{"id":"8dba972c-f5a0-4e84-929f-b168e0800cd1","arxiv_id":"2608.13456","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A causal world model is defined as a decision model over latent relational state, with identifiability guarantees stated component-wise up to admissible equivalences.","lead":"This paper proposes a formal definition of a Causal World Model as a Markov decision process over relational latent variables, connecting perception, representation, and causal structure. A smart generalist might read it to see how AI agents could be built with explicit, identifiable causal components rather than opaque deep-network predictors.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Definition 3's observation-level transition omits the posterior correction P(r_t|x_t,a_t) when actions depend on latent state, so the foundational factorisation is not valid as stated.","rationale":"The reader's conditional verdict is appropriate: the paper is a conceptual synthesis whose formal claims are plausible but not fully worked out. My concern is narrower and more specific than the reader's stated weakest assumption. Causal sufficiency addresses unobserved confounders between action and outcome given r_t; the issue I raise persists even under causal sufficiency whenever the behaviour policy depends on the latent state r_t, which the paper's own action model P(a_t|r_t) explicitly allows. This makes Definition 3's factorisation a present mathematical gap rather than merely an omitted proof. I do not think this requires changing the verdict to rejection: the framework is easily repairable by either writing the transition with do() and treating P(r_t|x_t) as the encoder under an interventional action regime, or by explicitly including the inverse action model P(r_t|x_t,a_t). Because the paper's central contribution is a perspective and a component-wise identifiability programme, and the reader already marked it CONDITIONAL, I keep the verdict unchanged. A concrete derivation or simulation of the conditional would settle whether the factorisation is intended as observational or interventional, and would clarify the exact assumptions the CWM definition must carry.","tokens_in":8344,"tokens_out":13360,"duration_ms":144019,"concrete_test":"Re-derive the conditional from the full joint implied by Figure 2 and Remark 1: P(x_t)P(r_t|x_t)P(a_t|r_t)P(r_{t+1}|r_t,a_t)P(x_{t+1}|r_{t+1}). Marginalise over r_t and r_{t+1} and condition on a_t; the resulting expression for P(x_{t+1}|x_t,a_t) is ∫ P(x_{t+1}|r_{t+1})P(r_{t+1}|r_t,a_t)P(r_t|x_t,a_t) dr_t dr_{t+1}, with P(r_t|x_t,a_t) proportional to P(a_t|r_t)P(r_t|x_t). A direct numerical check: instantiate a linear-Gaussian CWM with a policy P(a_t|r_t) that depends on the latent state, generate samples, and compare the empirical conditional density P(x_{t+1}|x_t,a_t) with Definition 3's factorised expression; the difference will be nonzero except when P(a_t|r_t) is independent of r_t.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Definition 3 writes P(x_{t+1}|x_t,a_t) = ∫ P(x_{t+1}|r_{t+1})P(r_{t+1}|r_t,a_t)P(r_t|x_t) dr_t dr_{t+1}. This expression follows only if P(r_t|x_t,a_t)=P(r_t|x_t). The paper, however, introduces P(a_t|r_t) as the action model (Remark 1, item 2) and describes the CWM as an MDP over the formal state (x_t,r_t) (Remark 1, item 1). In such an MDP the policy may depend on the latent relational state; conditioning on a_t then creates a dependence between x_t and r_t through the path x_t→r_t→a_t, so the factor P(r_t|x_t) should be replaced by P(r_t|x_t,a_t) ∝ P(a_t|r_t)P(r_t|x_t). Equivalently, if the expression is intended as an interventional transition P(x_{t+1}|x_t,do(a_t=a)), it should be written with do, not as the observed conditional P(x_{t+1}|x_t,a_t). As stated, the definition is internally inconsistent: the action model P(a_t|r_t) and the factorised conditional cannot both hold in general. Since the component-wise identifiability discussion in Section 3 takes Definition 3 as its object, this undermines the formal ground of the paper's central claim. The conclusion's admission that rigorous formalisation is future work does not remove this present mathematical issue.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a unifying formalization of Causal World Models (CWMs) as structured decision models, defined as a tuple W=(X,A,{R_O}_O,P,U), where the observation-level transition is factorized through relational latent variables. It argues that identifiability guarantees for CWMs should be stated component-wise, each component up to an admissible equivalence that preserves the downstream reasoning task, and it connects this view to causal representation learning, causal discovery, and model-based decision-making. The paper's stated contributions are (1) a formal definition of a CWM as a Markov decision process linking observations, latent states, actions, transitions, and utility, and (2) a component-wise identifiability framework with a table of admissible equivalences and compatibility conditions. The paper also explicitly discusses causal sufficiency, the conditions under which action-conditioned transitions support interventional interpretation, and acknowledges open formalization challenges.","tokens_in":1625,"tokens_out":1664,"duration_ms":58310,"significance":"If the framework is made rigorous, it would provide a valuable common language for world-model research, connecting representation learning, causal discovery, and decision-making under a single component-wise identifiability lens. The paper is useful in synthesizing a broad literature and in being explicit about assumptions such as causal sufficiency and the difference between predictive conditionals and interventional transitions. It also gives credit where due: the conceptual decomposition is clear, the table of equivalences is a concrete starting point, and the authors honestly flag the lack of rigorous proofs and the unproven causal identifiability of their compositional construction. However, as it stands, the formal core contains a mathematical inconsistency in the main factorization, and the central composition claim is asserted without proof. These issues are load-bearing for the paper's main thesis, so substantial revision is needed.","major_comments":[{"comment":"The factorization P(x_{t+1}|x_t,a_t) = ∫ P(x_{t+1}|r_{t+1})P(r_{t+1}|r_t,a_t)P(r_t|x_t) is not the observed conditional under the paper's own action model P(a_t|r_t). With the generative structure x_t → r_t → a_t → r_{t+1} → x_{t+1}, the correct conditional is ∫ P(x_{t+1}|r_{t+1})P(r_{t+1}|r_t,a_t)P(r_t|x_t,a_t) dr_t dr_{t+1}, where P(r_t|x_t,a_t) is proportional to P(a_t|r_t)P(r_t|x_t). As written, the expression holds only if actions are independent of r_t given x_t (or randomized), which contradicts Remark 1 item 2 and the MDP reading of the CWM. If the intended quantity is interventional, it must be written as P(x_{t+1}|x_t, do(a_t=a)). This is load-bearing because Section 3's component-wise identifiability analysis targets the components of this factorization.","section":"Section 2, Definition 3"},{"comment":"The claim that compatible component-wise equivalences compose into a policy guarantee (aligned trajectory distributions, same expected cumulative utility, and corresponding optimal policies) is asserted without proof. The paragraph cites Li et al. 2006 for MDP abstractions, but it does not verify that the conditions in Table 1 imply the abstraction conditions used in that work, such as stochastic bisimulation. This composition claim is central to the paper's 'component-wise identifiability' thesis, so it needs at least a theorem statement with explicit assumptions, or it should be clearly labeled as a conjecture.","section":"Section 3"},{"comment":"Several rows of Table 1 are labeled 'compatibility conditions' and asserted to preserve the downstream reasoning task, but no derivation or proof of sufficiency (or necessity) is given. In particular, the 'local inverse consistency' condition for the prediction model and the equivariance condition for the transition model are introduced as bespoke interface conditions without justification. The paper itself concedes in the paragraph after Table 1 that the compositional construction of Kori et al. 2025 is used 'without establishing their causal identifiability.' Since Table 1 is the concrete content of Definition 5, this leaves the paper's main formal claim unsupported.","section":"Section 3, Table 1"}],"minor_comments":[{"comment":"The notation P(r_{t+1}|r_t) is introduced as a transition model, but it is not a component of Definition 3; clarify its relationship to P(r_{t+1}|r_t,a_t) and P(a_t|r_t).","section":"Section 2, Remark 1"},{"comment":"The maps for the state-abstraction pipeline are used without specifying their domains and codomains; in particular, the assembly map depends on the time-varying entity set O_t, but this dependence is not formalized.","section":"Equation (1)"},{"comment":"The symbol 'd=' for equality in distribution is nonstandard; write 'equal in distribution' at first use or use a properly typeset notation.","section":"Section 3"},{"comment":"The statement that 'future work should turn this perspective into a rigorous formalisation' is in tension with the abstract's claim of providing a formal definition; either add the missing proofs or soften the contribution claims.","section":"Section 4"},{"comment":"Abbreviations such as CWM, MEC, and the composition notation appear in the figure and table but are not defined in the captions; define them in the captions or at first use.","section":"Figure 1 and Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a perspective piece with a promising conceptual framework, but the formal gaps are currently load-bearing. The incorrect factorization in Definition 3 can be fixed by rewriting it as an interventional transition with do(), or by adding the missing posterior correction; the composition claim needs a proof or a clear conjecture. The heavy reliance on two self-citations for Table 1 is acceptable if those works are available, but the explicit admission that the compositional construction lacks causal identifiability should be addressed in the revision. I recommend major revision rather than rejection, as the conceptual direction is valuable and the issues are fixable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper is worth your time if you work on world models, causal representation learning, or model-based RL. It gives the field a clean formal object — the CWM tuple — and, more importantly, a component-wise account of identifiability that separates representation, transition, graph, and utility equivalence. Table 1 is a practical checklist, and the separation of predictive from causal claims is exactly what the area needs.\n\nThe genuinely new pieces are Definition 3 and the component-wise equivalence table. The authors are honest that the formalisation is not yet rigorous, and they say so in the conclusion. They also cite the relevant literature, including two of their own papers that are directly on point.\n\nThe soft spots are real, and one is more than soft. Definition 3 writes\n\nP(x_{t+1}|x_t,a_t) = ∫ P(x_{t+1}|r_{t+1}) P(r_{t+1}|r_t,a_t) P(r_t|x_t) dr_t dr_{t+1}.\n\nAs an observational conditional, this only follows if P(r_t|x_t,a_t)=P(r_t|x_t). But the paper explicitly introduces P(a_t|r_t) and describes a CWM as an MDP over (x_t,r_t), so actions can depend on the latent state. Then conditioning on a_t should change the posterior over r_t. Either the expression should be written with do(a_t), so the middle factor is an interventional transition and the r_t distribution is unchanged, or the first factor in the integrand should be P(r_t|x_t,a_t). As stated, the foundational factorisation is not valid in general. This is not a deep philosophical flaw; it is a concrete mathematical imprecision that can be repaired. But it sits in the definition on which the rest of the paper builds, so it should be fixed before the framework is taken as given.\n\nThe other soft spots are milder. The composition guarantee in Section 3 — that compatible per-component equivalences yield aligned policies and equal expected utility — is asserted without proof. The authors flag that they defer rigorous formalisation, so this is a known limitation, but it does mean the central claim is not yet a theorem. Table 1 also contains several arrows that are 'bespoke interface conditions' rather than proven results.\n\nThe causal sufficiency assumption is explicit and standard for this kind of work. I do not read it as a hidden weakness; it is a declared scope choice.\n\nOverall: the paper is a useful synthesis, honest about its limits, and the main flaw is fixable. I would send it to referees, with a request that they address the factorisation and at least sketch the composition argument.\n\nYours,","headline":"A genuinely useful synthesis of world models and causal identifiability, but Definition 3 has a real, fixable mathematical slip in its action-conditioned factorisation.","tokens_in":9155,"tokens_out":3973,"would_cite":true,"duration_ms":39279,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A causal world model should be a structured decision model over relational latent states, not a monolithic next-observation predictor.","keywords":["causal world models","causal representation learning","relational latent state","identifiability","causal discovery","Markov equivalence","model-based decision making","structural causal models"],"falsifier":"Fit the CWM factorisation to a tabletop environment in which an unrecorded common cause, such as a latent stickiness that affects both the block's contact state and the effect of a push, influences actions and outcomes. Compare the learned $P(r_{t+1}|r_t,a_t)$ on observational trajectories with the same quantity estimated under random assignment of actions. A systematic mismatch would show that the factorisation alone does not deliver causal transitions when causal sufficiency is violated.","tokens_in":8156,"feed_emoji":"🧠","tokens_out":9872,"duration_ms":88188,"temperature":0.7,"pith_summary":"This paper argues that a world model worthy of the name causal cannot be a single monolithic predictor of next observations. It proposes a formal definition, a Causal World Model, as a tuple of observation space, action space, relational latent states, transition distributions, and utility, with the action-conditioned observation transition factored as $P(x_{t+1}|r_{t+1})P(r_{t+1}|r_t,a_t)P(r_t|x_t)$. The central claim is that the identifiability of such a model should be stated component by component: the latent representation, the entity partition, the relational causal graph, the action-conditioned transition, and the utility each admit their own admissible equivalence, and the correct equivalence is whatever preserves the downstream reasoning or control task. This gives the scattered world-model literature a common formal language that connects representation learning, causal discovery, and decision-making. If the paper is right, debates about whether a model really has a world model become replaceable by precise claims about which components are recoverable and up to which transformation.","feed_headline":"Causal world models need more than next-frame prediction","feed_subtitle":"A formal definition splits world models into representation, causal graph, and utility, each with its own recoverability guarantee.","key_machinery":"The load-bearing object is the structured relational state $r_t$, assembled by an entity-inference map $\\phi$ from observations $x_t$ to entity-indexed latent variables $v_t$ and an assembly map $\\psi_O$ that stacks diagonal blocks, which carry entity attributes, and off-diagonal blocks, which encode inter-entity relations, into $r_t$. Around this state, the CWM factorisation organises the modelling problem into inference, transition or intervention, and prediction. The other central device is the component-wise identifiability table, which assigns each component an admissible equivalence, such as affine or permutation equivalence for representations, Markov equivalence for the causal graph, equivariance in distribution for transitions, and utility preservation for $U$, and states interface conditions, such as local inverse consistency of encoder and decoder, under which these per-component guarantees compose into a policy-level guarantee.","core_discovery":"On the paper's own terms, the discovery is that a Causal World Model is a decision model, not just a predictive model: $W=(X,A,\\{R_O\\}_O,P,U)$, with the one-step observation transition factorised through a structured relational latent state $r_t$. Each factor is a separate modelling commitment: $P(r_t|x_t)$ is inference from raw observations to entity-indexed latent variables, $P(r_{t+1}|r_t,a_t)$ is the action-conditioned transition whose causal reading requires consistency, positivity, and no unmeasured action-outcome confounding given $r_t$, and $P(x_{t+1}|r_{t+1})$ is prediction back to observations. The paper then claims that identifiability of this model is inherently component-wise: representation components are identifiable up to admissible equivalences such as invertible affine maps or entity permutations, the relational causal graph $G_r$ only up to Markov equivalence, and the full model is usable for control only when these per-component equivalences align through a common transformation $T$ of states, actions, and utilities. The argument is that exact identity of learned components is the wrong target; the right target is the weakest equivalence that preserves the intended prediction, intervention, or decision objective.","pith_inferences":["If the component-wise view is adopted, a natural testable protocol emerges for benchmarking world models: evaluate each component separately under its own equivalence class, then test the interface compatibility conditions, rather than scoring end-to-end prediction error.","The alignment transformation $T$ is treated as something that must be satisfied; an extension the paper leaves implicit is treating $T$ as an object to be learned, connecting CWM identifiability to representation-alignment and object-centric metrics.","The paper's causal-ladder framing suggests that tasks requiring counterfactual reasoning inherit stronger identifiability demands than tasks requiring only interventional control, because Markov-equivalent graphs can disagree on counterfactual quantities.","A partially observable or confounded extension would need to relax causal sufficiency; the natural test is whether the factorisation can be augmented with confounder nodes while preserving the component-wise equivalence structure."],"forward_implications":["A predictive latent-dynamics model that lacks an explicit relational state and utility is, by this definition, not a causal world model; adding causal support requires committing to entity-level variables and their interaction structure.","Identifiability claims for world models should be stated per component and per data regime; a single claim that the representation is identifiable is too coarse, and a single claim that the graph is identifiable is too strong when only observational data are available.","When causal sufficiency holds and actions are randomised or unconfounded given $r_t$, the learned action-conditioned transition supports interventional planning; otherwise it is only a predictive conditional and should not be used for do-calculus reasoning.","The component-wise equivalences compose into a control guarantee only when encoder, decoder, transition, action targets, and utility respect the same alignment; aligned models then have corresponding utility-maximising policies.","For unstructured observations such as pixels, the paper's position implies that the causal target is the latent relational graph $G_r$, not a causal graph over raw pixels."],"supporting_citations":[{"why":"Supplies the do-calculus and the semantics of the do-operator that the CWM needs to read action-conditioned transitions as controlled interventions.","marker":"[Pearl, 2009]"},{"why":"Source of the causal-sufficiency, Markov, and faithfulness assumptions that ground causal discovery over the relational graph.","marker":"[Spirtes et al., 2000]"},{"why":"Defines the Markov-equivalence class used as the admissible equivalence for the relational causal graph $G_r$.","marker":"[Verma and Pearl, 1990]"},{"why":"Provides the identifiability results that justify affine or permutation equivalence for the latent representation components.","marker":"[Kivva et al., 2022]"},{"why":"Supplies the entity or slot permutation equivalence used for the entity-block components of the state.","marker":"[Locatello et al., 2020]"},{"why":"Underpins the identifiable object-centric representation result used for entity-attribute blocks.","marker":"[Kori et al., 2024]"},{"why":"Motivates the compositional assembly of relational blocks that the structured state $r_t$ relies on.","marker":"[Kori et al., 2025]"},{"why":"Gives the conditions under which the action-conditioned transition is identified as a controlled effect rather than a predictive conditional.","marker":"[Hernan and Robins, 2020]"},{"why":"Supplies the policy-preserving state-abstraction result that the composition of per-component equivalences into a control guarantee is modelled on.","marker":"[Li et al., 2006]"},{"why":"Provides the causal influence-diagram formalism used to place decisions and utilities alongside the CWM variables.","marker":"[Everitt et al., 2021]"}],"fun_headline_variants":["Causal world models are decision models, not just predictors","Unifying causal world models: representation, graph, utility","World models get a formal definition for causal reasoning","From observations to structure: causal world models defined","Component-wise identifiability for causal world models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is causal sufficiency: every variable that jointly drives prediction and utility must be available to the model, either as a recorded observation component or as an inferred latent factor, so that no unobserved confounder distorts causal inference; if this fails, the action-conditioned transition cannot be read as an interventional effect unless actions are randomised.","fun_headline_variants_meta":{"raw":{"variants":["Causal world models are decision models, not just predictors","Unifying causal world models: representation, graph, utility","World models get a formal definition for causal reasoning","From observations to structure: causal world models defined","Component-wise identifiability for causal world models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000538,"raw_usage":{"total_tokens":2601,"prompt_tokens":981,"completion_tokens":1620,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":1545}},"tokens_in":597,"tokens_out":1620,"duration_ms":12586,"temperature":1.0,"reasoning_tokens":1545,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:36:01.178512+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fit the CWM factorisation to a tabletop environment in which an unrecorded common cause, such as a latent stickiness that affects both the block's contact state and the effect of a push, influences actions and outcomes. Compare the learned $P(r_{t+1}|r_t,a_t)$ on observational trajectories with the same quantity estimated under random assignment of actions. A systematic mismatch would show that the factorisation alone does not deliver causal transitions when causal sufficiency is violated.","supporting_citations":[],"review_version":1}