acceptodds
Under review as a conference paper at ICLR 2027

Predictive Closure in Coarse-Grained World Models: Memory and Resolution

Abstract

Coarse-grained 3D scene simulators—including volumetric latent world models, BEV occupancy representations, and voxelised scene dynamics—replace high-resolution scene states with compact macro representations. Predictive insufficiency in such representations can arise from two distinct sources: inadequate temporal context or loss of predictive information during spatial aggregation. We introduce a two-axis predictive closure audit that separates these effects. Track A evaluates memory closure by measuring whether extending a length- macro history reduces out-of-sample prediction error. Track B evaluates feature-relative resolution closure by testing whether a declared micro-level feature provides additional predictive information beyond the same macro context, using a strictly nested, cross-fitted residual-augmentation procedure. We further establish two theoretical boundaries for predictive closure: the same observed macro law can arise from both strongly lumpable and non-lumpable micro processes, and finite recursive-state dimension does not imply finite observation-window memory. These results motivate testing predictive closure directly rather than inferring it from macro prediction accuracy or state dimension.

Primary area:
learning theory
Keywords:
predictive closure, world models, temporal memory, spatial aggregation, coarse-graining, predictive sufficiency

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_NfttOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies predictive sufficiency in coarse-grained world models by examining the value of additional temporal context and micro-level information separately. Track A compares prediction risks across history lengths, while Track B tests whether micro-level features improve prediction beyond the same macro history through residual augmentation. Two theoretical constructions illustrate that macro observations alone do not identify strong lumpability, and that a finite-dimensional recursive state does not imply a sufficient finite observation window. The main experiments use a controlled linear voxel system to compare the two empirical prediction gaps under different aggregation regimes.

Critical strengths

1.The paper addresses a useful question. Low prediction error alone does not establish that a representation retains sufficient information. Examining the gains from additional history and micro-level features separately could help determine how a model should be improved.

2.The theoretical examples clearly explain the limits of inference from macro observations. In particular, the construction of identical macro laws with different strong-lumpability properties shows why macro prediction performance cannot by itself establish properties of the underlying micro dynamics.

Critical weaknesses

W1. The dependence on predictor and input choices is not evaluated, leaving the practical value of the diagnostics unclear. Both tracks depend on whether the chosen predictor can use the additional information, but the experiments mainly use regularized linear models. The paper does not compare predictors with different modeling capabilities or systematically examine the effect of micro-level feature selection. It is therefore difficult to tell whether the absence of a gain means that the existing information is sufficient or that the predictor cannot use the additional information effectively. The authors explicitly limit their conclusions to the chosen models and features, but the practical impact of this restriction still needs to be evaluated. Track A also fixes the reference history length at 8 without showing whether the conclusions hold with longer reference windows. There is no comparison with simple alternatives, such as selecting history length on a validation set or directly comparing models with and without micro-level features. It remains unclear whether the two-track audit provides reliable guidance or adds value beyond standard model selection. W2. The experiments are insufficient, with limited validation in the intended applications and little visual analysis. The main synthetic experiments are limited to linear dynamics on a 4 × 4 × 4 voxel grid. Both tracks yield non-positive estimates in all four joint settings in Section 6.4 and Table 3, and the other synthetic Track B experiments also provide no positive examples. These results do not establish that the method can detect additional history or micro-level information when it is genuinely predictive. Nor do they show that the diagnostic recommendations lead to effective model improvements. The experiments therefore do not substantiate the applications to 3D scenes, BEV representations, or voxel world models emphasized in the paper. The results are also presented mainly as numerical tables, with few visual explanations connecting the diagnostic values to actual prediction behavior. Missing analyses include error curves with uncertainty intervals across history lengths, visualizations of how the diagnostics change with spatial aggregation, and side-by-side views of micro states, coarse representations, and predictions. There are also no concrete success or failure cases showing what changes after extending history or retaining more spatial detail. This makes it difficult to interpret the numerical differences in a practical setting. At a minimum, the paper needs validation on controls with known gains and known null effects, together with an application showing that the diagnostics correspond to actual improvements.

Questions for authors
  1. How do the diagnostic results change when linear predictors are replaced with nonlinear predictors, or when different micro-level features are used? Can you distinguish a lack of additional predictive information from a predictor’s inability to use it?
  2. Are the Track A conclusions stable with reference histories longer than 8? What additional information or practical benefit does the audit provide over selecting history length and feature inclusion directly on a validation set?
  3. Can you provide controlled cases with known positive gains and known null effects to establish the detection ability of each track? How does performance change with sample size?
Reviewer_D9itOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces an Analysis technique, referred to as 'Two-axis predictive clusure edit'. It aims to distinguish the cause of predictive insufficiencies of networks as either due to (Track A) a lack of temporal context or (Track B) a lack of spatial aggregation in coarse-grained latent BEV or voxel grids. The authors argue that categorization in these tracks allows for directing improvement actions towards extending spatial context (Track A) or revising spatioal representation (Track B).

Critical strengths
  • The gaps D_A(L) and D_B(L,phi) are intuitive and mathematically well defined.
  • The theoretical propositions in Sec. 4, namely 'Macroobservations do not identify strong lumpability' and 'Finite recursive state does not imply finite observation Memory' are useful and clarify the underlying problem.
Critical weaknesses
  • Applicability in real networks. It seems that application of the method to existing networks will involve heavy retraining. Hence an indication of superiority of established ablation studies should be presented.
  • Insufficient indication of practical impact: While the distinction into Track A and B may provide hints to improve learning algorithms, the practical use of the analysis to improve ML-networks is not sufficiently shown. The usability therefore stays a little hypothetical.
  • Synthetic Experiments only: Experiments are limited to synthetic and over-simplified toy examples. At least one convincing improvement in a realistic application would be appreciated. How would this method be applied and what improvements could be achieved in modern networks, e.g. diffusion or transformer networks?
Questions for authors
  • How are your tolerance Parameters selected?
  • Your Experiments basically Show Track A reasons. Can you report an example where Track B is signalized?
  • How can you avoid that temporal information is actually aggregated by enhancing (e.g enlarging) the micro state?
Reviewer_xxXkOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper presents a diagnostic for coarse-grained 3D scene simulators, such as volumetric latent world models. It asks whether insufficient prediction comes from missing temporal context or from information lost through spatial aggregation. The proposed two-axis predictive closure audit compares predictors out of sample. In particular, Track A tests whether a longer macro history improves prediction, and Track B tests whether a declared micro-level feature does. Two propositions state that the macro law does not identify strong lumpability and that a finite recursive state does not imply a finite observation window.

Critical strengths
  • The paper distinguishes between predictive information recoverable from additional temporal context and information available only through finer-grained observations.
  • The point of the paper is relevant: does the prediction error of a coarse model come from not enough temporal context or from information lost in aggregation.
Critical weaknesses
  • The title, the abstract and the introduction focus on world models. However, no learned world model, occupancy model or image data is referenced. The only simulator is a linear model on a 4×4×4 grid, divided into eight blocks, and all the predictors are linear autoregressions. The real-data study only appears in the appendix. The related work on world models covers video generation, memory and urban forecasting, but not the focus of this paper, e.g. the BEV world model.
  • The writing in this paper is really bad and hard to understand. Moreover, there are multiple unresolved citation placeholders in the manuscript, marked with (?). Most of the concepts such as “memory closure”, “resolution closure”, “lumpability”, and many more introduced in the paper are not explained or are poorly explained.
  • The experiments do not validate the claim of the paper. The paper aims to distinguish temporal insufficiency from information lost through spatial aggregation, but the experiments do not demonstrate this capability. Both gaps are non-negative by definition (Eq. 1 and 3), yet every reported Track A and Track B estimate is negative, with confidence intervals below zero (Tables 1–3 and 5). The paper acknowledge that their four-setting experiment cannot separate the four closure combinations (Table 3). No experiment therefore shows that the audit detects a gap when one exists, and the conclusion is not supported by the results.
  • The reproducibility statement says that extended proofs and derivations are in the supplementary material. The supplement, however, contains a folder with no actual proofs.
  • Overall, the paper is not up to the level of an ICLR publication at all.
Questions for authors
  • Is it possible to show a configuration where the true memory or resolution gap is positive and the audit detects it?
  • How would the audit be applied to an actual learned world model, such as a BEV-occupancy model?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.