acceptodds
Under review as a conference paper at ICLR 2027

Variable binding is not a property of representations

Abstract

Variable binding is the capacity of a network to hold each attribute of an object as a separable part of the object, so that attributes and objects can be recombined into novel arrangements. The field has long believed that binding is an intrinsic property of a representation and thus, sought to measure and certify it. However, we demonstrate it is not such a property. We first prove that how cleanly the attributes of an object can be separated is fixed by the observation itself. This is done before any encoder is trained, so no optimizer or architecture can improve the separability. If the objects in an observation are separable, a representation can bind them and where they are not, it can’t. We then show that the leakage metric the field currently uses to certify that a network binds changes from zero to positive under a lossless relabeling, and thus, no architecture-agnostic certificate can exist. Across seven architectures, including the tensor-product representation, an information-preserving relabel is able to change each model's binding verdict showing that variable binding is truly not an intrinsic property and the object entanglement predicts the out-of-distribution gap on real photographs, and reducing this entanglement closes the binding gap.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Variable binding, Neural networks, Representation Learning, Compositional generalization

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

The paper is studying binding and the extent to which it is a property of the representation. The first result that the authors show is a "binding limit" that show that conditioned on the roles of the items and the observation the representation cannot reduce the correlation between fillers without losing information. The second result shows that an "information preserving relabeling" can appear to bind without losing information. The results of the paper are evaluated experimentally.

Critical strengths

I find the problem interesting in general and the formulation could lead to an interesting modeling of the problem.

Critical weaknesses

The paper is hard to follow, e.g. it uses notation that wasn't defined before. As far as I can tell the proofs are rather simple and the resuolts not particularly surprising. Thus, I assume that the most significant part of the paper is the modeling of the problem. Unfortunately, not enough attention is given to explain why this is a good formalization of the problem. In general it seemed to me that there was a disconnect between what the paper claims to show in the title, abstract and intro and what the theoretical result is. The model does not seem to me to faithfully capture the scenario discussed in the intro: I don't quite understand why the color of a the car and the color of an apple become correlated once someone sees the observation XX (I assume that is supposed to mean the image).

I believe that the paper would benefit a lot from a more in depth discussion of the assumptions and modeling choices and why this captures the problem of interest. If these assumptions are standard a reference would be useful (I didn't see any).

Questions for authors

What is the intuition behind the correlation of fillers conditioned on the observation?

It is especially unclear to me in the car and apple example. Clarifying that would help me understand the modeling of the problem much better.

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

The paper studies binding (i.e., how fillers are associated with roles) and shows that binding depends on the input data's entanglement (e.g., overlapping objects) and on how the representation is used by the reader (i.e., how it is used downstream). Binding is thus not an intrinsic property of a representation, and it cannot be assessed from the representation alone.

The authors show two main results:

  • The first is the binding limit: the input data sets a limit on how well the fillers of different roles can be separated. In other words, it shows that when the model keeps all the information about the fillers, better training or a stronger architecture cannot fix this; only discarding information can. This is measured using total correlation, i.e., how much worse you do, in log-loss, when predicting the fillers from the representation separately, role by role, instead of jointly.
  • The second question challenges the test commonly used for demonstrating binding, which works slot by slot (where a slot is a part of a representation reserved for one role). This test asks whether one slot carries information about another role's filler. They show that this test is not reliable, as the representation can be rewritten reversibly (without discarding information) such that the verdict changes (e.g., from no leakage to leakage). In contrast, total correlation is invariant to such rewrites.

Experimentally,

  1. In many real models, a subspace that tracks which filler goes with which role can be identified. Swapping it between roles doesn't discard information, but flips the output of the (unchanged) reader.
  2. In a toy experimental setting (where the right answer can be computed exactly), they train models such that they keep all the filler information and check if the total correlation between fillers matches the value predicted by the binding limit, and find that it does.
  3. On image datasets (modified versions of CLEVR and MNIST), the authors show that the amount of overlap between objects is correlated with the drop in performance for unseen combinations. This effect is strongest in settings where the authors control the overlap, and weaker on natural examples. On the Visual Genome dataset, reducing the overlap in the training set mitigates the performance drop.
Critical strengths

None: there is no critical strength, but the paper does formally show a new result (to the reviewer's knowledge): namely, that slot-wise leakage is not invariant to information-preserving rewrites of the representation, so it is not sufficient to certify binding.

Critical weaknesses
  • In its current form, the paper is very difficult to follow. It relies on numerous non-standard notions/word choices for the machine learning community (e.g., role and filler instead of object and attribute, rendering instead of data-generation process, reader instead of probe/head, minimal slot with a not mathematically accurate meaning of "minimal", etc.). Furthermore, it uses several names for one notion (e.g., the invertible edit is called relabel, reparametrization, information-preserving, and lossless), which further complicates the reading. Finally, the abstract relies on multiple terms without defining them first (binding verdict, binding gap, etc.), and further never uses those terms in the paper.
  • Some of the theoretical results are close to trivial. The binding limit (Lemma 3.1) follows almost directly from the definition of joint sufficiency; the invariance of total correlation holds by construction, etc.
Questions for authors

Could the authors consider choosing more standard wording, using one term per notion, adding a glossary, and making the abstract self-contained?

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

This paper proposed a view against the long-believed view that variable binding is an intrinsic property of a representation, which presents several conclusions. Firstly, the authors show that residual dependence among fillers which are measured by role-conditioned total correlation, is fixed by the observation for representations that preserve information and can be reduced by discarding at least as much joint information. Moreover, the paper also constructs an invertible information-preserving relabeling that mixes role representations and raises the measured leakage on a representation from zero. Various experiments are constructed covering Language, Vision-Language, State-Space, Text-to-Image Models, and etc.

Critical strengths

The structure of this paper is clear and the content is easy to understand. The experiment covered a broad range of model families was well designed, visualized and provided with detailed explanations, including targeted subspace swaps and random-subspace controls.

This paper shows that the preservation of all information does not ensure the unchanged slot-level leakage, which also distinguishes filler from joint sufficiency. These results and conclusion provide meaningful theoretical guidance for related aspects.

Critical weaknesses

Although experiments contain various model families, the task mainly focused on controlled role–filler retrieval and attribute binding, which may ignore the demonstrated practical relevance of the conclusions.

The practical benefits of reducing entanglement remain insufficiently established, where the de-occlusion intervention requires object annotations and introduces substantial image modifications.

Questions for authors

While the paper demonstrates impressive results, the lack of open-source code hinders other researchers from reproducing the experiments or building upon this work. Could the authors please provide a link to the code to improve the reproducibility of the research?

Does de-occlusion improve performance on unseen combinations, or reduce the difference between in-distribution and out-of-distribution performance?

Lack of model-access, computational requirements and annotation for the proposed audit.

Comments on the paper

1 comment

Thank you for the careful responses, with the code availability, the audit’s model-access, and annotation requirements, address several of my problems. The explanation of de-occlusion, serving as a controlled test of the theory rather than a method, addressed my concerns. The responses clarified the methodology and experimental intent, but do not fully resolve my concerns about the interpretation and practical relevance of the results. I will therefore maintain my original score of 2 (Weak Reject).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.