on the data manifold. ℙ(t+1∣t)P(r_t+1 r_t): Transition model Equivariance in distribution under state alignment: ~t+1∣T()=T(t+1)∣ r_t+1 T(r) d=T(r_t+1) r. rG_r: Relational-variable causal graph Markov equivalence class (MEC) over the relational-variable DAG: same skeleton and same v-structures after admissible variable alignment (27). ℙ(t+1∣t,t)P(r_t+1 r_t,a_t): Action-conditioned transition Actions and their affected components correspond under T, and ~t+1|T(),a=T(t+1)|,a r_t+1 T(r),a d=T(r_t+1) r,a. Causal interpretation additionally requires consistency, positivity, and either randomised actions or no unmeasured action–outcome confounding given tr_t (10); cf. Remark 1. U: Utility model Utility preservation under transformation T: ~(a,T())=(a,) U(a,T(s))=U(a,s). Table 1: Component-wise equivalences and compatibility conditions for causal world-model identifiability. The representation conditions form a hierarchy of semantic commitments. The coordinates of tv_t need not have fixed names, whereas ti\v_t^i\ must preserve entity correspondence. Since 15 identify each whole entity vector tiv_t^i, interpreting B,tiv_B,t^i as a named attribute block additionally requires the transformation to preserve that partition. For ti,jr_t^i,j, the same entity correspondence must act on both indices and retain the diagonal aliases. The compositional construction of 14 motivates combining such blocks, without establishing their causal identifiability. Compatibility concerns whether the aligned representation can be used coherently by the rest of the CWM. Equivariance of ψO _O makes entity relabelling yield the same assembled relational state, while local inverse consistency of EO=ψO∘ϕE_O= _O φ and f f prevents the encoder–decoder interface from discarding relevant state information. The passive and action-conditioned dynamics and U must respect the same T so that aligned states produce aligned rollouts and values. The causal graph rG_r has a different status: its MEC is itself a graph equivalence (27), not an interface condition, and need not preserve action effects. Matched actions must affect components carried to one another by T: entity-directed actions follow the entity permutation, whereas relation-directed actions follow both relational indices. Action-conditioned transitions support causal planning only under the stated identification conditions (cf. Remark 1). Thus, equivalences specify what may change within individual components; compatibility conditions ensure that these changes propagate coherently through the CWM, yielding an admissible equivalence of the full model. Component-wise guarantees compose into a policy guarantee only when these interfaces are compatible. In particular, if the encoder and decoder satisfy the stated local inverse-consistency condition, the action-conditioned transition commutes with T in distribution, and actions, their affected components, and utilities are preserved under the same alignment, then corresponding policies induce aligned trajectory distributions and the same expected cumulative utility, for the same horizon and discount. Their utility-maximising policies therefore correspond under T, as in policy-preserving MDP abstractions (17). If transition effects, intervention targets, or utilities fail to commute with the same alignment, the per-component equivalences do not compose into a control guarantee. In summary, the relevant target is the weakest component-wise equivalence whose compatible interfaces preserve the intended prediction, intervention, or decision objective. This separates what can be identified from data from what must be supplied as structural assumptions or expert knowledge. 4 Conclusion This paper argues that CWMs should be understood as structured decision models rather than as monolithic predictors. We formalised a CWM as a model linking observations, representations and structure, and used this formalisation to connect representation learning, causal discovery, and model-based decision-making. The central contribution is a component-wise view, where each part of the model may be identifiable only up to admissible equivalence that enables the downstream reasoning. This decomposition clarifies what can be learned from data, what must be supplied by interventions or domain knowledge, and when predictive models can support causal reasoning. Future work should turn this perspective into a rigorous formalisation, followed by practical algorithms and empirical testing, while extending it to partial observability and latent confounding. Acknowledgements. Russo was funded by the ERC under the ERC-POC programme (grant number 101189053). References Bareinboim et al. (2022) E. Bareinboim, J. D. Correa, D. Ibeling, and T. Icard On pearl’s hierarchy and the foundations of causal inference. In Probabilistic and Causal Inference: The Works of Judea Pearl, p. 507–556. Cited by: Figure 1, Figure 1, §1, §3. Bengio et al. (2013) Y. Bengio, L. Yao, G. Alain, and P. Vincent Generalized denoising auto-encoders as generative models. Advances in neural information processing systems 26. Cited by: §1. Ceriscioli and Mohan (2025) M. Ceriscioli and K. Mohan Agents robust to distribution shifts learn causal world models even under mediation. In Proc. of NeurIPS, Cited by: §1. Ding et al. (2025) J. Ding, Y. Zhang, Y. Shang, Y. Zhang, Z. Zong, J. Feng, Y. Yuan, H. Su, N. Li, N. Sukiennik, et al. Understanding world or predicting future? a comprehensive survey of world models. ACM Computing Surveys 58 (3), p. 1–38. Cited by: §1. Everitt et al. (2021) T. Everitt, R. Carey, E. D. Langlois, P. A. Ortega, and S. Legg Agent incentives: a causal perspective. In Proc. of AAAI, Vol. 35, p. 11487–11495. Cited by: §2. Ge et al. (2026) L. Ge, H. Cai, R. Wan, Y. Xu, and R. Song A review of causal decision making. Journal of Artificial Intelligence Research 85. Cited by: §1. Guan et al. (2023) L. Guan, K. Valmeekam, S. Sreedharan, and S. Kambhampati Leveraging pre-trained large language models to construct and utilize world models for model-based task planning. In Proc. of NeurIPS, Vol. 36, p. 79081–79094. Cited by: §1. Ha and Schmidhuber (2018) D. Ha and J. Schmidhuber World models. arXiv:1803.10122 2 (3). Cited by: §1. Hafner et al. (2025) D. Hafner, J. Pasukonis, J. Ba, and T. Lillicrap Mastering diverse control tasks through world models. Nature 640, p. 647–653. Cited by: §1. Hernan and Robins (2020) M. A. Hernan and J. M. Robins Causal inference: what if. Chapman & Hall/CRC. Cited by: item 2, Table 1. Khemakhem et al. (2020) I. Khemakhem, R. Monti, D. Kingma, and A. Hyvarinen Ice-beem: identifiable conditional energy-based deep models based on nonlinear ica. Advances in Neural Information Processing Systems 33, p. 12768–12778. Cited by: §3. Kivva et al. (2022) B. Kivva, G. Rajendran, P. Ravikumar, and B. Aragam Identifiability of deep generative models without auxiliary information. In Proc. of NeurIPS, Vol. 35, p. 15687–15701. Cited by: Table 1, Table 1, §3. Klindt et al. (2026) D. Klindt, Y. LeCun, and R. Balestriero When does LeJEPA learn a world model?. arXiv:2605.26379. External Links: 2605.26379 Cited by: §1. Kori et al. (2025) A. Kori, B. Glocker, B. Schölkopf, and F. Locatello Unifying causal and object-centric representation learning allows causal composition. In Proc. of ICLR Workshop on Deep Generative Model in Machine Learning: Theory, Principle and Efficacy, Cited by: Table 1, §3. Kori et al. (2024) A. Kori, F. Locatello, F. Toni, B. Glocker, and F. D. S. Ribeiro Identifiable object centric representations via probabilistic slot attention. In Proc. of NeurIPS, Cited by: Table 1, §3. LeCun (2022) Y. LeCun A path towards autonomous machine intelligence version 0.9. 2, 2022-06-27. Open Review. Cited by: §1. Li et al. (2006) L. Li, T. J. Walsh, and M. L. Littman Towards a unified theory of state abstraction for MDPs. In Proc. of the International Symposium on Artificial Intelligence and Mathematics, Cited by: §3. Locatello et al. (2020) F. Locatello, D. Weissenborn, T. Unterthiner, A. Mahendran, G. Heigold, J. Uszkoreit, A. Dosovitskiy, and T. Kipf Object-centric learning with slot attention. In Proc. of NeurIPS, Vol. 33, p. 11525–11538. Cited by: Table 1. Matsuo et al. (2022) Y. Matsuo, Y. LeCun, M. Sahani, D. Precup, D. Silver, M. Sugiyama, E. Uchibe, and J. Morimoto Deep learning, reinforcement learning, and world models. Neural Networks 152, p. 267–275. Cited by: §1. Pearl and Mackenzie (2018) J. Pearl and D. Mackenzie The book of why: the new science of cause and effect. 1st edition, Basic Books, Inc., USA. Cited by: Figure 1, Figure 1. Pearl (2009) J. Pearl Causality. 2 edition, Cambridge University Press. External Links: Document Cited by: item 2, §3. Richens et al. (2025) J. Richens, T. Everitt, and D. Abel General agents need world models. In Proc. of ICML, Cited by: §1. Richens and Everitt (2024) J. Richens and T. Everitt Robust agents learn causal world models. In Proc. of ICLR, Cited by: §1. Schölkopf et al. (2021) B. Schölkopf, F. Locatello, S. Bauer, N. R. Ke, N. Kalchbrenner, A. Goyal, and Y. Bengio Toward causal representation learning. Proc. IEEE 109 (5), p. 612–634. Cited by: §1. Spirtes et al. (2000) P. Spirtes, C. Glymour, and R. Scheines Causation, prediction, and search, second edition. Adaptive computation and machine learning, MIT Press. Cited by: §1, §2. Taniguchi et al. (2023) T. Taniguchi, S. Murata, M. Suzuki, D. Ognibene, P. Lanillos, E. Ugur, L. Jamone, T. Nakamura, A. Ciria, B. Lara, et al. World models and predictive coding for cognitive and developmental robotics: frontiers and challenges. Advanced Robotics 37, p. 780–806. Cited by: §1. Verma and Pearl (1990) T. Verma and J. Pearl Equivalence and synthesis of causal models. In Proc. of UAI, p. 255–270. Cited by: §1, Table 1, §3, §3. Von Kügelgen et al. (2021) J. Von Kügelgen, Y. Sharma, L. Gresele, W. Brendel, B. Schölkopf, M. Besserve, and F. Locatello Self-supervised learning with data augmentations provably isolates content from style. Advances in neural information processing systems 34, p. 16451–16467. Cited by: §3. Wang et al. (2024) Z. Wang, C. Wang, X. Xiao, Y. Zhu, and P. Stone Building minimal and reusable causal state abstractions for reinforcement learning. In Proc. of AAAI, Cited by: §1. Wang et al. (2022) Z. Wang, X. Xiao, Z. Xu, Y. Zhu, and P. Stone Causal dynamics learning for task-independent state abstraction. In Proc. of ICML, Vol. 162, p. 23151–23180. Cited by: §1, §1. Xu et al. (2026) Y. Xu, L. Liu, W. Zhou, M. Feng, and H. Li Self-supervised hierarchical visual reasoning with world model. arXiv:2605.17537. Cited by: §1. Zanga et al. (2022) A. Zanga, E. Ozkirimli, and F. Stella A survey on causal discovery: theory and practice. Int. J. Approx. Reason. 151, p. 101–129. Cited by: §1.