acceptodds
Under review as a conference paper at ICLR 2027

Position: Not Every Action Is an Intervention – Toward Causally Grounded World Models

Abstract

World models (WMs) have become increasingly capable of predicting and generating how environments evolve, yet predictive accuracy alone does not guarantee reliable reasoning about how the world responds when its variables or mechanisms are deliberately changed. We argue that addressing this gap requires treating causal interventions as a first-class consideration in world modeling. Importantly, actions and interventions should not be conflated: actions can have valid interventional semantics under appropriate assumptions, but the intervention space relevant to understanding a world may differ from and extend beyond the agent's action space. We propose viewing next-generation WMs as intervention engines: systems that represent where interventions enter a causal system, how their effects propagate through its mechanisms, and how informative interventions can reduce uncertainty about those mechanisms. This perspective shifts the emphasis from predicting trajectories under observed conditions toward reasoning about the consequences of deliberate changes. We develop a conceptual framework and identify open problems spanning causal representation learning, intervention-driven exploration, mechanism uncertainty, counterfactual reasoning, and interventional evaluation. Our goal is not to prescribe a particular causal architecture, but to argue for a broader formulation of world modeling centered on meaningful intervention spaces. Such a formulation provides a path toward WMs that combine predictive capability with more reliable reasoning under interventions, with implications for planning, embodied AI, and scientific discovery.

Primary area:
foundation or frontier models, including LLMs
Keywords:
World Model, Causal Intervention

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_9aWHOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper argues that world models should distinguish agent actions from causal interventions and should reason over intervention spaces that may extend beyond the agent’s ordinary action space. It proposes viewing future world models as “intervention engines” that represent intervention targets, causal mechanisms, propagation of intervention effects, and uncertainty over competing mechanisms. The paper discusses several open problems in causal representation learning, counterfactual reasoning, intervention-driven exploration, and evaluation, and sketches a possible intervention-centric world-model architecture based on a causal encoder, latent causal graph, intervention module, causal dynamics model, and decoder.

Critical strengths
  1. The paper highlights a valid distinction between agent actions and causal interventions, and correctly notes that action-conditioned prediction can have interventional semantics only under appropriate assumptions.

  2. The manuscript collects several relevant open problems at the intersection of world models and causal reasoning.

Critical weaknesses
  1. The central position is too broad and insufficiently novel.

The main claim is that world models should distinguish actions from interventions and reason over an intervention space that may be broader than the agent’s action space. However, this distinction largely follows directly from standard SCM semantics: an action is a decision variable entering through the agent’s control interface, while an intervention may modify any variable or mechanism in the causal system. I found it difficult to identify what new scientific claim is established beyond applying these existing causal concepts to world models. A position paper can be valuable without introducing a new algorithm, but in that case the position itself should offer a sufficiently sharp or non-obvious conceptual advance. I do not think this manuscript currently reaches that bar.

  1. The proposed “intervention engine” is not operationally defined.

The paper states that “a causal WM is an intervention engine,” but it does not provide clear necessary or sufficient criteria for what qualifies as such a model. It is unclear whether an intervention engine must explicitly represent a causal graph, support unseen do-operations, perform counterfactual inference, maintain uncertainty over mechanisms, identify informative interventions, or satisfy some combination of these capabilities. Without an operational definition, the term remains largely rhetorical. This also makes the taxonomy in Table 1 difficult to interpret consistently, because the paper itself later states that an explicit causal graph or symbolic do-operator is not required.

  1. The proposed formulation leaves nearly all of the difficult technical problems unresolved.

Section 4 presents a five-module architecture consisting of a causal encoder, latent causal graph, intervention module, causal dynamics model, and decoder. However, the paper does not provide a learning objective, identification assumptions, training procedure, or mechanism for learning the intervention space and the action-to-intervention mapping. The manuscript itself acknowledges that recovering causally meaningful latent variables is a difficult identification problem requiring assumptions, supervision, or interventional variation. Consequently, the “possible formulation” functions more as a conceptual block diagram than as a developed methodological framework.

  1. The paper provides no concrete validation that the proposed perspective changes what a world model can actually do.

Again, I do not view the absence of experiments as automatically disqualifying for a conceptual paper. However, because the proposed position is broad and the technical framework remains underspecified, there is also no empirical, theoretical, or even worked-out synthetic demonstration showing that distinguishing actions from a broader intervention space leads to a capability that existing world-model formulations cannot already provide. The result is that the paper remains almost entirely at the level of verbal argument and illustrative examples.

Questions for authors
  1. What is the single new scientific claim of this paper beyond the standard causal distinction between conditioning, actions, and interventions?

  2. What are the necessary and sufficient capabilities for a world model to qualify as an “intervention engine”?

  3. How are the causal variables, intervention targets, and action-to-intervention mappings in the proposed architecture actually identified from data, and under what assumptions is this possible?

  4. Can the authors provide a concrete setting in which an existing action-conditioned world model provably or empirically fails, while the proposed intervention-centric formulation resolves that failure?

  5. If explicit causal graphs, symbolic do-operators, and a particular causal architecture are all optional, what precisely distinguishes the proposed class of causal world models from sufficiently expressive existing world models evaluated under intervention-style perturbations?

3 comments

Thank you for the detailed feedback. We agree that the previous version did not state the central scientific claim sharply enough.

Q1. What is the new scientific claim beyond standard SCM distinctions?

We agree that the distinction between conditioning, actions, and interventions is established. We revised the paper to make the central claim more precise:

Action-space adequacy does not imply intervention-space adequacy. A WM may correctly predict every action in its ordinary action space A\mathcal{A} while remaining non-identified for interventions in a broader domain I∗\mathcal{I}^*.

This is not a claim that action-conditioned prediction is inherently non-causal. Under appropriate assumptions,

p(zt+1∣zt,at=a)=p(zt+1∣zt,do(at=a)).p(z_{t+1}\mid z_t,a_t=a)=p(z_{t+1}\mid z_t,\mathrm{do}(a_t=a)).

The issue is instead reliability over $\mathcal{A} $ versus reliability over a broader $\mathcal{I}^* $. We added both a formal construction and a controlled benchmark showing that distinct mechanisms can agree on all action-conditioned transitions yet disagree under interventions outside the action space.

Q2. What qualifies as an "intervention engine"?

We replaced the architecture-centered characterization with a behavioral definition. Given an intervention family $\mathcal{I}^* $, we call a WM intervention-capable if it can reliably predict the consequences of interventions I∈I∗I\in\mathcal{I}^* while preserving mechanisms not targeted by the intervention.

An explicit causal graph or symbolic do(⋅)\mathrm{do}(\cdot) operator is not required. Counterfactual reasoning, mechanism uncertainty, and active intervention selection are stronger capabilities rather than defining requirements.

Q3. How are causal variables and intervention mappings identified?

We clarified that they are generally not identifiable from passive or ordinary action-conditioned trajectories alone. Identification may require assumptions, supervision, multi-environment variation, known intervention targets, or interventional data.

We also revised Sec. 4 so that the five-module decomposition in Fig. 2 is presented as one possible implementation rather than a solved identification procedure. In particular, the causal encoder is now explicitly described as an open causal representation learning problem.

Q4. Can you give a concrete failure case?

Yes. We added a controlled benchmark designed to isolate the distinction between action-space prediction and intervention-space identification.

For

v_{t+1}=v_t+\Delta t\,\frac{g a_t-cv_t}{m},$$ the transition in a fixed environment depends only on $g/m$ and $c/m$. Thus, $(m,c,g)=(1,0.2,1)$ and $(2,0.4,2)$ induce exactly the same action-conditioned dynamics. We verified zero discrepancy over $10^6$ random samples and 200-step rollouts. However, applying the same intervention to one mechanism breaks this equivalence; Fig. 1 in the revised paper visualizes this directly. We also trained learned predictors over five seeds: | Condition | Action-only | Intervention-aware | Descriptor control | |---|---:|---:|---:| | Original environment | $5.8\times10^{-9}$ | $7.1\times10^{-7}$ | $1.3\times10^{-8}$ | | Held-out actions | $1.3\times10^{-8}$ | $1.3\times10^{-6}$ | $2.0\times10^{-8}$ | | $\mathrm{do}(m=2.0)$ | $4.4\times10^{-2}$ | $8.0\times10^{-7}$ | $6.0\times10^{-2}$ | | Combined intervention | $4.3\times10^{-2}$ | $1.6\times10^{-6}$ | $6.5\times10^{-2}$ | *50-step rollout MSE, mean over five seeds.* The action-only model generalizes to held-out action values but fails under mechanism interventions. Crucially, the descriptor control uses the same architecture and the same $(m,c,g)$ inputs as the intervention-aware model, but is trained only on action-equivalent environments with fixed $g/m=1$ and $c/m=0.2$. It fails similarly to the action-only model. Thus, the key factor is **identifying variation**, not simply providing additional inputs. We emphasize the limited scope: the intervention-aware model is given the post-intervention values explicitly, so the benchmark does not test intervention detection or online adaptation. It supports the narrower identifiability claim that perfect prediction in a fixed action-conditioned environment does not determine responses to independent mechanism changes. ### Q5. What distinguishes this from a sufficiently expressive existing WM? The distinction is behavioral, not architectural. An existing WM can satisfy our definition if it is reliable over the relevant intervention family. Our claim is only that success on ordinary action-conditioned trajectories does not establish that capability.

Thank you for the detailed rebuttal and additional clarifications. I appreciate the authors' efforts to refine the central position, clarify the definition of intervention capability, and provide supporting theoretical and experimental evidence.

However, my primary concern remains the novelty and significance of the central position. The distinction between action-space and intervention-space adequacy largely follows from established causal identifiability principles. While the additional evidence strengthens the argument, I remain unconvinced that the paper offers a sufficiently novel conceptual contribution or a concrete research perspective beyond existing causal reasoning frameworks.

Although I acknowledge the improvements, I believe these fundamental concerns remain unresolved. I therefore maintain my original rating of 1 (Clear Rejection).

Thank you for the follow-up. We agree with the reviewer on one important point: the causal identifiability principle underlying our argument is not itself new, and we do not intend to claim a new causal inference framework.

Our intended contribution is instead a position about how world models should be conceptualized and evaluated. In current WM research, strong predictive performance over an agent's action space is often treated as evidence of increasingly general world understanding. Our claim is that this inference is not justified: even when action-conditioned prediction has valid interventional semantics, correctness over the agent's action space does not establish correctness over a broader intervention domain involving changes to environmental variables or mechanisms.

The contribution is therefore not a new SCM formalism, but three things:

  1. A world-model-specific distinction: we separate action-space adequacy from intervention-space adequacy and argue that the latter should be stated explicitly as part of a WM's intended domain of reliability.

  2. An operational criterion: we define intervention capability behaviorally, relative to a specified intervention family I∗\mathcal{I}^*, rather than requiring a particular causal graph, architecture, or symbolic do(⋅)\mathrm{do}(\cdot) operator.

  3. A research and evaluation perspective: we show concretely that ordinary predictive and action-generalization tests can fail to reveal intervention-space non-identifiability, and argue that WM benchmarks should therefore test mechanism interventions, identifying variation, counterfactual correctness, and uncertainty over competing mechanisms.

The new benchmark is intended to make this distinction concrete rather than to establish a new causal theorem. It demonstrates that two mechanisms can be indistinguishable under ordinary action-conditioned prediction while diverging under interventions, and that simply adding intervention descriptors is insufficient without identifying variation.

We therefore agree that our position builds on established causal reasoning principles. Our claim is that these principles imply a specific and, in our view, underemphasized consequence for the design and evaluation of modern world models.

Reviewer_rmHuOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper is actually a position paper to address a problem that how a world model gives reliable reasoning when its variables or mechanisms are deliberately changed. It claims that causal interventions should be considered as a first-class topic in world modeling. To this end, it identifies open problems on causal representation learning, intervention-driven exploration, mechanism uncertainty, counterfactual reasoning and interventional evaluation. A possible formulation of intervention aware world model and implications on some downstream tasks are also discussed.

Critical strengths

The paper discusses an important topic about causal world modeling.

Critical weaknesses

1.In Section 3, four position statements are proposed, while the novelty of such statements is limited even though I agree with most of them. Almost all of concepts already have been discussed in causal representation learning and structural causal model (SCM). The paper seems merely inserting them in to world models. The paper does not give a new conceptual insight, which significantly reduces the novelty. 2.In Section 3, S1 and S2 proposes to represent the intervention explicitly and structure representations causally. If so, why not just expand conditioning action space that includes such constructed and explicit interventions instead of designing an intervention aware world model formulation.

Questions for authors

Other than above Critical Weaknesses, the architecture of Figure 1 in Section 4 is not that novel and could be misleading even though it claims the proposed architecture is a possible formulation. For example, the proposed encoder outputs a structured latent state in terms of an intervention space. It seems impractical in a pure vision scenario using the proposed encoder. Also, the perfect encoder and simplified intervention module could weaken the novelty.

2 comments

Thank you for highlighting the relationship between our proposal and existing SCM and causal representation learning literature.

1. Limited novelty relative to existing causal concepts

We agree that SCMs, interventions, causal representations, and counterfactual reasoning are established concepts. We revised the manuscript so that the novelty does not rest on these ingredients themselves.

The central claim is now:

Correctness over an agent's action space does not imply correctness over a broader intervention space.

We added both a formal construction and a controlled benchmark showing that two systems can produce identical transitions for every state and action in the original environment, yet disagree after interventions on environmental variables or mechanisms.

2. Why not simply expand the conditioning/action space?

We agree that this can be a valid implementation. Our distinction is semantic rather than syntactic.

An intervention may be represented as an additional conditioning variable. The key questions are whether that variable corresponds to a meaningful mechanism change and whether the training data contain sufficient variation to identify its consequences.

We therefore changed the emphasis from

p(zt+1∣zt,at)→p(zt+1∣zt,do(at))p(z_{t+1}\mid z_t,a_t)\rightarrow p(z_{t+1}\mid z_t,\mathrm{do}(a_t))

to

\rightarrow \text{reliability over }\mathcal{I}^*.$$ The new benchmark directly tests this point: | Condition | Action-only | Intervention-aware | Descriptor control | |---|---:|---:|---:| | Original environment | $5.8\times10^{-9}$ | $7.1\times10^{-7}$ | $1.3\times10^{-8}$ | | Held-out actions | $1.3\times10^{-8}$ | $1.3\times10^{-6}$ | $2.0\times10^{-8}$ | | $\mathrm{do}(m=2.0)$ | $4.4\times10^{-2}$ | $8.0\times10^{-7}$ | $6.0\times10^{-2}$ | | Combined intervention | $4.3\times10^{-2}$ | $1.6\times10^{-6}$ | $6.5\times10^{-2}$ | *50-step rollout MSE, mean over five seeds.* The descriptor control uses the same architecture and the same $(m,c,g)$ inputs as the intervention-aware model, but is trained only on action-equivalent environments $k\cdot(1,0.2,1)$. It still fails under interventions. For example, under $\mathrm{do}(m=2)$ its rollout MSE is $6.0\times10^{-2}$, compared with $8.0\times10^{-7}$ for the intervention-aware model trained with identifying variation. Thus, simply expanding the conditioning space does not resolve the identification problem without informative variation. The intervention-aware model is given the post-intervention descriptor explicitly; our claim is therefore about **identifiability from variation**, not intervention detection. ### 3. Concerns about Fig. 2 and the causal encoder We agree. We revised Sec. 4 to make clear that the five-module decomposition in Fig. 2 is neither a novel nor a necessary architecture. The causal encoder is now explicitly described as an unresolved causal representation learning problem. Recovering meaningful variables from high-dimensional observations may require additional assumptions, supervision, multi-environment variation, known intervention targets, or interventional data. The benchmark also does not support a claim that a factorized architecture is necessary. A simple intervention-conditioned predictor performs well once it is trained with identifying variation, so we use the experiment only to support the identification claim, not a particular architecture.

I appreciate authors' effort on the rebuttal. However, it doesn't fully address my concern about the novelty as listed by Reviewer 9aWH W1 and W3. I will keep my score.

Reviewer_aDscOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper argues that world models should be designed and evaluated around causal interventions, not just their ability to predict future states or generate plausible trajectories. The authors argue that predictive accuracy alone is insufficient for world models intended to support robust planning, embodied AI, and scientific discovery.

The authors propose a conceptual architecture with five components: a causal encoder, latent causal graph, intervention module, causal dynamics model, and a final observation-generation component. This is presented as one possible formulation, and not a fully implemented or validated system.

Critical strengths
  1. The paper identifies a limitation in current world modelling, namely that success at predicting observed trajectories does not necessarily imply reliable predictions under changes to the environment's causal mechanisms.

  2. The paper brings together causal representation learning, intervention-driven exploration, counterfactual reasoning, mechanism uncertainty, and evaluation. These are often treated as separate problems, and the paper makes a reasonable case for considering them as parts of a broader research agenda for world models.

  3. The call for benchmarks that separately measure predictive accuracy, interventional correctness, counterfactual accuracy, and uncertainty calibration is useful.

Critical weaknesses
  1. The novelty compared with causal modelling literature is not clear. Much of the paper's conceptual foundation comes from established causal inference: structural causal models, the distinction between conditioning and intervention, invariant mechanisms, counterfactual reasoning, and active experimental design. The five-module architecture also appears to be a fairly direct translation of these established concepts into a world-model setting.

  2. The claims about existing world models need to be grounded – they are currently very broad. The comparative table is useful but does not include sufficient criteria for assigning the categories. The claim that intervention spaces are almost universally implicit and that counterfactual reasoning is largely absent requires a more systematic account of what counts as sufficient evidence for these classifications.

  3. The evaluation proposal lacks a sufficiently concrete plan. The authors make a compelling case for interventional benchmarks, including tests of novel interventions, counterfactual correctness, and mechanism uncertainty. However, the discussion remains largely a list of desirable evaluation properties and open questions. The paper would be stronger if it developed at least one benchmark design in enough detail to illustrate how these principles could be evaluated.

  4. The paper does not provide empirical evidence for its central claim. There are no experiments establishing that current world models fail on the proposed intervention criteria, nor that the suggested formulation improves interventional or counterfactual reasoning. The claims about limitations of existing systems could be demonstrated through a small experiment, a systematic analysis of published results, or a worked benchmark example.

Questions for authors

Please address points raised in Weaknesses.

2 comments

Thank you for the constructive feedback. We revised the paper to address each of the four weaknesses directly.

1. Novelty relative to existing causal literature

We agree that the individual ingredients—SCMs, interventions, causal representations, counterfactuals, and experimental design—are established. We revised the paper so that the contribution does not rest on introducing these concepts themselves.

Our central claim is now:

Action-space adequacy does not imply intervention-space adequacy.

A WM may correctly model every action available to the agent while remaining non-identified for interventions on environmental variables or mechanisms outside that action space. We formalize this distinction and provide a constructive example showing that two mechanisms can induce identical action-conditioned dynamics while making different post-intervention predictions.

2. Claims about existing WMs were too broad

We agree and have narrowed the framing. We no longer argue that action-conditioned WMs are inherently non-causal, nor that causal adequacy requires an explicit graph or symbolic do(⋅)\mathrm{do}(\cdot) operator.

Under appropriate assumptions,

p(zt+1∣zt,at=a)=p(zt+1∣zt,do(at=a)),p(z_{t+1}\mid z_t,a_t=a)=p(z_{t+1}\mid z_t,\mathrm{do}(a_t=a)),

so ordinary actions can already have valid interventional semantics.

Our revised claim is instead behavioral: a WM should be evaluated relative to the intervention family $\mathcal{I}^* $ over which it is expected to remain reliable. Existing architectures may satisfy this criterion; predictive success over A\mathcal{A} alone simply does not establish reliability over a broader I∗\mathcal{I}^*.

3. The evaluation proposal was not concrete enough

We added a controlled benchmark that directly operationalizes the distinction. The dynamics are

v_{t+1}=v_t+\Delta t\,\frac{g a_t-cv_t}{m}.$$ In a fixed environment, the transition depends only on $g/m$ and $c/m$. Therefore, $(m,c,g)=(1,0.2,1)$ and $(2,0.4,2)$ are exactly action-equivalent. We verified zero discrepancy over $10^6$ random samples and 200-step rollouts. However, applying the same intervention to $m$, $c$, or $g$ breaks this equivalence. Fig. 1 in the revised paper visualizes this directly. This provides a concrete evaluation principle: test not only held-out actions or trajectories, but also whether predictions remain correct under specified interventions on variables or mechanisms. ### 4. Lack of empirical evidence We added learned experiments over five seeds: | Condition | Action-only | Intervention-aware | Descriptor control | |---|---:|---:|---:| | Original environment | $5.8\times10^{-9}$ | $7.1\times10^{-7}$ | $1.3\times10^{-8}$ | | Held-out actions | $1.3\times10^{-8}$ | $1.3\times10^{-6}$ | $2.0\times10^{-8}$ | | $\mathrm{do}(m=2.0)$ | $4.4\times10^{-2}$ | $8.0\times10^{-7}$ | $6.0\times10^{-2}$ | | Combined intervention | $4.3\times10^{-2}$ | $1.6\times10^{-6}$ | $6.5\times10^{-2}$ | *50-step rollout MSE, mean over five seeds.* The action-only model remains highly accurate in the original environment and on held-out action values, but its error increases by several orders of magnitude under mechanism interventions. The intervention-aware model remains accurate; across all tested interventions, its mean rollout MSE remained below $3\times10^{-5}$ across seeds. Importantly, the descriptor control uses the same architecture and receives the same $(m,c,g)$ inputs as the intervention-aware model, but is trained only on action-equivalent environments with fixed $g/m=1$ and $c/m=0.2$. It fails similarly to the action-only model. This shows that the improvement cannot be explained merely by adding intervention descriptors; identifying variation in the relevant mechanisms is required. We intentionally keep the empirical claim narrow. The intervention-aware model is given the post-intervention values $(m,c,g)$ explicitly, so the benchmark does not test intervention detection or online adaptation. It demonstrates that **near-perfect predictive performance in a fixed action-conditioned environment does not by itself identify responses to independent interventions on the underlying mechanisms**.

Thank you for the clarifications. I appreciate the authors' efforts to refine where the novelty lies, and address both the comments around claims and evaluations.

While these changes have strengthened the paper, having read the other reviews and discussion, I will not change my grade.

Comments on the paper

2 comments

Dear authors and reviewers,

The author-reviewer discussion period closes on November 18 (AoE). Please engage in response periods so that we have our issues discussed thoroughly.

AC

Thank you for overseeing the discussion. We appreciate the reviewers' feedback and have substantially revised the paper to address the main concerns around novelty, scope, operational definition, and empirical support.

The central claim is now stated more precisely as:

Action-space adequacy does not imply intervention-space adequacy.

We do not argue that action-conditioned world models are inherently non-causal. Under appropriate causal-sufficiency assumptions, ordinary actions can already have valid interventional semantics. Our claim is narrower: correctness over the agent's action space A\mathcal{A} does not in general establish correctness over a broader intervention domain I∗\mathcal{I}^* containing changes to environmental variables or mechanisms.

To make this claim concrete, we added both a formal construction and a controlled benchmark. We consider a one-dimensional controlled system in which xtx_t is position, vtv_t is velocity, ata_t is the agent action, mm is object mass, cc is a damping/friction coefficient, gg is actuator gain, and Δt\Delta t is the time step:

v_{t+1}=v_t+\Delta t\,\frac{g a_t-cv_t}{m}.$$ In a fixed environment, the transition depends only on the ratios $g/m$ and $c/m$. Thus, distinct mechanisms such as $(m,c,g)=(1,0.2,1)$ and $(2,0.4,2)$ are exactly action-equivalent, yet make different predictions after independent interventions on $m$, $c$, or $g$. We verified zero discrepancy over $10^6$ random samples and 200-step rollouts in the original environment. We also added learned experiments over five seeds: | Condition | Action-only | Intervention-aware | Descriptor control | |---|---:|---:|---:| | Original environment | $5.8\times10^{-9}$ | $7.1\times10^{-7}$ | $1.3\times10^{-8}$ | | Held-out actions | $1.3\times10^{-8}$ | $1.3\times10^{-6}$ | $2.0\times10^{-8}$ | | $\mathrm{do}(m=2.0)$ | $4.4\times10^{-2}$ | $8.0\times10^{-7}$ | $6.0\times10^{-2}$ | | Combined intervention | $4.3\times10^{-2}$ | $1.6\times10^{-6}$ | $6.5\times10^{-2}$ | *50-step rollout MSE, mean over five seeds.* The action-only model generalizes well to held-out action values but fails under mechanism interventions. Crucially, the descriptor control uses the same architecture and receives the same $(m,c,g)$ inputs as the intervention-aware model, but is trained only on action-equivalent environments with fixed $g/m$ and $c/m$. It fails similarly to the action-only model. This isolates the role of **identifying variation**, rather than merely adding more conditioning variables. We intentionally keep the empirical claim narrow: the intervention-aware model is given the post-intervention descriptor explicitly, so the benchmark tests intervention-space identifiability rather than intervention detection or online adaptation. We also revised the conceptual framing in response to the reviewers. "Intervention-capable" is now defined behaviorally: a WM is intervention-capable with respect to $\mathcal{I}^*$ if it reliably predicts the consequences of interventions in that domain while preserving mechanisms not targeted by the intervention. This definition does not require an explicit causal graph, factorized architecture, or symbolic $\mathrm{do}(\cdot)$ operator. The architecture in Fig. 2 is now presented only as one possible implementation. We clarified that identifying causal variables, intervention targets, and mechanism structure from high-dimensional observations is itself an open problem. Such identification may require assumptions, supervision, multi-environment variation, known intervention targets, or interventional data; the revised paper no longer treats the causal encoder or graph as a solved component. Finally, we do not claim a new causal inference framework. The novelty of the position lies in identifying a specific mismatch in current world-model practice: predictive or action-space success is often used as evidence of general world understanding, although causal identifiability only supports such conclusions within the intervention domain that has actually been identified. Our contribution is to make this distinction explicit for WMs, define intervention capability behaviorally, and derive concrete implications for benchmark design, training variation, and model evaluation.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.