acceptodds
Under review as a conference paper at ICLR 2027

Conditioning Is Not Control: Scientific Endpoint Controllability for Tensor-Conditioned Generative Models

Abstract

Property-conditioned generators are increasingly used for scientific inverse design, but a conditioning variable need not be an action that controls the requested physical response. We formalize scientific endpoint controllability for tensor-valued targets by composing a generator with an independently specified scientific evaluator and quotienting physically equivalent representatives. This formulation separates three notions that are often conflated: structural recoverability, conditioning-path visibility, and local endpoint reachability; direct target tracking further requires local inverse-action calibration. On a matched clamped-piezoelectric benchmark, tensor orbits are strongly recoverable from structure (CEITNet proper-orbit cosine 0.883), yet conditioning interventions reveal a distinct actuation geometry: scalar conditioning is locally rank one, whereas full-tensor routes expose roughly twelve numerically visible directions while concentrating most response energy in about two stable-rank directions. Common-noise swaps further show that a nonzero field response can remain nearly orthogonal to the requested correction. Independently, response-state analysis shows that endpoint specification materially changes the target, while DFPT separates many-mode lattice dynamics from the pointwise Born-charge-visible response channel. These results provide a mechanism-level diagnostic criterion for locating which stage limits endpoint control despite condition responsiveness.

Primary area:
generative models
Keywords:
generative models, controllability, scientific machine learning, inverse design, geometric deep learning, equivariant learning, symmetry quotients, tensor-valued properties, materials generation

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

This paper discusses property-conditioned generators for scientific inverse design like material design. The authors claim to formalize controllability for tensor-valued targets by composing a generator with an independently specified scientific evaluator and qoutienting physical SO(3) symmetry. They experimentally evaluated their method comparing some baselines like EATGNN and GMTNet and show merits of their method.

Critical strengths

(S1) Control or guidance of generative modeling is important and interesting problem. Especially, considering the symmetry could be a starting point of the discussion. Also, the problem setting is understandable.

(S2) Qoutienting physical symmetry is straightforward and easy to understand.

Critical weaknesses

(W1) For me, the organization of paper is not clean. As a result, it is difficult to catch the motivation (e.g. example of target problems), contribution, and novelty.Thus, I struggle to understand the main concepts of this paper.

(W2) It would be quite appreciated if authors would clarify the novelty of the analytical contribution. In particular, to clarify the novelty and contribution of Theorem 1 might be important.

(W3) The figures are not so effectively used. For example, non-expert readers (like me) struggle to catch the main concept of this work from Figure 1. It would be better to reconsider the organization of components of Figure 1.

(W4) Experimental evidence is somehow limited.

(W5) Related to (W1), some sentences are not easy to read. For example, many sentences have specific numbers like numbers of dataset or accuracy. Personally, I think that kind of information is too detailed and reduces the readability.

Questions for authors

(Q1) Please explain again the motivation (e.g. target problem), main concept, and contribution in language that is easy to understand even for broad ML readers.

(Q2) Please reconsider the organization of this paper and refine the figures to clarify the concept of this work.

(Q3) Could you discuss the novelty of the analytical contribution (in particular Theorem 1) and physical meaning of the theorem. Since my understanding might be limited, I cannot judge how valuable and interesting the discussion in this work is.

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

This paper identifies the "conflation of property conditioning with physical controllability" and the lack of "endpoint-aware evaluation" in previous works. Based on their insight that conditioning sensitivity does not necessarily imply target-directed control, they propose a scientific endpoint controllability framework that combines "symmetry-aware physical endpoint evaluation" and "conditioning-path Jacobian" analysis. The major contribution is a combination of endpoint controllability and conditioning-path visibility for diagnosing whether tensor-conditioned generative models can realize desired physical responses.

Their experiments on JARVIS tensor benchmarks show that structural properties can be accurately recovered, while scalar conditioning is limited to rank one and full-tensor conditioning, demonstrating that increased conditioning sensitivity does not guarantee target alignment or physical endpoint control.

Critical strengths
  • [Motivation] The proposed research question is a general and important question in scientific discovery, ie, conditioning guidance in generative models is hard to correctly guide the generation with expected physical meanings.

  • [Theoretical contribution] The proposed theoretical framework combines physical evaluation and symmetry to conduct endpoint controllability is a natural direction to address the problem, and the theorem proposed can be seen as a general method.

Critical weaknesses

Weaknesses:

  • [Presentation] However, the presentation for the methods and proposals is poor. Specifically:
    1. The language uses some method-specialized words without clear explanations, such as "locally visible actuation directions", DFPT, making the work hard to understand (although the words can be understood with strong context in this paper).
    2. Many notations are used without definition when they first appear, making readers confused when first reading.
    3. The work is not well-structured; the background is only some related work but lacks necessary definition or explanation for the work, such as the definition of "local controllability/ quotient reachability", and the background research objective (the crystal-related background), etc.
    4. Many Figures and Tables are not used (and cited) in the main text, making the results descriptions difficult to read.
  • [Methods] The paper introduces a theoretical framework, analyzes the recoverability, condition trajectories, and endpoint transfer, but it lacks concrete, unified methods.
  • [Results] The main claims in the results part are not clear. The authors do not successfully and explicitly present their points and observations based on these experiments, although the figures show some meaningful results.
  • The work lacks experimental and implementation details.
Questions for authors

Please address the Weaknesses if possible.

Reviewer_za6yOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposed a perspective for evaluating the failure causes of conditional generative models for piezoelectric material design. I found the way the authors interpret the generation failures, that separating the generator-side Jacobian and endpoint-side Jacobian, is interesting and conceptually useful. However, I think the current paper is quite unfinished. The several theoretical and empirical outcomes (conditioning sensitivities, physical reachability, etc.) are distinct from one another. I understand the author's attempt to break the black-box generation failure by some testable claims, but to me, they are not easy to transfer to other regions, and the paper failed to provide any general insights into the reasons for the failures and how to avoid them (e.g., by advancing the training pipeline or enhancing the training data). Related to this, the paper claimed itself to be a general theory of “scientific controllability for tensor-conditioned generative model, but the study is entirely centered on one piezoelectric material case, and quite a few analyses are specific to that case (e.g., clamped/ionic/relaxed response definitions, DFPT factorization, etc). I consequently see this work/contribution more suited to a domain-specific journal instead of the iclr conference.

Critical strengths

The conceptual separation/distinction between the two Jacobians is potentially useful for identifying the cause of failure for physical generation models.

Critical weaknesses
  1. The paper measures generator-side trajectory geometry for trained models and endpoint-side physics separately, but fails to explain how these factors contribute to the end-to-end failure of a generative model. The analysis of root causes and mitigation directions is also limited and too separated (e.g., high-dimensional condition input).
  2. The paper overclaims to present a general theory, but the experiments and analysis are too specific to the piezoelectric material case and fail to provide useful insights transferable to other cases.
Questions for authors
  1. Can the authors quantify the relative contribution of the proposed failure modes (e.g., conditioning visibility, endpoint controllability, etc.) to the final inverse-design error?
  2. Can the paper demonstrate that the proposed diagnostics lead to practical improvements that may be useful beyond the piezoelectric case?
  3. What is the reason for the strong anisotropy under full-tensor conditioning? Is it caused by degenerated training data or the model’s representability?
1 comment

I thank the authors for their effort to address my comments and concerns. I acknowledge the contribution of the decomposition and the added attribution of generation errors. But I still found this to be “one step away” from a real contribution. To my understanding, the physical-side error cannot be improved directly, since it comes from a real physical solver, so it would be super interesting if one could use this signal to guide better generator selection. My concern about case generalization also remains. Although the authors claim their formulation is case-irrelevant, I still recommend including another real case (even a very simple one) when aiming for an ML conference, especially considering the error decomposition can be totally different from case to case (the strong anisotropy shows that the case in the manuscript may be an extreme case for inverse design; please correct me if I’m wrong). Based on the above points, I don’t think the authors would make substantial improvements that may change my opinion in the discussion stage, and the score would remain (=2).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.