acceptodds
Under review as a conference paper at ICLR 2027

Disentangling Observation–System Coupling in Physical Reconstruction via Active Sensing and Intervention

Abstract

Physical reconstruction for industrial diagnosis, embodied interaction, and physical twins aims to recover object-specific properties that support prediction under new physical inputs. Existing methods often rely on a fixed passive protocol, where the acquisition scheme is predetermined rather than adapted to the observed object. However, distinct physical systems can remain compatible with the same measurements because intrinsic physical effects and observation effects are entangled. In this paper, we propose the PARAD (Physics-Aware Reconstruction through Active Disentanglement) framework, which models physical reconstruction through active sensing and intervention and separates physical response generation from measurement formation to analyze such ambiguities. Specifically, PARAD maintains a joint belief over physical and observation parameters from accumulated evidence, and sequentially selects intervention–sensing pairs to distinguish unresolved, task-relevant physical alternatives. Experiments across thermal, vibration, and wave-scattering systems show improved prediction under new inputs, transfer of informative evidence to an unchanged inverse solver. We also propose a foundation for physical reconstruction that actively seeks the evidence needed for prediction, interpretation, and verification.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Physical Reconstruction, Active Sensing, Experimental Design, Inverse Physics

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_X6nXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PARAD, a modular framework that reframes physical reconstruction as active disentanglement: interventions control which physical differences are expressed, sensing controls which remain measurable, and intervention–sensing pairs are selected to distinguish task-relevant physical alternatives under a limited acquisition budget, with interchangeable models and existing inverse solvers. The authors conduct fairly extensive experiments across thermal, vibration, and wave-scattering systems, demonstrating improved reconstruction and held-out prediction.

Critical strengths
  1. The distinction between improving an estimate on fixed data and acquiring evidence that rules out alternative physical explanations is useful and clearly developed. Separating physical response generation from observation formation provides an intuitive way to understand why different physical systems can remain consistent with the same measurements.

  2. The framework is demonstrated with different reconstruction components and physical systems, including an existing VVT inverse core and a frozen diffusion prior. The same-core replay, unchanged-history controls, and known-observation controls help separate some of the effects of acquisition, inference, and calibration. The combination of generated thermal data, public vibration simulations, real vibration video, and scattering experiments also gives the study useful breadth.

Critical weaknesses
  1. The benefit of feedback over a strong fixed design remains uncertain. The thermal comparison includes only four objects and one adaptive decision per object, while the larger scattering confirmation uses fixed acquisition. More detailed per-object results and planning costs would help establish the practical value of feedback.

  2. The reliability of the approximate acquisition scores is unclear. The planner relies on local linearization and an approximate joint covariance, but the paper does not establish how sensitive candidate rankings and acquisition choices are to these approximations.

Questions for authors
  1. For the thermal confirmation, could you report the per-object differences between feedback and the universal fixed design on the policy-unused new-task metric in Table 1, alongside the additional planning time? Appendix C.2 already provides per-object target-error reductions, but these do not directly establish how the new-task improvement is distributed across objects. Please also distinguish offline fixed-design selection costs from per-object online planning costs.

  2. How sensitive are the selected experiments to the local linearization and covariance approximation used by the planner? For example, whether a higher-fidelity estimate changes the candidate ranking or selected protocol—would help establish when the approximate utility is reliable. Can the authors relate this sensitivity to cases where feedback helps and cases where it does not?

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

The paper argues that in physical reconstruction, two explanations can be hard to tell apart, either because the experiment barely excites the difference or because an unknown measurement process can compensate for it. PARAD separates each experiment into a physical response operator F and an observation operator H, and keeps a joint belief over the physical and measurement parameters. A nuisance-projected sensitivity flags when the measurement can explain a difference away, and the method then picks interventions and sensing to resolve the alternatives that matter. The experiments cover thermal reconstruction, vibration tomography against an adapted VVT pipeline (including one real specimen), scattering with unknown receiver gains, and a historical exclusion test.

Critical strengths
  1. The framing is clean and I think people will reuse it. Separating F from H turns "the data can't tell these apart" into something you can check. The exclusion test convinced me most: the relaxed bound in Eq. 30 rules an alternative out for every admissible H, not just the estimated one, and Table 8 shows alternatives that fit under some legal H being excluded after one more record.

  2. The experiments are careful and honest. Methods are frozen before the confirmation runs, the replay controls reproduce the baseline exactly when the acquired data don't change, and negative results are reported plainly, for example the real-video frequency error getting worse. The controlled scattering study is solid, improving on 23 of 24 objects.

Critical weaknesses
  1. The headline number doesn't measure the mechanism the paper is about. The "up to 90.89% lower" error comes from a comparison that, per App. D.3, changes the representation, mesh, fitting window, inference and acquisition at once, without matched compute. PARAD also gets a prior matched to the inclusion family. With that many differences I can't tell how much of the gain comes from F/H-aware selection, which is the central claim.

  2. The evidence that targeted selection beats good fixed designs is thin. The thermal test is a single adaptive decision on four objects, feedback picks the same fixed protocol on two of them, and the gain has no interval while field error goes up. In the vibration replay, the acquired data only change on two objects, one of them outcome-gated, and on the other material error gets slightly worse. Scattering fixes acquisition by design, so it tests recovery rather than selection. Robustness to the observation model is also uneven (App. D.4). The paper doesn't overclaim (Sec. 5), but for now the core idea rests on a few small effects.

Questions for authors
  1. On Table 2A, what do your own inclusion model and the fixed sequence without feedback achieve? And how does the adapted VVT do with a comparable parameterization? That would separate the representation gain from the acquisition gain.
  2. Can you compare fixed and feedback acquisition under identical inference on more objects, with several adaptive steps and seeds, and against simple baselines such as random designs or treating the measurement parameters as known?
  3. Can the replay be rerun fresh on all six vibration objects with more than one noise seed?

Other suggestions:

  • The development exposure of the vibration objects, the single seed in the new scattering confirmation and the lack of a multiplicity correction deserve a more visible place.
  • The Fig. 4 exclusion result is post hoc on development objects.
  • Will the code, the thermal benchmark and the adapted VVT pipeline be released?
  • The identifiability discussion could be positioned against profile-likelihood identifiability (Raue et al., 2009) and observability-aware self-calibration (Hausman et al., 2017).
  • Curves of accuracy against acquisition budget would make the trade-off clearer.
Reviewer_VQCTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents PARAD, a framework that separates physical response from measurement formation and selects interventions and sensing configurations to reduce task-relevant ambiguity. Experiments cover thermal reconstruction, vibration tomography, and wave scattering, including prediction under unobserved inputs.

Critical strengths

Evaluating frozen reconstructions on new physical inputs is a useful test of recovery quality. The evidence-replay experiment through an unchanged VVT solver also provides a concrete check that better measurements can help independently of the reconstruction method.

Critical weaknesses
  1. I am not convinced that the methodological contribution is sufficiently distinct from goal-oriented experimental design with nuisance parameters. The explicit F/H decomposition is useful, but the covariance-reduction utility in Eq. (17) appears to be a standard local design criterion. A direct comparison with that formulation is missing.
  2. The evidence for adaptive acquisition is much thinner than the headline improvements suggest. Thermal feedback is tested on four objects with one adaptive step; the 90.89% vibration improvement also changes representation and inference; and the new 24-object scattering comparison fixes acquisition. These experiments support useful reconstruction components, but provide limited independent evidence for the central acquisition framework.
Questions for authors
  1. What changes algorithmically relative to nuisance-aware, goal-oriented experimental design using the same inference model and candidate experiments?
  2. Does feedback consistently outperform the DEV-selected fixed protocol on a larger independent test set with the solver held fixed?

Comments on the paper

1 comment
AuthorsOpenReview ↗

Urge for reply

Dear reviewers and ACs:

I hope this message finds you well, as the discussion period is near its end with less than 1 day, I want to ensure we have addressed your concerns satisfactorily. If there are any additional concerns, please let us know. We really appreciate your advice and comments.

Thank you again for your time and efforts.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.