PaperlayerView exact revision on arXiv

Source-linked AI summary

Do-JEPA: From Masking to Intervention in Latent World Models

Hossein Resani, Javen Qinfeng Shi

arXiv:2609.37378v1cs.CVcs.AIcs.LG

TL;DR

Latent world models predict logged next states, but their objectives do not separate action-caused changes from co-occurring ones. DO-JEPA compares paired futures under an action and a reference action, using the difference to supervise effects and intervention structure. Paired supervision improves pixel effect prediction over a same-data control and supports recovery of intervention structure in controlled settings.

  • Problem

    Logged-data prediction does not distinguish what an action caused from what merely co-occurred with it, so a model can credit a variable the action does not control without prediction error revealing the mistake.

    Source: §1 · p. 1

  • Method

    DO-JEPA compares latent futures from one saved simulator state under an action and a reference action, deriving effect, support, propagation, and invariance training signals from the paired rollouts.

    Source: §1 · pp. 1–2

  • Results

    Paired pixel supervision improves effect prediction over a same-data control, while onset supervision raises propagation edge AUROC from 0.624 without Ledge to 0.975 with Ledge.

    Source: §1 · p. 2, §4.1 · p. 6

  • Takeaways & Limitations

    A centered adapter lets paired supervision fine-tune a pretrained LeWM planner with success equivalent to the base planner within ±3 percentage points.

    Source: §5 · p. 9

    Exact pairs require saving and restoring a simulator’s full state, and each pair costs two rollouts; real trials lack shared noise and exact state, so their differences may be noisy and biased.

    Source: §5 · p. 9

Abstract

from arXiv · show

Latent world models are trained to predict what happens next, so nothing in their objective separates what an action caused from what merely co-occurred with it. Object-masking models such as C-JEPA intervene on what the predictor can see; we intervene on what physically happens. From one saved simulator state we run the dynamics under an action $a$ and under a reference action $a_{\varnothing}$, and train the model to predict the difference $\Delta z=z^{a}-z^{a_{\varnothing}}$ between the two latent futures. The resulting objective, Do-JEPA, has an effect loss, a support loss (where the action enters), a propagation loss (where its effect travels) and invariance losses (what must not change). In a synthetic system with object-aligned variables, support supervision finds the directly intervened object in 99.95% of test cases, where a sparse action mask sends the action to a nuisance slot in every case, and response-onset supervision recovers the ring-shaped propagation graph (edge AUROC 0.975 vs. 0.624). From pixels, the effect loss beats a control trained on exactly the same data: it lowers latent effect error by 28.4% on an end-to-end LeWM model and physical effect error by 13.5% when trained and tested on natural action sequences, and on three independently generated CausalWorld benchmarks it lowers responsive effect error by about 20% under physics shifts and the latent context sensitivity of predicted effects by 66%. Trained from scratch it costs factual accuracy; fine-tuning an existing model with it removes this cost. Together, these results show that intervening on the world, rather than on what the model sees, helps latent world models predict what their actions cause.

1 INTRODUCTION

Latent world models predict logged outcomes, but their objectives do not identify what an action caused. DO-JEPA compares futures from paired interventions on the world and adds signals for action effects, entry, propagation, and invariance.

1 INTRODUCTION

  • Logged next-state prediction can fit outcomes by crediting a variable the action does not control, and prediction error may not reveal that error.The paper gives a hidden cause shared by the behavior policy and rendered scene as an example.Source: §1 · p. 1
  • DO-JEPA contrasts latent futures from one saved simulator state under an action and a reference action, then supervises effect, support, propagation, and invariance.The paired runs share state, context, history, and noise.Source: §1 · pp. 1–2
  • 99.95–100% support detection on three seeds contrasts with sparse masks selecting the wrong slot at chance; onset supervision recovers propagation, and paired pixel supervision improves effect prediction over a same-data control.The support and propagation results are from controlled settings; the pixel comparison uses matched information.Source: §1 · p. 2

2 RELATED WORK

The paper situates DO-JEPA among latent-prediction world models and work on interventions and causal representations. Its contrast with C-JEPA is that masking changes observability, whereas DO-JEPA changes the rollout.

2 RELATED WORK

  • JEPAs predict in embedding space and have been used in action-conditioned video models and planners; LeWM provides the paper’s end-to-end pixel backbone.The paper also uses LeWM with its CEM planner unchanged for a planning study.Source: §2 · p. 2
  • Prior work uses interventions to identify latent causal variables, but this paper does not recover causal variables or graphs from pixels.The related work also covers modular mechanisms, interaction graphs, action selection for causal hypotheses, and intervention-centric evaluation.Source: §2 · p. 2
  • C-JEPA masks object-slot trajectories, while DO-JEPA’s comparison changes the world rollout rather than what the predictor observes.The paper re-implements C-JEPA on the same slots.Source: §2 · p. 2

3 METHOD

The method contrasts paired simulator rollouts to obtain action effects and response labels, then uses a slot predictor to model action entry and propagation. Its training signals are derived from rollouts and context twins rather than graph labels or object identities.

3.1 SETTING

  • A JEPA predicts the next observation’s latent, but its observational target does not identify which part of the observed change the action caused.The next-frame latent target is accompanied by a collapse-prevention term.Source: §3.1 · p. 3

3.2 PAIRED INTERVENTIONS AND WHAT THEY REVEAL

  • Paired trajectories replay one saved state under the action and reference action with shared context and noise, making the latent difference the effect of replacing the reference action in that state.The pair is exact on replay; shared latent components cancel, exactly for additive context contributions but not for context interactions with state.Source: §3.2 · p. 3
  • Immediate slot responses label direct support, later responses identify downstream responders, and response onset supplies propagation labels when events are temporally resolvable.A response one step after another slot makes the earlier slot a candidate cause; interactions unused in the rollout receive no positive label.Source: §3.2 · p. 3
  • A context twin changes context while holding state and action fixed, so differences between predicted effects under the two contexts are attributable to context.The paper defines nuisance, mechanism, and composed shifts as out-of-distribution test conditions.Source: §3.2 · p. 4

3.3 MODEL AND OBJECTIVE

  • The slot predictor combines a history-based action-free state prediction, a mask that routes the action into a slot, and gated message passing that propagates its residual.The gates depend on slot states and are not renormalized; the LeWM-class single-global-latent backbone has no mask or gates.Source: §3.3 · p. 4
  • The effect loss compares predicted and paired-rollout effects, while the training targets are derived from paired rollouts or context twins rather than simulator graphs or object identities.The figure summarizes the other losses as supervising where the action enters, how its effect travels, and what must not change.Source: §3.3 · p. 4, Figure 2 · p. 4

2. The support, propagation, invariance and context

The objective combines losses for the predicted effect, action support, propagation, invariance, and context consistency. Its variants separate paired-effect supervision from using the same rollout data as ordinary examples.

2. The support, propagation, invariance and context

  • Lsup matches the action-entry mask to immediate response energy, Ledge matches gates to onset labels, and Linv suppresses predicted effects and gates on non-responders.Lctx asks predicted effects to remain the same across context twins; the full objective also includes the paired effect loss.Source: §2 · p. 5
  • AUG uses factual and reference branches as ordinary examples without comparing them, PAIRED adds the effect loss, and full DO-JEPA adds context and slot losses where applicable.OBS uses only the factual branch; MASK re-implements C-JEPA on the same slots.Source: §2 · p. 5
  • At test time the model predicts an effect from an observation history and action, while training requires an exactly restorable simulator and different losses additionally require aligned slots or changeable context.PAIRED and AUG receive the same branches, making their comparison matched in information; planning uses only the factual prediction.Source: §2 · p. 5

4 EXPERIMENTS

Across these experiments, paired interventions identify direct support and propagation in controlled settings, while the effect loss improves effect prediction from pixels against matched controls. Its factual-accuracy trade-off depends on whether it fine-tunes an existing model or trains from scratch.

4.1 Q1–Q2: WHERE DOES THE ACTION ENTER, AND WHERE DOES ITS EFFECT TRAVEL?

  • 99.95% support accuracy contrasts with a sparse mask selecting a nuisance in every test case; the immediate paired response, not sparsity, identifies the direct target.Prediction and effect error do not reveal the support: MASK+L∆ has the lowest of both despite selecting the wrong entry slot.Source: §4.1 · p. 6
  • Edge AUROC rises from 0.624 without Ledge to 0.975 with it, recovering the four ring edges; coarser observations support reachability rather than one-hop edges.As propagation stages overlap within one observed step, one-hop edge F1 collapses while reachability remains recoverable.Source: §4.1 · p. 6

4.2 Q3: WHAT MUST NOT CHANGE?

  • With given variables, Lctx lowers OOD error by 43.9% and 55.2% on two seeds, while learned latents require care because generic isotropic penalties can cost accuracy.On learned latents, the penalty lowers latent context shift but does not lower decoded pixel shift on frozen slots; the takeaway is to measure invariance in physics-carrying directions.Source: §4.2 · p. 7

4.3 Q4: DOES THE EFFECT LOSS HELP FROM PIXELS?

  • On frozen VideoSAUR slots, DO-JEPA beats C-JEPA-style masking on all six metrics across three seeds, including 12% lower responsive effect error and 31% lower pixel context shift.The comparison is on C-JEPA’s own slots of pixel Push-T.Source: §4.3 · p. 7
  • 28.4% lower latent effect RMSE on LeWM’s training-distribution holdout accompanies 0.066 higher cosine, while the outcome-unfiltered holdout has 11.2% lower RMSE.These are latent metrics; on natural expert action sequences, the matched PAIRED control lowers responsive physical effect error by 13.5%.Source: §4.3 · p. 7, Figure 3 · p. 8
  • About 20% lower responsive effect error under mechanism and composed shifts accompanies a 66% lower latent context shift, but factual error is 51–53% higher on three CausalWorld benchmarks.Fine-tuning removes this factual cost; across the pixel experiments, the effect loss improves effect prediction against information-matched controls.Source: §4.3 · p. 8, Table 2 · p. 8
  • Fine-tuning an existing model keeps factual accuracy in CausalWorld or improves it on Push-T, whereas training from scratch trades factual accuracy and response AUROC for larger effect gains.Reducing the scratch effect weight to α = 0.25 cuts the factual cost from 52% to 10%.Source: §4.3 · p. 8

5 DISCUSSION AND LIMITATIONS

DO-JEPA is presented as an intervention-based objective for latent world models, with signals for effects, support, propagation, and invariance. Its scope and evaluation depend on simulator access and, for some signals, object-aligned variables and temporally separable responses.

5 DISCUSSION AND LIMITATIONS

  • The supervision captures intervention propagation in observed rollouts under the training intervention distribution and observation rate, without claiming to identify the underlying mechanism.The paper uses “do(·)” to name interventions, not do-calculus; paired effects are counterfactual only in the sense that simulator branches share state and noise.Source: §5 · p. 9
  • Exact paired rollouts require saving and restoring a simulator’s full state, cost two rollouts, and cannot be reproduced exactly in real trials, where differences may be noisy and biased.The intervention and its reference must also be known, although the direct target need not be.Source: §5 · p. 9
  • Support and onset labels require object-aligned slots, a known slot count, one direct target per action, and a frame rate that separates response onsets.Context twins also assume context can change without changing physics, while benchmark nuisances are rendered or synthetic.Source: §5 · p. 9

6 CONCLUSION

DO-JEPA contrasts same-state simulator rollouts to learn action effects and their support, propagation, and invariance. The conclusion also identifies context-sensitive invariance and factual accuracy as evaluation considerations, while the appendix provides benchmark, result, notation, and metric references.

6 CONCLUSION

  • Paired effects identify action entry and onset order identifies propagation in object-aligned synthetic systems, subject to the time resolution at which events separate.From pixels, paired effect loss improves prediction over a same-data control in two environments, and fine-tuning preserved factual accuracy in the reported experiments.Source: §6 · p. 9
  • Context twins provide nuisance invariance with given variables, but learned-latent invariance must be measured in directions that carry the physics.Prediction error revealed little about these properties, motivating interventional metrics as a routine part of world-model evaluation.Source: §6 · p. 9

CONTENTS OF THE APPENDIX

  • The appendix contents list benchmark descriptions, training and evaluation details, additional results, planning-study details, and boundary-condition diagnostics.The conclusion defines intervention as paired branches sharing state and noise; the terminology section distinguishes direct support, downstream propagation, response onset, invariance, and nuisance context.Source: CONTENTS OF THE APPENDIX · p. 12, A NOTATION AND TERMINOLOGY · p. 13

B METRICS GUIDE

The metrics guide defines how the paper measures prediction, effects, response detection, intervention support, propagation, and robustness. It also specifies metric direction, interpretation, and caveats that affect comparison.

B METRICS GUIDE

  • Effect error measures the predicted action-induced difference, while factual prediction error measures the predicted outcome of the factual branch.Effect MSE is the square of effect RMSE; responsive effect RMSE restricts physical error to responsive pairs and depends on the response threshold and probe accuracy.Source: B METRICS GUIDE · p. 15
  • Effect cosine measures direction while ignoring size, and response AUROC measures response detection rather than effect size or direction.Cosine can be noisy for near-zero effects, and response AUROC can be inflated by sampling shortcuts.Source: B METRICS GUIDE · pp. 15–16
  • Direct-target accuracy and F1 assess whether the action is routed to the right slot, while nuisance-mask leakage measures action mass on nuisance slots.Object-only accuracy can look like chance when mask mass goes to nuisances, and a clean mask does not guarantee clean predicted effects.Source: B METRICS GUIDE · p. 16
  • Nuisance effect measures predicted effects on variables the action cannot move, but its scale-dependent value must be read alongside effect error.It is zero for a model that predicts no effects at all.Source: B METRICS GUIDE · p. 16
  • Edge AUROC scores whether gates rank rollout-used edges above other slot pairs, whereas structural and event-level F1 compare a thresholded graph with different reference relations.Temporal resolvability measures whether onset order is visible; its results depend on the onset threshold, and equal onsets are unresolved.Source: B METRICS GUIDE · p. 16
  • Context shift measures changes in predicted effects across contexts that should not matter, while latent and physical versions can disagree.The true latent effect itself can depend on context.Source: B METRICS GUIDE · p. 17
  • The OOD/IID error ratio measures degradation under a shift, but can fall when IID error rises, so both errors should be reported.Relative reduction compares a lower-is-better metric with its baseline; rates use percentage-point differences, while cosine and AUROC use absolute differences.Source: B METRICS GUIDE · pp. 13, 17
  • Planning success measures goal completion, while rollout ratio measures how adaptation changes base-model dynamics; neither alone establishes causal structure.Planning success depends on the planner and its cost, and one-step rollout proxies can poorly reflect planning performance.Source: B METRICS GUIDE · pp. 15, 17

C BENCHMARKS

The benchmark suite spans synthetic structural and temporal systems, state- and vision-based Push-T, LeWM intervention corpora, and CausalWorld pushing. The settings vary in nuisance shifts, action distributions, observations, and what can be evaluated.

C BENCHMARKS

  • The benchmark suite includes synthetic SCMs, two Push-T settings, LeWM intervention corpora, and CausalWorld pushing, with settings and losses summarized in Table 8.Figure 4 depicts nuisances and shifts; in each benchmark a training nuisance-outcome correlation reverses or physics changes at test, while context twins share physics by construction.Source: C BENCHMARKS · p. 17, Table 8 · p. 17, Figure 4 · p. 18

C.1 SYNTHETIC SCMS

  • The hard SCM uses four object slots on a directed ring and three nuisance slots, with hidden-context action correlation reversed at test.Actions specify a noisy contact point and impulse rather than naming the target; propagation proceeds around the ring, while the easy SCM was retired because variants saturated at target accuracy 1.0.Source: §C.1 · pp. 17–18

C.2 STATE-BASED PUSH-T

  • State-based Push-T branches a saved simulator state under a factual action and no-op, with weak or strong nuisance confounding and a strong-confound multi-step corpus.The replay restores contact and warm-start state exactly; the strong confound also adds input-only sensor corruption.Source: §C.2 · p. 18

C.3 VISION PUSH-T

  • Vision Push-T uses RGB frames with a rendered border nuisance whose contact-and-action correlation reverses at test without changing physics.The frozen C-JEPA encoder’s slots decode physical state, and the nuisance moves slots more than one step of physics.Source: §C.3 · pp. 18–19

C.4 LEWM PUSH-T PAIRS AND PLANNING CORPORA

  • LeWM studies use five primitive controls grouped into a 10-D action and compare corpora that differ in responsiveness and action sampling.The enriched corpus is direction-biased and stratified rather than representative; outcome-unfiltered holdouts sample primitive actions independently from the expert marginal, removing temporal structure.Source: §C.4 · p. 19, Table 9 · p. 19

C.5 CAUSALWORLD PUSHING

  • CausalWorld varies visual shortcuts, physics, and context-effect correlation across OOD splits, but its single manipulated block cannot test direct-support recovery.Its mechanism split also makes context independent of outcome, so it is not a physics-only shift; the rebuilt benchmark assigns response labels from rollout outcomes after sampling geometry and action.Source: §C.5 · pp. 19–20

D TRAINING AND EVALUATION DETAILS

Across synthetic, state-based, and vision settings, the experiments test whether paired interventional supervision recovers action support and propagation while improving effect prediction under shifted context. Results vary by setting: context weighting improves state-based OOD error, while vision comparisons show both gains and trade-offs against AUG.

D TRAINING AND EVALUATION DETAILS

  • Training used AdamW, IID validation for checkpoint selection, and held-out evaluation after selection, with seed means and standard deviations or per-seed results reported.Settings varied by task; the synthetic SCM and temporal experiments used predictor and loss configurations distinct from the Push-T and LeWM runs.Source: D TRAINING AND EVALUATION DETAILS · pp. 20–21, Table 10 · p. 21

E.1 SYNTHETIC SCMS

  • In the hard SCM, the gated model achieved target F1 0.9995 ± 0.0003 and structural precision and recall of 1 across seeds 7, 17, and 27.Its mean ring-edge gate was 0.947–0.953, while the mean gate on other pairs was 0.0018–0.0053.Source: §E.1 · p. 22, Table 12 · p. 22

E.2 TEMPORAL RESOLVABILITY

  • The multi-timescale model reached one-hop F1 0.977 and schedule F1 0.955/0.977/0.909 at κ = 2/3/4, while a shared gate reached one-hop F1 ≈0.509.Tying heads improved cross-scale consistency by +0.083, +0.070, and +0.022 F1 over independent heads, with one-hop F1 0.977 versus 0.999.Source: §E.2 · p. 22, Table 13 · p. 23

E.3 STATE-BASED PUSH-T

  • With λctx = 0.25, state-based Push-T OOD error fell 43.9–59.0% across seeds, while IID error rose 3.5–7.0% and edge AUROC remained 0.994.Seed 7’s OOD split informed the λctx choice on the IID–OOD trade-off curve.Source: §E.3 · p. 23, Figure 5 · p. 23, Table 14 · p. 23

E.4 VISION PUSH-T

  • On vision Push-T’s reversed-pattern OOD split, DO-JEPA beat C-JEPA-style masking on every metric and seed, including 30% lower factual error and 12% lower responsive effect error.Response AUROC rose from 0.49 to 0.69, and predicted-effect context shift fell 18% in latent space and 31% in pixels.Source: §E.4 · p. 24, Table 15 · p. 24
  • Against AUG on vision Push-T, DO-JEPA had AUROC +0.025 and 20% less latent context sensitivity, but 9.3% higher factual error and 6.5% higher responsive effect error.In pixels, AUG’s predicted effects were less context-sensitive: 12.6 versus 13.7 px, with 0/3 seeds favoring DO-JEPA.Source: §E.4 · p. 24, Table 15 · p. 24

E.5 LEWM PUSH-T: PER-SEED RESULTS

On LeWM Push-T, the effect-loss comparison depends on which pairs and physical read-out are considered: gains are clearer for natural-action training than for the responsive subset of an outcome-unfiltered holdout. The results also distinguish latent metrics from post-hoc physical measurements.

Per-seed results

  • On the outcome-unfiltered holdout, responsive effect RMSE changed by −0.5% and cosine by −0.068, while the overall RMSE gain was −17.3% from non-responsive pairs.The holdout evaluator reproduced the original evaluation exactly.Source: §E.5 · p. 24, Table 17 · p. 24

Physical read-out

  • On training-distribution pairs, PAIRED lowered responsive physical effect error by 7.0% on average and by 8.2% over all pairs, but was 2.6–3.2% worse on responsive holdout pairs.On training-distribution pairs, responsive cosine changed from −0.19 to +0.23; the read-out was a post-hoc diagnostic using privileged labels for evaluation.Source: §E.5 · p. 25

Training on natural action sequences

  • 13.3–13.6% lower responsive physical effect error on the natural test set and 5.3–5.7% lower on the holdout followed natural-sequence training, while enriched-corpus training was 6% and 4–5% worse.Natural-trained models had 10–12 px responsive error versus 19.7 px for predicting no displacement, with cosine 0.80.Source: §E.5 · p. 25

E.6 CAUSALWORLD

CausalWorld results separate effect-loss benefits from factual-accuracy costs: training from scratch improves responsive effects but can miss the factual-error tolerance, whereas fine-tuning a trained AUG model avoids that cost. Context sensitivity and the effect-weight trade-off further qualify the results.

Training budget and confirmation

  • The earlier seed-7, 10k-update pass was single-seed and not fully blind to OOD results, while the fresh-seed confirmation used a frozen selector.OOD values for β = 0 and β = 1 were known in the earlier run.Source: §E.6 · pp. 25–26

Fresh-seed confirmation

  • Fresh-seed H1 results showed 11–24% lower responsive effect error and cosine +0.06, but factual cost of 15–18% exceeded the 10% tolerance, so H1 was not confirmed.Only seed 67 passed; the frozen selector changed none of the nine checkpoints.Source: §E.6 · p. 26, Table 20 · p. 27

Diagnosis

  • On IID data only, the full model’s effect norm fell relative to PAIRED by 23% late in the encoder, 43% after projection, and 48% at predictor output, where decoded physical error was 29% higher (3/3 seeds).The context gradient was 2.25 times larger, with 58% coming from the target side. Diagnostics used decoders fitted with privileged labels.Source: §E.6 · p. 26

IID results on three benchmark instances

  • On IID validation across three instances, responsive effect error changed from 8.76 to 7.96 mm and cosine from 0.482 to 0.537, while context shift fell from 2.38 to 0.79.Factual error changed from 15.07 to 22.84 mm and AUROC from 0.880 to 0.817 (AUG → PAIRED).Source: §E.6 · p. 26

Where the factual cost comes from

  • Fine-tuning converged AUG models with the effect loss kept factual error within −4.4 to +2.0%, left response AUROC unchanged, and improved responsive effect error, cosine, and latent context shift on every instance and split.This used 2,000 further updates from each instance’s converged AUG checkpoint.Source: §E.6 · pp. 26–27, Table 21 · p. 27

Effect-weight trade-off

  • From scratch with β = 0.25, responsive effect error fell 10.2/6.5/10.5% IID and 10–15% under physics shifts, while factual error rose 3.7/11.5/15.8% IID.Against PAIRED with α = 1, factual error was 25–30% lower on every instance and split; the effect weight therefore sets the trade-off when training from scratch.Source: §E.6 · p. 27, Table 20 · p. 27

F PLANNING STUDY DETAILS

The planning study tests whether paired supervision can be added to a pretrained planner without sacrificing planning performance. A centered adapter with rollout preservation retains the base success rate, but the reported components were introduced together, so centering alone is not established as sufficient.

Can paired supervision be added to a pretrained planner?

  • The centered-adapter approach kept planning success at 94%, while direct fine-tuning lowered effect RMSE by 29.1% relative to AUG but reduced planning success from 94% to 54%.The adapter’s three components were introduced together, so centering alone is not shown to suffice.Source: F PLANNING STUDY DETAILS · p. 27, Figure 7 · p. 28

Properties of the adapter

  • The adapter adds a shared action-effect feature difference to the frozen predictor, making its correction exactly zero for a no-op action.A 5-step rollout-preservation loss and a checkpoint rule capping rollout drift accompanied the adapter.Source: F PLANNING STUDY DETAILS · pp. 27–28, Figure 7 · p. 28

Rollout drift

  • The measured rollout drift rose from 0.28 to 0.42 over k = 1–5, but the estimated Lipschitz value was only a lower bound, not a certified constant.Because the rollout state carries over two history frames, L̂ < 1 does not establish that rollouts contract.Source: F PLANNING STUDY DETAILS · p. 29

Confirmation

  • In the fresh-seed confirmation, all relative protocol criteria passed, while the absolute ≥90% criterion was out of reach because the base model scored 83.0% on these 200 episodes.Source: F PLANNING STUDY DETAILS · p. 29

Training distribution

  • Adapters gained most on the distribution they were trained on: natural-trained gain was 2.47% on natural tests, while enriched-trained gain was 5.43% on enriched tests.The natural-trained gain was below the protocol’s 5% threshold but above zero.Source: F PLANNING STUDY DETAILS · pp. 29–30

Planning equivalence

  • At n = 600, natural-trained PAIRED planning success was 87.7% versus 87.8% for the base, and all six natural-trained and enriched-trained adapters were equivalent within ±3 pp.At n = 200, equivalence passed at ±5 pp but not ±3 pp.Source: F PLANNING STUDY DETAILS · pp. 27, 30

G BOUNDARY CONDITIONS AND DIAGNOSTIC FINDINGS

The section tests where the method’s assumptions hold and reports diagnostic findings, including boundaries imposed by benchmark data and the scope of pixel-slot propagation evidence.

Benchmarks outside the method’s assumptions

  • CLEVRER could not support named interventions because it lacks actions and paired videos, and its object-named counterfactuals could not be matched to anonymous slots.Soft IoU was at most 0.075.Source: G BOUNDARY CONDITIONS AND DIAGNOSTIC… · p. 30
  • In PHYRE, support-trained models located responding objects better than a matched control (support AUROC 0.810 vs. 0.583) but predicted their motion worse (responsive cosine −0.110 vs. +0.284).The results distinguish support identification from effect prediction.Source: G BOUNDARY CONDITIONS AND DIAGNOSTIC… · p. 30
  • In OGBench cube, exact MuJoCo restoration produced exact pairs, but undertrained preliminary models did not support a conclusion.The anonymous-slot diagnostic’s context term reduced context shift by 14–15%, short of its 20% target.Source: G BOUNDARY CONDITIONS AND DIAGNOSTIC… · p. 30

Propagation from anonymous pixel slots (single-seed IID diagnostic)

  • Edge AUROC rose from 0.823 without Ledge to 1.000 with it on anonymous pixel slots, while effect RMSE changed by only 2.8%, within the 5% margin.Direct-target top-1 was 1.0 on 583 evaluable cases; this single-seed result supports onset-order gates on IID visual slots, not a propagation claim beyond object-aligned variables.Source: G BOUNDARY CONDITIONS AND DIAGNOSTIC… · p. 30
  • The anonymous-slot propagation diagnostic used one seed and IID data, so the reported propagation claim remains limited to object-aligned variables.Onset supervision trained gates on anonymous visual slots that reproduced onset order.Source: G BOUNDARY CONDITIONS AND DIAGNOSTIC… · p. 30
Loading 2609.37378v1…