acceptodds
Under review as a conference paper at ICLR 2027

NAILONG: Freeze the Shared Lie Algebra of Generators

20 views

Abstract

Two operator families can share a continuous symmetry and still differ in the residual that breaks it. The method extracts the shared Lie algebra of infinitesimal generators and freezes it on the target. A jointly equivariant predictor is capped by the group average; a trunk plus a free residual is already outside that class, so the trunk is a comparison, not a corollary of the ceiling. With the shared generators known in the control chart, freezing them and fitting the residual drops matched-design target OLS risk from to . The algebra can be recovered from a finite source by identifying the span, aligning it to the target chart, and freezing it. On noisy generator OLS that procedure approaches the oracle residual rate at , against oracle and scratch . On a separate labeled spherical field probe, generator regression followed by a freeze of reaches residual-of- on the lab axis; a jointly equivariant residual stops at the group-average fraction .

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
transfer learning, Lie algebra, equivariance, operator learning, infinitesimal generators, sample complexity

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_saZPOpenReview ↗
Rating
1
Confidence
1(−1)
Informativeness
1(−1)
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies operator transfer when the source and target share continuous symmetries but have different symmetry-breaking residuals. The authors propose NAILONG, which identifies the shared Lie-generator span from source data, aligns it to the target coordinates, and freezes it while learning a target residual. The theoretical analysis covers target OLS risk reduction with correctly calibrated shared generators, group-projection limits for jointly equivariant predictors, and error floors when the residual cannot compensate for an incorrect freeze. Experiments on noisy generator regression and spherical fields examine finite-source identification, alignment, target sample efficiency, and adaptation to symmetry-breaking targets.

Critical strengths

The paper provides a useful theoretical distinction between the variance reduction from a correctly calibrated freeze and the bias from an incorrect freeze. Under the stated matched-design OLS and residual-class assumptions, this clarifies the conditions under which transferring shared generators can be beneficial. This gives a helpful basis for assessing the proposed transfer mechanism.

Critical weaknesses
  1. The practical benefit of transferring the identified generators is not sufficiently isolated. Section 6.4 states that generator OLS uses analytic velocity supervision and ordered control paths, while the Wigner baseline uses endpoint supervision and control summaries. These different information conditions affect the comparison. Therefore, the reported field gap does not establish the advantage of the identify-align-freeze strategy under matched transfer conditions.
  2. The inspected code raises additional consistency concerns. The resolution lift uses an oracle branch and true higher-resolution generators, while the main unknown-intertwiner grid reuses target samples for alignment and residual fitting. These choices differ from the stated learned-generator and disjoint-sample protocols. The authors need to clarify the version used for the reported results and verify the corresponding experimental claims.
Questions for authors
  1. Can you provide one matched field comparison with a source-trained trunk plus a free residual, using the same supervision, control information, residual capacity and total target budget?
  2. Which code version produced the reported resolution-lift and unknown-intertwiner results? Please clarify whether the lift uses identified generators and whether alignment samples are disjoint from residual-fitting samples.
  3. Are the ceiling and method risks evaluated under the same Haar-symmetrized joint law? How do you verify joint equivariance of the baseline to which Theorem 2 is applied?
  4. Please improve the figure layout and readability. On page 7, the spacing around Figures 3–4 and Table 1 could be adjusted to enlarge Figure 3, especially its legends and heatmap labels. The x-axis tick labels in Figure 2 are also crowded and should be simplified. These changes would make the results easier to check.
Reviewer_aWRYOpenReview ↗
Rating
1
Confidence
1
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

I have spent three hours reading this paper, but I am not able to summarize it now. I have 20 years of experience in research in ML and during these 20 years I worked on many different problems, and even this experience was not sufficient for understanding the paper. I was even not able to understand the formulation of the problem the paper is focusing on. Presumably, if I was chosen as a reviewer for this paper, it is because there were no more qualified reviewers. Since I did not manage to understand it, I believe that ICLR is not the appropriate venue for this manuscript.

Critical strengths

None.

Critical weaknesses

The paper is impossible to understand.

Questions for authors

I have no questions. Just a remark: it is common practice in ML community to define the notions and notations that are used in the paper. Adopting this would greatly help in making the paper more understandable.

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

This paper analyzes operator families in the context of shared continuous symmetries. The authors appear to demonstrate improved outcomes resulting from their methods.

Critical strengths

None. I may be under-appreciating the contribution, but I leave it to the authors to address the weaknesses and questions below.

Critical weaknesses

Weakness 1: the paper severely lacks precision regarding the problem definition and methods employed. For example, in line 120, the authors allude to a shared quantity across "source and target families," but the authors have provided no meaning for these terms. Basic key terms like "operator family" and the symmetry "residual" are not defined: it is also not made clear what the precise meaning of freezing is. I also cannot access the anonymous GitHub link provided, casting doubts on well-definition and reproducibility.

Weakness 2: discounting a lack of precision, the quality of the paper is lacking. The related work section does not serve the paper and is the functional equivalent of a hastily-combined and arbitrary selection of potentially-related literature, the very brief summaries of which are not always precise: for example, an apparent distinction between discovery via Lie derivatives and recovering infinitesimal generators is made. Figure 1 is also a prime example of the paper's clear lack of suitability for publication: unreadable fonts, missing or misplaced figure attributes, and eggregiously small print. Throughout, the writing is not clear, though this is more especially relevant to weakness 1.

Questions for authors

Question 1: What, in complete, precise, and unambiguous terms, is the objective of this paper?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.