PAPER REPORTENAll readings ↗

On the Capability Separation Between World-Model Policy Learning and Imitated World-Action Models

English reading report: Method, equations, original figures, experiments and reproducibility.

Authors: Yang Yu

Affiliations: Nanjing University

Source: 2608.22197 ↗ · Catalog record

Reading: 56 / 558 · 3 original figures & tables · ~17 min ·

1. Paper overview

In one sentence: Future-action factorization can improve learning while preserving the ideal imitation target; stronger decisions require identified action effects and a deployment rule that compares their utility. e-identitye-paradigmse-flatteninge-equivalencee-regrete-formal-scope

At a glanceWhat to know
Research problem
Source description

Future-conditioned action prediction can benefit from temporal supervision without identifying what happens under a specified alternative action. The paper asks which differences arise from architecture, observational learning targets, or available information. Control capability is defined through complete trajectory distributions, so agreement on one task reward is insufficient to establish equivalence. e-frameworke-flatteninge-benefits

Core mechanism
Source description

The flattening theorem equates unrestricted external policy classes; the population theorem separately equates action marginals selected by ideal direct and world-action imitation. These results have different assumptions. e-flatteninge-equivalence

A key reported resultFormal value of one informative action intervention: Proved: every observational population learner has worst-case regret at least 1/4; exact direct and world-action imitation optima have 1/2; one informative intervention permits zero.

Worst-case regret across two environments; lower is better. Proved one-step family, one history, actions 0 and 1, utility U=Y. Demonstrations always choose action 0 and observe 1/2; action 1 yields 1 in M+ and 0 in M−. No dataset split applies.

The same utility and unrestricted two-action policy class are used. Only the interventional learner additionally observes action 1’s outcome. An information separation in a deterministic construction, not an empirical success rate or universal one-intervention learning guarantee. e-regret

Reading caution
Source description

Unrestricted-kernel equivalence does not equate fixed-size neural networks. Parameter, latency, memory and compute constraints can favor world-action factorization. There are no finite-sample bounds or optimization guarantees for practical networks. e-limitse-extensions

Core contributions

  • Source description

    The flattening theorem equates unrestricted external policy classes; the population theorem separately equates action marginals selected by ideal direct and world-action imitation. These results have different assumptions. e-flatteninge-equivalence

  • Source description

    An interventional KL metric isolates irreducible error when a predictor lacks the current candidate action. A separate result shows how the full future-action joint can encode an observational forward conditional. e-gape-recovery

  • Source description

    Counterexamples distinguish missing action support from hidden confounding. A further construction proves that one informative intervention can resolve a decision ambiguity left unresolved by an exact observational distribution. e-supporte-confoundinge-regret

Table 1. Three learning paradigms distinguished by what they fit and how they choose actions. Original paper, p. 11 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Read each row from the learned object to the decision principle. The direct policy estimates actions given history. The world-action row fits a joint: q_F predicts a future and q_I associates that future with an action. Its population target is an observational future-action distribution, so predicting a future does not by itself introduce utility optimization. The world-model row instead conditions an outcome or next-history kernel on a specified action and compares predicted consequences through a utility. The words “when identified” qualify its interventional target. Equations (12)–(13) on page 10 supply the world-action loss and deployment marginal; page 17 explains how explicit action conditioning can support a different decision rule. e-paradigmse-objectivese-scopee-equivalencee-identificatione-extensions

What it supports. The middle row changes the learned representation, but standard distribution-preserving deployment still recovers the behavior-action conditional at the exact population optimum. The bottom row requires a decision operation beyond that marginalization: specify candidate actions, evaluate their consequences and select through utility. An exact observational joint can support either use when the causal conditions hold.

Where the evidence stops. This is a conceptual comparison, not a neural architecture diagram or performance table. Equality assumes matched information and ideal observational training; fixed parameter or compute budgets can make the practical representational comparison different.

2. Motivation

2.1 The problem and the proposed response

Source description

Future-conditioned action prediction can benefit from temporal supervision without identifying what happens under a specified alternative action. The paper asks which differences arise from architecture, observational learning targets, or available information. Control capability is defined through complete trajectory distributions, so agreement on one task reward is insufficient to establish equivalence. e-frameworke-flatteninge-benefits

2.2 What this reading follows

Imagine a controller that predicts what a demonstrator will see next and then infers which action goes with that future. Does this already let it choose better actions than imitation? Yang Yu’s paper answers by separating three questions: what policies a representation can express, what distribution observational training recovers, and what information identifies the consequences of alternatives. The distinction is clearest in the paper’s small potential-outcome tables. Two environments can look identical to an observational learner yet demand different decisions. This reading follows the formal argument and its limits; the source supplies proofs and counterexamples, with no empirical robot benchmark. e-identitye-paradigmse-flatteninge-equivalencee-regrete-formal-scope

3. Research context

We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.

Catalog dimensionRecorded classification
Major categoryFoundational work
ArchitectureNot applicable
Prediction paradigmNot applicable
QuadrantNot applicable

This table preserves the labels recorded at reading time. The current major category is Related resources. View the current classification.

3.1 Evidence-based assessment

Supports the recorded classification

Reader analysis

Foundational work and Theory & planning / Classical world models & model-based RL fit this formal comparison. Architecture, prediction paradigm and quadrant are appropriately Not applicable: the work analyzes controller classes rather than introducing an implemented WAM. The future-plus-inverse factorization can describe a joint generated by one network, so factorization alone cannot establish a One Model architecture label. e-formal-scopee-paradigmse-objectivese-flattening

This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.

4. Problem formulation

4.1 Inputs and outputs

InputsOutputs
  • History H: context G, observations and past actions
  • Observational tuples (H, A, Y), with recorded future label Y
  • For model-based selection: candidate actions or policies, bounded utility and identified action effects
  • An executed stochastic action policy
  • For comparison: trajectory distributions, interventional prediction risk and regret

4.2 Equations and their role

LWA(qF,qI)=E(H,A,Y)pμ[logqF(YH)logqI(AH,Y)]\mathcal{L}_{\mathrm{WA}}(q_F,q_I)=\mathbb{E}_{(H,A,Y)\sim p_\mu}\left[-\log q_F(Y\mid H)-\log q_I(A\mid H,Y)\right]
Equation (12): p_mu is the demonstration distribution; H is history, A the action and Y the future. q_F predicts the future and q_I its associated action. Direct cloning instead minimizes expected negative log likelihood of its action policy (Eq. 9). e-objectives
πWA(ah)=qF(yh)qI(ah,y)dyπWA(ah)=pμ(ah)=πA(ah)\pi_{\mathrm{WA}}(a\mid h)=\int q_F(y\mid h)q_I(a\mid h,y)\,dy\quad\Longrightarrow\quad\pi_{\mathrm{WA}}^*(a\mid h)=p_\mu(a\mid h)=\pi_A^*(a\mid h)
Equations (13) and (28): integrate over possible futures; lowercase letters denote realized values and stars exact population optima. The implication requires the observational-training assumptions and holds almost everywhere under demonstration histories, not automatically at unsupported histories. e-objectivese-equivalencee-scope
TqWA(yh,a)=qF(yh)qI(ah,y)πˉq(ah)=pμ(yh,a)T_q^{\mathrm{WA}}(y\mid h,a)=\frac{q_F^*(y\mid h)q_I^*(a\mid h,y)}{\bar\pi_q(a\mid h)}=p_\mu(y\mid h,a)
Equations (49)–(52): bar-pi_q is the joint summed over futures and must be positive at the chosen action. T_q^WA is the extracted forward conditional. Equating it with T_a, the true interventional outcome kernel, needs the causal conditions in Eqs. (53)–(57). e-recoverye-identification
Eint(qg)=Iν,ρint(A;YH)+EHν[DKL(Tρ(H)g(H))]\mathcal{E}_{\mathrm{int}}(q_g)=I^{\mathrm{int}}_{\nu,\rho}(A;Y\mid H)+\mathbb{E}_{H\sim\nu}\left[D_{\mathrm{KL}}\left(T_\rho(\cdot\mid H)\Vert g(\cdot\mid H)\right)\right]
Equation (46): nu weights histories and rho candidate actions. g ignores the current action; q_g repeats g for all actions. T_rho mixes true action effects using rho. E_int is expected interventional KL error; I_int is conditional mutual information under this evaluation distribution, the unavoidable first term. e-gap

5. Method in detail

5.1 Marginalize the future before comparing controllers

Source description

The world-action policy has an internal random future, but the environment ultimately receives an action. Condition on the same history, multiply the future probability by the associated inverse-action probability, and integrate over futures. This gives an ordinary stochastic action kernel. Repeating that equality at each common history preserves the closed-loop trajectory law, which proves flattening without claiming that a small direct network implements the integral efficiently. Training is a second question. With realizable models and exact population log-loss optimization, the two factors become the observational conditionals. Their product integrates to the behavior-action conditional by the law of total probability. Thus a useful training-time future label can change how actions are learned without changing the ideal distribution selected by distribution-preserving imitation. e-objectivese-flatteninge-equivalencee-extensions

5.2 Reuse the joint by fixing an action

Source description

The limitation of the future predictor alone is not a limitation of every use of the joint. For a supported action, multiply the exact future and inverse factors, then divide by their action marginal. The result is the observational future distribution conditional on that action. This division is legitimate only where the denominator is positive. A causal interpretation requires more: consistency relates the observed outcome to its potential outcome, and conditional exchangeability rules out remaining action-outcome confounding given history. With those assumptions, the recovered conditional becomes an interventional model. Deployment still matters. Sampling the joint reproduces behavior; fixing each candidate action, evaluating its expected utility and taking an optimizer implements action comparison. The same stored representation can therefore support different decision rules. e-gape-recoverye-identification

5.3 Locate the information that changes the decision

Source description

Appendix A.8 sharpens non-identification into a decision consequence. The learner knows that action 0 always gives one half, but demonstrations never reveal whether action 1 gives one or zero. Those two environments demand different optimal actions. Any observational learner sees the same distribution and utility and consequently uses the same probability p of action 1 in both. Its two regrets are (1−p)/2 and p/2, leaving a worst-case lower bound of one quarter. Ideal imitation chooses p=0 and has a larger worst-case regret of one half. Intervening on action 1 reveals the environment because its outcome is deterministic in this construction. The final policy can then choose optimally. The proof gives an information advantage while keeping the external policy class fixed. e-regret

5.4 Training and inference

During training

Source description

Both imitation objectives minimize expected negative log likelihood under the same true demonstration distribution. Realizability and exact global population optimization remove estimation and optimization error from the main comparison. Future labels can reduce training-time action uncertainty, but predicting them at deployment introduces another error source. Practical efficiency benefits are discussed without measured experiments. e-objectivese-scopee-benefits

During inference

Source description

Distribution-preserving deployment samples a future and its associated action, or exactly marginalizes the future. At the ideal optimum, the resulting kernel is the observational behavior conditional. Matching the demonstrator’s entire trajectory law additionally requires that no omitted private variable controls its actions. e-objectivese-equivalencee-scope

Source description

Planning fixes candidate actions and compares their predicted utilities. Even an identified joint reproduces behavior under standard marginalization. MAP decoding or verifier reranking changes the action marginal and falls outside population imitation equivalence, while remaining flattenable as a policy. e-identificatione-deploymente-flattening

5.5 Implementation flow

  1. Fix the information and environment

    A finite-horizon partially observed process supplies the same history to both controllers. A future image, state, latent or video chunk is a demonstration label but has not occurred when the next action is selected. Observations following execution extend the history. e-framework

  2. Factor the observational joint

    Direct cloning models actions given history. World-action imitation models a future given history and an inverse action conditional given history and that future. Past actions may enter history; the future predictor does not receive the current candidate action. e-objectives

  3. Separate representation from identification

    For finite action and outcome spaces, normalize an exact joint at a supported action to recover the observational forward conditional. Consistency, conditional exchangeability and positivity are additionally required to interpret it as an action intervention. e-recoverye-identification

  4. Compare candidate consequences

    World-model control evaluates predicted one-step utility, or optimizes a policy through next-history model rollouts. This is a formal decision interface; the source supplies no particular neural planner, training schedule or frozen-module design. e-planninge-formal-scope

6. Experiments & results

This theoretical paper asks whether predicting a future before decoding an action enlarges control capability. It separates representable closed-loop behavior, the target selected by ideal imitation training, and information about alternative actions. A world-action controller can be marginalized into a direct stochastic policy; under explicit population assumptions, both imitation objectives recover the same behavior-action conditional. Action-conditioned control additionally needs identified consequences and a utility-based decision rule. The strongest separation is a proved information gap in a two-environment construction, not a robot benchmark or an equality claim for finite neural architectures.

Source and visual limitations
Reader analysis

The complete supplied paper contains one paradigm comparison table and two unnumbered tabulated counterexamples, with no original architecture figure, empirical quantitative benchmark table or experimental ablation. The edition therefore uses three original table crops: Table 1 supplies the mechanism comparison, Appendix A.7 supplies a formal diagnostic rather than an ablation experiment, and Appendix A.8 supplies a theoretical result construction rather than measured performance. No fourth suitable figure or table is present; equations and body prose have not been relabeled as experimental visuals. e-formal-scopee-paradigmse-supporte-regret

6.1 Read the original evidence

Appendix A.8, unnumbered environment table for Theorem 5.6. Identical demonstration outcomes conceal which action is optimal. Original paper, p. 25 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Use the paper’s utility U=Y, so larger entries are preferable. In the top row action 1 is optimal; in the bottom row action 0 is optimal. The behavior policy always selects action 0, revealing only the common value 1/2. An observational learner must therefore select the same action distribution in both environments. The proof names p as its probability of choosing action 1. Read the formulas below the table on the inspected source page: regret is (1−p)/2 in M+ and p/2 in M−. Finally, an intervention on action 1 reveals either 1 or 0, identifying the row and the correct subsequent decision. e-regrete-framework

What it supports. Every observational learner has worst-case regret at least 1/4 in this family. Exact direct cloning and world-action imitation choose action 0, giving worst-case regret 1/2. One informative intervention permits zero subsequent policy regret. The bound measures an inability to distinguish environments from observational information, not a limitation of the unrestricted policy class.

Where the evidence stops. This is a deterministic one-step existence construction with known utility and environment family. Zero regret describes the selected policy after information is obtained; it does not establish a cumulative exploration-cost guarantee or one-sample identification in noisy settings.

6.2 Results and evaluation conditions

Task & protocolReported resultComparison & interpretation
Formal value of one informative action intervention

Proved one-step family, one history, actions 0 and 1, utility U=Y. Demonstrations always choose action 0 and observe 1/2; action 1 yields 1 in M+ and 0 in M−. No dataset split applies.

Proved: every observational population learner has worst-case regret at least 1/4; exact direct and world-action imitation optima have 1/2; one informative intervention permits zero.

Worst-case regret across two environments; lower is better

The same utility and unrestricted two-action policy class are used. Only the interventional learner additionally observes action 1’s outcome.

An information separation in a deterministic construction, not an empirical success rate or universal one-intervention learning guarantee. e-regret

Population closed-loop equivalence

Matched data/history, realizability, exact optimization, distribution-preserving deployment and behavior sufficiency for trajectory equality.

Proved distance 0 between exact direct and world-action imitation policies.

Decision-complete control distance, equal to trajectory total variation

Both recover the behavior policy. With uniform component errors, the continuity bound is at most 1−(1−epsilon)^T, hence T epsilon.

T is the horizon; epsilon=min(1, epsilon_F+epsilon_I+epsilon_A) combines future, inverse-action and direct-policy errors. The bound must hold at every reachable history; it is not a sample-complexity result. e-flatteninge-equivalencee-sensitivity

Action-specific outcome prediction

Specified evaluation histories and candidate-action distribution; prediction is compared with true interventional outcome kernels.

Proved minimum for an action-unconditioned predictor: I_int(A;Y|H); exact action-conditioned prediction: 0.

Expected interventional KL divergence

The best action-unconditioned predictor is the candidate-action mixture T_rho.

The bound is positive when action effects differ on the evaluation support. It concerns the future predictor alone, not all information in the future-action joint. e-gape-recovery

6.3 Ablations and diagnostic examples

Read component removals and qualitative examples within their stated evaluation conditions.

Appendix A.7, unnumbered support-failure table. The unobserved action column contains an ambiguity that exact imitation cannot resolve. Original paper, p. 24 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Columns give the potential outcome Y(a) if action a is selected; rows give two possible environments. Start with Y(0): both entries are zero. The surrounding construction fixes the history and makes the behavior policy always choose action 0, so every recorded action-outcome pair is identical across rows. Now read Y(1). The first environment gives zero and the second gives one, although no observational example distinguishes them. The difference is a failure of action support: the missing column is precisely what a planner would need to predict before comparing that alternative. These entries are defined potential outcomes, not estimated means or trial success rates. e-supporte-confounding

What it supports. An exact fit to the observed joint cannot identify the unsupported action’s consequence. The example isolates information missing from the demonstrations, independently of model capacity or future-prediction accuracy on observed behavior. It supports the first part of Theorem 5.4 rather than an empirical claim about the size of a robot’s generalization error.

Where the evidence stops. This support-failure table is distinct from the hidden-confounding construction later on the same page. In that second example both actions are observed; omitted demonstrator information still prevents causal identification. Coverage and causal sufficiency address different failures.

7. Analysis & limitations

7.1 What the evidence leaves open

Source description

Unrestricted-kernel equivalence does not equate fixed-size neural networks. Parameter, latency, memory and compute constraints can favor world-action factorization. There are no finite-sample bounds or optimization guarantees for practical networks. e-limitse-extensions

Source description

Positive action coverage does not eliminate hidden confounding. Appendix A.7 uses a private Bernoulli variable controlling the demonstrated action; identical observed pairs remain compatible with different consequences of externally setting that action. e-confounding

Source description

Y must retain decision-relevant information. Identification on behavior support does not guarantee accuracy at histories visited by an optimized policy. Exact improvement over a suboptimal demonstrator assumes correct transitions throughout the candidate policies’ reachable support and exact optimization. e-limitse-extrapolatione-suboptimal

Reader analysis

No empirical ablation, robot execution, quantitative benchmark comparison or uncertainty estimate is reported. The tables and regret values describe mathematical constructions, leaving practical effect sizes untested. e-formal-scopee-regret

7.2 Questions for discussion

  1. Under matched data, history and compute, how much practical advantage comes from representation efficiency versus a changed deployment rule?
  2. What evidence establishes causal identification and adequate support for every action and history evaluated by a planner?

8. Reproducibility audit

8.1 Requirements and known gaps

Source description

Reproduction starts with probability identities and finite counterexamples: specify histories, action/outcome spaces, demonstration probabilities, potential outcomes, utility and deployment rule. Neural datasets, checkpoints, hardware budgets and hyperparameters are not provided as an implementation recipe. e-frameworke-objectivese-regrete-formal-scope

Reader analysis

Proposed check: enumerate Appendix A.8, sweep the probability of action 1, and compare worst-case regret before and after revealing that action’s outcome. Hold the utility and policy class fixed to isolate interventional information. e-regret

Reader analysis

Proposed check: hold Appendix B.1’s 0.6/0.4 joint fixed and compare marginalization, posterior sampling and MAP-future decoding. Compare complete action probabilities, not only means, to detect changes introduced by deployment. e-deployment

8.2 Proposed reproduction checks

The following checks are proposals motivated by the paper. They have not been run as part of this reading.

Check 1: Keep the joint fixed and change only deployment

Reader-proposed check, not run: use Appendix B.1’s exact joint, with masses 0.6 at (Y=0,A=0) and 0.4 at (Y=1,A=1). Obtain the two conditionals algebraically. Compare direct action marginalization, future sampling followed by inverse sampling, and MAP-future decoding without changing the learned probabilities. Marginalization must give action probabilities 0.6/0.4; repeated sampling should approach that distribution with sampling uncertainty reported. MAP must always choose action 0. Agreement of all three full distributions would signal that the deployment rules were implemented incorrectly or that the example was changed. e-objectivese-deployment

Check 2: Reveal one action outcome while holding the decision problem fixed

Reader-proposed check, not run: enumerate Appendix A.8’s two environments and compute regret for a sweep of p, the probability of action 1. Check the worst-case minimum of 1/4 and the imitation value 1/2 at p=0. Then compare two information controls: reveal another action-0 outcome, or reveal one action-1 outcome. The former cannot distinguish the environments; the latter must permit the correct policy in each and zero subsequent policy regret. Keep the utility, candidate policies and deterministic outcomes identical. Report intervention acquisition separately from the selected policy’s evaluation. e-regret

8.3 Reading coverage

Visual audit: Visually inspected the title/version page; formal framework, objectives, deployment, equivalence, prediction metric, identification, regret and limitations on pages 8–20; and the counterexamples and deployment/support remarks on pages 24–27. Reviewed Table 1 and both unnumbered potential-outcome tables as original PDF images, then inspected every final crop. Supporting assumptions and numerical derivations outside the crops are on these declared pages. All 29 pages were read in the six supplied text chunks; unlisted pages were not visually inspected. Separate supplements were not supplied or verified. No empirical architecture figure, benchmark or ablation was available in this theoretical source.

PDF pages inspected for this edition: 1, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 24, 25, 26, 27. Appendix coverage: reviewed.

Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.

Text reading scope & known omissions
  • Abstract (p. 1)
  • 1 Introduction (pp. 2–4)
  • 2 Related Work, Sections 2.1–2.3 (pp. 4–5)
  • 3 Framework and Scope, Sections 3.1–3.5.3 (pp. 5–11)
  • 4 Observational Equivalence of Direct and World-Action Policies, Sections 4.1–4.5 (pp. 11–14)
  • 5 Interventional Identification and Decision Separation, Sections 5.1–5.5 (pp. 14–19)
  • 6 Conclusion and limitations; Acknowledgments (pp. 19–20)
  • A Proofs, Sections A.1–A.8 (pp. 21–25)
  • B Additional Technical Remarks, Sections B.1–B.5 (pp. 25–27)
  • References [1]–[22] (pp. 27–29)

Outside the original text pass

  • Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
  • Separate supplemental material availability has not been fully verified.
  • Identity: the observed title exactly matches the catalog. The title page credits Yang Yu, Nanjing University, and identifies arXiv:2608.22197v1 [cs.LG], 23 August 2026. The catalog author string Yu, Yang uses inverted name order. Its submittedDate is 22 August 2026, one day earlier than the PDF stamp; this date discrepancy remains unresolved. Only the supplied v1 artifact was reviewed; no other edition was compared.
  • All six supplied text chunks were read completely. Text extraction does not reconstruct original visuals; the retained PDF was therefore visually inspected on pages 1, 8–20, and 24–27, including every final table crop. Other pages were read as extracted text, not visually inspected.
  • Separate supplemental material availability has not been fully verified; none was supplied.
  • Code was not inspected and no experiments or computational reproductions were run.

The visual audit above records the subsequent illustrated pass.

8.4 Traceable evidence

e-identityPDF p. 1, title block and arXiv margin stampInspect

The exact title is On the Capability Separation Between World-Model Policy Learning and Imitated World-Action Models. The author is Yang Yu, Nanjing University. The stamp reads arXiv:2608.22197v1 [cs.LG], 23 Aug 2026.

Go to primary source ↓
e-frameworkPDF pp. 2–3, Introduction; pp. 6–8, Sections 3.1–3.3, Eqs. (1)–(7) and Definition 3.4Inspect

Histories contain context, observations and past actions in a finite-horizon partially observed process. Future labels are recorded trajectory functions. Controller classes encode behaviors; learner classes distinguish observational from additional interventional information.

Go to primary source ↓
e-scopePDF pp. 8–9, Assumptions 3.5–3.6; p. 12, Assumption 4.5Inspect

The exact comparison assumes common history/data, realizability, global population optimization and distribution-preserving deployment. Behavior sufficiency is needed when identifying the learned marginal with the demonstrator’s trajectory-generating policy.

Go to primary source ↓
e-objectivesPDF pp. 9–10, Sections 3.5.1–3.5.2, Eqs. (9)–(15); pp. 21–22, A.3Inspect

Direct cloning minimizes action log loss. World-action imitation fits q_F(Y|H)q_I(A|H,Y) using two log losses and marginalizes Y at deployment. One network can induce this joint. KL decompositions establish the population targets.

Go to primary source ↓
e-planningPDF pp. 10–11, Section 3.5.3, Eqs. (16)–(20)Inspect

An action-conditioned outcome kernel supports expected one-step utility. A next-history model supports rollouts and maximization of expected bounded trajectory utility over candidate policies.

Go to primary source ↓
e-paradigmsPDF p. 11, Table 1, all three rows and four columnsInspect

The table contrasts learned object, population target and decision principle for direct cloning, world-action imitation and world-model policy learning. The last targets an interventional kernel when identified and compares actions through utility.

Go to primary source ↓
e-flatteningPDF pp. 11–12, Definition 4.2, Proposition 4.3 and Theorem 4.4, Eqs. (22)–(27); p. 21, A.1–A.2Inspect

Supremum utility difference equals trajectory total variation. Marginalizing a world-action controller yields an identical external action kernel. A point-mass future and arbitrary inverse kernel represent any direct policy in unrestricted classes.

Go to primary source ↓
e-equivalencePDF p. 13, Theorem 4.6, Eqs. (28)–(30), and Remark 4.7; pp. 21–22, A.3Inspect

Exact observational optima recover p_mu(A|H) almost everywhere. Behavior sufficiency extends equality to full trajectories. Omitted demonstrator information can invalidate equality with the original demonstration trajectory law.

Go to primary source ↓
e-sensitivityPDF p. 13, Section 4.4, Eqs. (31)–(35); p. 22, A.4Inspect

Uniform future, inverse and direct-policy TV errors sum to a capped action error epsilon. If the bound holds at every reachable history, trajectory TV is at most 1−(1−epsilon)^T and T epsilon. This is sensitivity, not a statistical rate.

Go to primary source ↓
e-benefitsPDF p. 14, Section 4.5, Eqs. (36)–(37)Inspect

A realized future can lower training-time action uncertainty. The source discusses video pretraining, temporal structure and representation benefits while emphasizing deployment future-prediction error.

Go to primary source ↓
e-inversePDF pp. 14–15, Section 5.1, Eqs. (38)–(41)Inspect

When action effects are identified, future prediction learns a behavior-weighted mixture. The inverse conditional depends on both behavior prior and action effect; averaging the posterior recovers the behavior policy.

Go to primary source ↓
e-gapPDF p. 15, Section 5.2, Theorem 5.1, Eqs. (42)–(47); p. 23, A.5Inspect

Interventional KL error for an action-unconditioned predictor decomposes into conditional mutual information plus KL to the candidate-action mixture. Its minimum is that mutual information; an exact action-conditioned kernel has zero error.

Go to primary source ↓
e-recoveryPDF p. 16, Proposition 5.2, Eqs. (48)–(52)Inspect

An exact joint normalized by its positive action marginal recovers p_mu(Y|H,A). This algebraic result is observational, not by itself causal.

Go to primary source ↓
e-identificationPDF pp. 16–17, Section 5.3, consistency condition and Eqs. (53)–(61); p. 23, A.6Inspect

Consistency, conditional exchangeability and positivity identify the recovered conditional as T_a. Standard deployment still marginalizes to behavior; action clamping and utility maximization support optimal one-step comparison.

Go to primary source ↓
e-supportPDF p. 18, Theorem 5.4(i); p. 24, A.7 Support failure, unnumbered table and following equationsInspect

With behavior always selecting action 0, M1 and M2 both yield Y(0)=0, but Y(1) is respectively 0 and 1. Observational distributions coincide although action-1 interventions differ.

Go to primary source ↓
e-confoundingPDF p. 18, Theorem 5.4(ii); p. 24, A.7 Hidden confounding under positive action support; p. 26, B.2Inspect

Let private Z be Bernoulli(1/2) and A=Z. Models Y=A and Y=Z have identical observations; under do(A=1), outcome-1 probabilities are 1 and 1/2. Deploying the action marginal loses dependence on Z.

Go to primary source ↓
e-suboptimalPDF p. 18, Corollary 5.5, Eqs. (62)–(64)Inspect

With behavior sufficiency, a suboptimal demonstrator, exact transitions on every candidate policy’s reachable support, and exact optimization, world-model control improves value by the demonstrator’s optimality gap Delta.

Go to primary source ↓
e-regretPDF p. 19, Theorem 5.6(i)–(iv); p. 25, A.8, unnumbered potential-outcome table and regret derivationInspect

M+ has outcomes (1/2,1), M− has (1/2,0); behavior always selects action 0 and utility is Y. Selecting action 1 with probability p gives regrets (1−p)/2 and p/2. Their maximum is at least 1/4; exact imitators incur 1/2 worst-case regret. One action-1 intervention identifies the environment and permits zero subsequent policy regret.

Go to primary source ↓
e-deploymentPDF p. 13, paragraph after Theorem 4.6; pp. 25–26, B.1 Point decodingInspect

Joint masses are 0.6 at (Y=0,A=0) and 0.4 at (Y=1,A=1). Distributional deployment preserves action probabilities 0.6/0.4; MAP-future decoding always chooses action 0. Conditional-mean decoding in continuous spaces only preserves the mean in general.

Go to primary source ↓
e-limitsPDF p. 20, Section 6, seven limitation paragraphsInspect

The source limits its claims to class-level, population analysis with matched deployment information and distribution preservation, stressing behavior sufficiency, causal identification and a decision-relevant future variable.

Go to primary source ↓
e-extensionsPDF pp. 26–27, B.3 Action chunks and language conditioning; B.4 Restricted parametric classesInspect

Marginalization extends to aligned action/future chunks and common language context. A restricted direct architecture reproduces a factorized policy class only if closed under the required marginalization.

Go to primary source ↓
e-extrapolationPDF p. 27, B.5 Observational recovery and model exploitationInspect

A planner can leave behavior support even when observational identification is valid there. Support-level identification differs from model accuracy under the optimized policy.

Go to primary source ↓
e-formal-scopePDF pp. 3–4, Main results and Organization; pp. 11–19, Sections 4–5; p. 20, population-analysis limitation; pp. 21–27, Appendices A–BInspect

The work presents proofs, a paradigm comparison table and finite counterexamples. It reports no trained neural implementation, empirical benchmark table, experimental ablation, dataset split or measured compute budget.

Go to primary source ↓

8.5 Primary sources

Scroll across the image to inspect details. Press Esc to close.