acceptodds
Under review as a conference paper at ICLR 2027

Probe and Canonicalize: Active Identification of Hidden Physical Scale for Out-of-Distribution World Models

Abstract

Physical world models often fail to extrapolate when a physical factor leaves the training distribution, even when the governing interaction is unchanged. We study whether such a factor should be learned jointly with dynamics or identified separately and applied through known structure. Our active hidden-physics canonicalization predicts a factor-independent raw effect, estimates an unknown scale from a controlled visual probe, and applies it through a fixed analytical operator. In an object-centric environment where mass is absent from appearance, inverse mass is estimated from pixel displacement after a five-step probe calibrated only on training masses in [1, 4]. Across five backbone families and five seeds each, canonicalization reduces normalized final-position error by 58.8–60.9% on unseen masses in [6, 10] and improves every evaluated seed. For the mirror-attention backbone, error falls from 1.670 ± 0.028 to 0.680 ± 0.004, matching a true-mass oracle at 0.679 ± 0.003. Crucially, supplying the same estimate as a learned input yields only 1.585 ± 0.086, and supplying oracle mass as input gives the same result. Fixed analytical application therefore reduces error by 57.1% relative to learned conditioning: access to the factor alone does not ensure extrapolation. Interventions further show robustness to estimation noise while revealing that this benchmark depends primarily on recovering the unseen physical scale rather than precise per-object attribution. These results isolate a specific OOD entanglement that is prevented by placing identified physical structure outside the learned predictor.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
physical world models, out-of-distribution generalization, physics-informed learning, system identification, active perception, hidden physical factors, analytical canonicalization

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_fYj5OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper first points out that learned world models often struggle to generalize to unseen physical properties, such as object mass, because they implicitly learn the dependence between physical parameters and interaction dynamics. To address this, the paper proposes active hidden-physics canonicalization, which predicts a factor-independent motion effect, estimates the hidden physical parameter through a controlled visual probe, and applies the estimated parameter using a fixed analytical operator. The method aims to improve out-of-distribution generalization by explicitly separating known physical relationships from learned dynamics.

Critical strengths
  1. The paper investigates a meaningful problem in out-of-distribution generalization by examining how known physical relationships should be incorporated into learned world models.

  2. The experiments are well controlled, with consistent improvements across five backbone architectures. The additional ablation studies also help clarify the sources and limitations of the observed improvements.

Critical weaknesses
  1. The paper lacks sufficient baselines to isolate the source of the performance improvement. The learned-conditioning baseline only uses additive feature embeddings, while alternative conditioning mechanisms are not explored. More importantly, the paper does not compare against a standard world model with simple post-hoc scaling, making it unclear whether learning a mass-independent effect is necessary or whether the improvement mainly comes from correcting the predicted motion scale.

  2. The evaluation only considers a single direction of physical distribution shift, with models trained on lighter objects and tested on heavier ones. Since heavier objects generally produce smaller displacements under the assumed physical law, the observed improvement may primarily reflect motion scale reduction. Evaluation on lighter unseen masses, heterogeneous mass distributions, and longer prediction horizons would provide stronger evidence of generalization.

  3. The proposed approach mainly combines established techniques for physical parameter estimation and explicit physics-based modeling. The inverse-mass estimator uses simple linear regression, while the canonicalization mechanism applies a predefined physical relationship through fixed multiplication. Although the empirical improvements are substantial, the methodological contribution beyond existing physics-informed modeling approaches remains limited.

  4. Several important implementation details are insufficiently specified, including the simulator dynamics, action representation, training objective, and aggregation of normalized prediction errors. In particular, the assumptions underlying the inverse-mass factorization need further clarification. These details are necessary to interpret and reproduce the reported results.

  5. The overall research depth and experimental validation appear insufficient for a full conference paper. Several important aspects remain underexplored, including stronger baselines, the underlying reasons for the failure of learned conditioning, and generalization beyond the specific experimental setting. These limitations weaken the completeness and significance of the work.

Questions for authors
  1. Could the authors provide stronger baselines, such as alternative learned-conditioning methods and a standard world model with post-hoc scaling, to better isolate the contribution of the proposed canonicalization? Additionally, why does the current learned-conditioning baseline perform similarly with estimated and oracle mass?

  2. Could the authors evaluate the method on unseen lighter masses, heterogeneous mass distributions, and longer prediction horizons? Given the findings in Tables 4 and 5, it would also be useful to examine when accurate per-object identification is necessary and how probe length affects downstream prediction performance.

  3. Could the authors clarify the simulator dynamics, action representation, and training objective for the canonical raw effect? In particular, under what assumptions does the inverse-mass factorization hold, and how is the normalized prediction error computed?

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

The goal is to identify a hidden physical factor in a world model, based on viewing two frames of the output of the world model. Mass, friction and elasticity are mentioned as example hidden physical factors to be inferred, although only mass is tested.

Two methods to solve the problem are described: canonicalization (by observing the motion and then including an estimate of the inverse mass, or a canonical mass (please clarify the canonicalization branch--do you estimate mhat, or assert it?

The second method: active visual probing. A known force is applied to the object, and the resulting displacement is observed.

Critical strengths

The authors used 5 different backbones in the testing. Clear, if dense, writing.

Critical weaknesses

The domain of the paper is a somewhat toy domain. The standard model, which has no mass information, matches the mass-aware models in distribution (0.781 vs 0.780, l. 167–170). If mass were truly hidden and mattered over a 4× range, the standard model should do worse. Either mass can be inferred from the two input frames, or it barely affects the four-step outcome in distribution.

Questions for authors

What is m_ref, and where and when is m hat inverse computed and fed to Eq.1?

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

This paper proposes Active Hidden-Physics Canonicalization to improve the out-of-distribution (OOD) generalization of world models to unseen physical parameters. The method estimates hidden mass by applying a known force and observing pixel displacement, predicts mass-independent motion effects using a world model, and then applies inverse-mass scaling through a fixed analytical operator. Experimental results show that the proposed method consistently improves OOD prediction performance across different backbones and outperforms baselines that directly provide mass information to neural networks.

Critical strengths

The paper proposes a world model framework that combines active physical parameter identification with explicit physical constraints. By estimating hidden physical parameters through active probing, separating parameter identification from dynamics prediction, and applying their effects through a fixed analytical operator, the method provides a structured approach to improving OOD generalization in world models.

Critical weaknesses

W1. The method has limited novelty and relies on an overly simplified physical formulation. The core approach merely scales predicted displacement using the known inverse-mass relationship, offering limited technical novelty. Moreover, the authors assume that displacement is inversely proportional to mass without adequately considering initial velocity, collisions, or complex interactions, and provide insufficient theoretical justification for when this assumption holds. Consequently, the method's effectiveness may heavily depend on the simplified simulation environment, limiting its applicability to more complex physical scenarios.

W2. The experiments fail to demonstrate the necessity of per-object Active Identification. The authors estimate each object's mass through active probing and use it to correct motion predictions. However, Tables 4 and 5 show that prediction performance remains nearly unchanged even when mass estimates are shuffled across objects or replaced with scene-level or batch-level averages. Therefore, the experiments do not establish the necessity of the proposed per-object active mass identification mechanism.

W3. The paper suffers from poor presentation and limited readability. The method section lacks clear explanations, necessary mathematical derivations, and intuitive illustrations. Key concepts and experimental settings are also insufficiently explained, making it difficult for readers to understand the proposed method and its implementation.

Questions for authors

Q1. The authors assume that displacement is inversely proportional to mass. Does this relationship still hold in the presence of initial velocity, collisions, or complex interactions? Could the authors provide a more rigorous theoretical derivation and explain how the method could generalize to more complex physical systems?

Q2. Why does the learned-conditioning baseline use only simple additive feature fusion? Would the explicit physical operator still demonstrate significant advantages over other conditioning mechanisms?

Q3. Tables 4 and 5 show that prediction performance remains nearly unchanged even when mass estimates are shuffled across objects or replaced with average mass values. If accurate per-object mass estimates are unnecessary, how do the authors justify the necessity of per-object Active Identification?

Q4. The proposed method fails to achieve satisfactory results on Physion++. Could the authors further analyze the reasons for this failure and validate the method's effectiveness through more complex 3D scenarios and visualizations of predicted trajectories?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.