acceptodds
Withdrawn from ICLR 2027

When Does Temporal Distance Improve World-Model Planning?

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Latent world models plan by keeping candidate action sequences whose predicted final state lies closest to the goal in latent space. This latent distance, however, is not trained to reflect reachability: states that are close in latent space may still require many steps to connect. Learned temporal distances can address this mismatch, but it remains unclear when replacing the latent distance will improve planning, or whether the benefit can be predicted before planning. We propose a simple offline diagnostic using trajectories already collected to train the world model. A useful planning distance should rank states requiring fewer steps to reach the goal as closer; we therefore measure how much a learned temporal distance improves this ordering over the original latent distance. With the world model and search fixed and the execution schedule matched, planning improves substantially when the temporal distance greatly improves temporal ordering, while gains are smaller when the original latent distance already captures that ordering. We further apply the diagnostic before planning in held-out settings, where its predictions match the observed distinction between large and smaller planning gains. Controlled experiments show that the large gains come from learning temporal structure rather than simply changing the representation, search budget, or execution schedule. The results suggest that temporal ordering in offline trajectories provides a practical way to identify when latent distance is a bottleneck in world-model planning.

Primary area:
reinforcement learning
Keywords:
world models, model-based planning, latent planning, temporal distance, reachability, hierarchical world models, pre-registration

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_eHphOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates when learned temporal distance improves planning in latent world models. The authors train a temporal representation on top of a frozen LeWorldModel and propose an offline diagnostic that measures improvements in temporal ordering using rank correlation between trajectory step separation and representation distance. The primary claimed contribution is an empirical observation that improvements in temporal ordering can serve as an indicator of potential planning gains from replacing the original latent distance. The authors evaluate this relationship through experiments on navigation and manipulation tasks.

Critical strengths

None

Critical weaknesses

The main diagnostic is simply the difference in rank correlation between trajectory separation and representation distance. While intuitive, this metric does not establish a principled connection to planning performance. Moreover, from a control perspective, temporal separation along recorded trajectories is not equivalent to optimal cost-to-go, and the diagnostic does not explicitly account for the underlying dynamics.

The experiments are largely restricted to a single world model framework and a small set of relatively simple environments.

Consequently, the central claim remains largely empirical without a clear characterization of when improved temporal ordering leads to better planning.

Please correct the author ordering in references "Contrastive Learning as Goal-Conditioned Reinforcement Learning" and "Decision-Metric Alignment in Latent World Models" to match their official publications. Also, I think the citation to Table 1 in the Figure 7 caption should be corrected to Table 3.

Questions for authors

Since temporal separation along recorded trajectories does not necessarily reflect optimal reachability, what justifies using its rank correlation with representation distance as an indicator of planning performance? Under what assumptions would this diagnostic be informative?

To what extent does the observed relationship between temporal ordering and planning improvement depend on the specific world model or environments? What evidence supports its applicability beyond the evaluated settings?

Reviewer_RrnTOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces the idea of “temporal distance” in JEPA-based world models. The paper argues that the learned Euclidean distance in latent spaces learned by JEPA models using the based predictive loss and anti-collapse regularization does not reflect the metric required for goal-based planning. Put simply, the Euclidean distance learned by JEPAs predictive loss doesn’t take into account how many decisions it actually takes to reach a goal and, therefore, it is harmful to planning when the scoring function of the planner relies on such ill-defined distance.

Critical strengths
  • The paper, as presented, lacks the polishing of an ICRL level paper. The authors failed to include a literature review that places their contributions in the wider field of model-based sequential decision making (RL and planning) where these problems have been core research endeavors for a long while. I would encourage the authors to deeply review the literature and include a discussion and proper positioning of their contributions.
  • The evaluation for such a simple addition is limited to a small number of domains.
  • The proposed patch is to learn a re-encoding (a new latent space) where Euclidean distance is meaningful: now, it would reflect the distance with respect their behavioral policy used to collect the data. This is, in fact, the underlying idea behind goal-conditioned value functions which I would encourage the authors to review. I believe the paper’s novelty is limited and its presentation not to ICLR standard. I recognize this might have some novelty in the world of JEPA, but I do not believe in its current presentation has the right level of novelty for an ICLR publication.
Critical weaknesses

The core observation that Euclidean distance in an arbitrary latent space is not equal to “how many actions away a state is from the goal given this policy” is correct. Hence, their method show some benefit empirically.

Questions for authors

See above.

Reviewer_xAFDOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper studies whether replacing the L2 latent cost of a frozen world model with a learned temporal-distance head improves planning. The authors use the ranking of states by steps-to-go as an offline diagnostic to quantify the benefit of the learned head before running the planner. They show that the learned distance produces a smoother planning landscape, raises planning success, and keeps success more stable over longer task horizons.

Critical strengths

The authors clearly show that the performance gain is attributable to the proposed cost and not to confounding factors such as the execution schedule, search budget, or representation rescaling.

The training-target controls isolate temporal supervision: reconstruct-only (50.8/29.2) and shuffled-future (48.3/22.5) stay close to the raw latent cost (42.5/32.5), whereas 20-step prediction reaches 93.3/82.5 across three seeds. PCA and whitening recover only part of the gap.

Critical weaknesses

Temporal distance as a planning metric depends heavily on the quality of the data-collection policy. Recovering a true steps-to-go distance requires near-optimal trajectories, but the LeWM datasets (e.g., two-room) are collected with highly sub-optimal policies, so the learned temporal distance is itself sub-optimal. The paper does not quantify how much this limits the method.

Novelty is limited. Learning temporal distances for planning is an established line of work, and prior papers have already shown that it is effective.

There is no related-work section, and the literature review is missing key references. In particular, the Temporal Latent Straightening paper (Wang et al., 2026) does something very similar for JEPA models and is already published, yet it is not discussed. Beyond the missing literature, the experimental scope is narrow: the only backbone is LeWorldModel, with a single encoder per environment. Many relevant alternatives exist and none are used as baselines. This is not strong enough to support the generality of the claims.

Questions for authors

Could the authors run a more complete experimental setup that includes other relevant baselines, e.g., DINO-WM with temporal straightening, temporal goal-conditioned behaviour cloning, and other learned-distance costs?

Could the authors add a related-work section that clearly cites and credits the prior work in this area?

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

Related papers

Discussion (0)

Sign in to comment.