acceptodds
Withdrawn from ICLR 2027

TIDE: The Reachability–Learnability Gap in Temporal Intent Control

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Visual world models support goal-directed control through learned latent representations used for direct execution or decision-time search, yet search may succeed when direct execution fails. This discrepancy raises three questions: What behaviors can a frozen system execute? Which can finite-budget search discover? And which can a shared model recover without online search? Focusing on goal-conditioned visual control with frozen, action-conditioned jointembedding predictive world models, we introduce TIDE, a controlled empirical framework that equips each frozen representation with a standardized intentconditioned execution head and varies only the temporal structure of bounded intent corrections. Specifically, we compare SINGLE one-shot corrections, SHARED corrections reused across execution calls, and DYNAMIC call-specific corrections to study how temporal intervention structure affects reachability, searchability, and learnability. In a matched intervention evaluation based on Push-T, DYNAMIC call-specific corrections uncover more validated successful behaviors. Across four control tasks (Push-T, Cube, Reacher, and TwoRoom) and two frozen backbones (LeWM and Fast-LeWM), persistent SHARED corrections are more consistently effective under matched finite search budgets. Validated SHARED interventions can also be partially amortized into search-free control, although the gains remain modest and offline learning signals alone do not reliably predict closed-loop improvement. These findings expose a control-interface gap in this class of frozen visual world models: predictive latent representations do not by themselves ensure that reachable behavior is readily searchable or learnable.

Primary area:
applications to robotics, autonomy, planning
Keywords:
World Models, JEPA

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_BQE2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents TIDE, an empirical framework for studying reachability, finite-budget searchability, and search-free amortization in frozen visual world-model control. It compares bounded intent corrections applied once (Single), shared across execution calls (Shared), or varied across calls (Dynamic). A separate Push-T reachability study finds that Dynamic uncovers more validated successful behaviors, while experiments across four tasks and two backbones generally favor Shared under matched finite search budgets. The authors also train proposers from validated Shared interventions, obtaining modest search-free improvements. The results highlight that the existence of successful interventions, their discovery through search, and their generalization through learning are distinct properties.

Critical strengths
  1. Clear conceptual framing. The paper usefully distinguishes reachability, finite-budget searchability, and search-free amortization.
  2. Controlled empirical comparisons. Frozen components, matched search budgets, and temporal ablations support comparisons across multiple tasks and backbones.
Critical weaknesses
  1. The reachability and searchability findings are not fully aligned. Dynamic’s greater validated coverage comes from a separate, reference-search-conditioned INTACT Push-T cohort using Adam inversion, whereas the main searchability study uses standardized LeWM/Fast-LeWM heads and CEM. Table B provides an aligned analysis, but reports 16 Shared versus 14 Dynamic high-budget successes without cross-structure overlap. Thus, greater coverage alongside weaker finite-budget searchability remains unverified in the same standardized system.
  2. Learnability is evaluated only for Shared. Although the abstract links temporal structure to reachability, searchability, and learnability, the amortization experiments train only Shared proposers. They therefore leave temporal structure’s effect on learnability untested, including whether behaviors found by Dynamic but missed by Shared search can be amortized. A controlled Shared–Dynamic learning comparison would address this gap; otherwise, the stated research objective should be narrowed.
  3. Limited qualitative visualization. The paper provides aggregate results but no task snapshots or rollout comparisons showing how intent corrections change actual behavior. Matched visualizations of Direct, Single, Shared, and Dynamic executions, including representative rescues and failures, would help readers interpret the interventions and assess their behavioral effects.
Questions for authors
  1. Could the authors clarify why these particular papers are cited for Push-T, Cube, Reacher, and TwoRoom, respectively? Do they refer to the original task definitions, dataset sources, or the specific environment implementations used?
  2. See weakness
Reviewer_t1WSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

For a frozen visual world model with a fixed search budget, this work proposes a framework for evaluating the relationship between reachability, searchability, and learnability. The work introduces TIDE, a way to compare with an intent-conditioned executor that is modified in three different ways using bounded corrections (Single, Shared, Dynamic). Finally, the method is evaluated on four tasks (Push-T, Cube, Reacher, and TwoRoom) and two backbones.

Critical strengths
  1. Freezing the world model and executor makes for a robust and controlled evaluation setting for studying temporal intervention.
Critical weaknesses
  1. Gains from learnability are too small and limited to be able to draw decisive conclusions.
  2. Other factors, like the choice of algorithm, also influence the applicability for control.
Questions for authors
  1. If provided with a larger search budget or a better optimization algorithm, does Shared still have an advantage over Dynamic?
Reviewer_GvZMOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper asks why a search-free policy built on a frozen action-conditioned JEPA world model fails on states where search succeeds. TIDE attaches a standardized intent-conditioned execution head to frozen LeWM and Fast-LeWM backbones and perturbs only the intent m=zg−zm = z_g - z, with a correction bounded to 25% of its RMS. SINGLE corrects only the first of five head calls, SHARED reuses one normalized correction for every call, and DYNAMIC uses an independent correction per call. The paper separates three properties: reachability (some admissible correction succeeds), searchability (a specified finite-budget optimizer finds one), and learnability (a shared proposer recovers one without search). On a separate 136-state INTACT Push-T cohort conditioned on prior search success, gradient-based inversion validates 25, 36, and 58 successes for Single, Shared, and Dynamic. Across four tasks and two backbones, CEM with 256 evaluations rescues more with Shared than with Single in 7 of 8 cells, and more with Shared than with Dynamic in 6 of 8. A Shared proposer gains +0.90 pp in the four eligible cells and +0.84 pp on held-out units.

Critical strengths
  1. The evaluation protocol is careful and transparent. Structures are compared on the same frozen systems, cohorts, and budgets. Outcomes come from real closed-loop validation with causal-twin identity checks, and fixed denominators bound the effect of unresolved units. Preregistered eligibility rules are applied even when they exclude cells. A one-time held-out confirmation on 448 untouched units per task reproduces the main search pattern. This makes the observations credible within their scope.

  2. The Single–Shared comparison is a clean controlled contrast. Both expose dmd_m search variables and differ only in persistence. Shared rescues more in seven of eight cells, often by a wide margin (e.g., +15 on Push-T/Fast-LeWM). The Reacher/LeWM cell also illustrates a large performance gap between the tested intent interface and raw-action search: every intent structure rescues 0/64, while raw-action CEM rescues 51–52/64.

Critical weaknesses
  1. The diagnostic novelty and interpretation remain unclear. At each call, mj⋆=(zg+Δmj)−zjm^\star_j=(z_g+\Delta m_j)-z_j is equivalent to shifting the head's goal-conditioning input, motivating comparison with latent subgoal planning (LEAP). The shifted latent need not represent a realizable state, and the planning objective retains the original goal. Connections to frozen-policy steering (Parrot/DSRL) and search distillation (POPLIN/Expert Iteration) also need development. Dynamic's larger validated success count comes from the separate INTACT study; the aligned Table B does not establish a strict reachability gain for Dynamic. Shared and Dynamic search 192 and 960 variables under the same small CEM budget, leaving dimensionality and optimizer behavior as unresolved explanations for their ordering.

  2. The learnability study does not establish what limits amortization. Only Shared is evaluated in the standardized amortization matrix, gaining +0.90 pp from teachers collected on 750 train-side units per cell. No teacher-count learning curve or same-setting comparison of alternative learners separates limited teacher coverage from proposer limitations. Table C's additional learning probes use the separate INTACT setup. Thus, the results demonstrate modest amortization under one protocol, but offer limited evidence that the remaining gap persists across learners, data scales, or execution interfaces.

Questions for authors
  1. Does Dynamic expose additional validated reachable states on the standardized systems? Please clarify the inversion objective in A.7 and apply a comparable procedure on matched standardized-system cohorts, extending the Push-T/LeWM comparison in Tab. B. Please report class lower bounds from the union of admissible validated witnesses. For Tab. 3, are the four states rescued by Shared but missed by Dynamic included among the five Single-found states missed by Dynamic? Reporting this overlap would clarify whether the stated lower bound of 63 can be tightened.

  2. Is the searchability ordering due to temporal structure or to search dimension? Please report the achieved predicted costs. Please also test one or more of the following: budgets scaled with dimension, gradient-based search, or Dynamic restricted to a dmd_m-dimensional subspace. Relatedly, does the Single–Shared gap remain when the action-space displacement is matched?

  3. Is learnability limited by data or by the training signal? Please report teacher counts for the eligible cells and a learning curve over teacher count. If possible, please also report a proposer trained with closed-loop RL, or one trained on Dynamic corrections.

  4. What does TIDE establish beyond LEAP, Parrot/DSRL, POPLIN, and Expert Iteration? A discussion of these connections is needed to judge the novelty of the separation.

Reviewer_YzR4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper establishes a controlled empirical framework named TIDE to investigate how the temporal intervention structure affects intention control. Built on frozen, action-conditioned joint-embedding predictive world models (LeWM and Fast-LeWM), it fixes the backbone, standardizes the execution head and search budget, and only varies the temporal structure of bounded intention revision. Three paradigms are compared: single, shared and dynamic. It reveals the control interface gap that latent representations with accurate prediction do not inherently guarantee that reachable behaviors are easy to search or learn. The experiments cover four control tasks including Push-T, Cube, Reacher, and TwoRoom, along with two frozen backbones. Validated SHARED interventions can be partially amortized into search-free control, yet the gains are limited, and closed-loop improvements cannot be reliably predicted using offline learning signals alone. The results demonstrate that predictive latent representations themselves do not guarantee easy search or learning of reachable behaviors, exposing the control interface gap of such frozen visual world models.

Critical strengths

The paper freezes the backbones (LeWM/Fast-LeWM), adopts a unified execution head and a fixed search budget, and only varies the temporal structure of intent correction (SINGLE, SHARED and DYNAMIC ). Seeds, checkpoints and hyperparameters are fixed before results are inspected. By isolating the temporal structure as the sole variable, findings in the paper such as “SHARED is more stable and effective under a limited budget, and DYNAMIC reveals more verified successful behaviors” are not confounded by implementation differences. Negative results (modest amortized gains, offline signals failing to predict closed-loop improvements) also carry weight.

Critical weaknesses

The work is generally an empirical framework rather than a methodological innovation, and all adopted components (CEM, frozen world models of the JEPA family, action heads) are pre-existing technologies.

Questions for authors

What is the essential difference between the intent correction mechanism of TIDE and the intervention on object-level latent variables by Causal-JEPA during the representation learning stage? Where lies the boundary of their respective advantages and disadvantages?

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

Related papers

Discussion (0)

Sign in to comment.