acceptodds
Under review as a conference paper at ICLR 2027

Position: The Linear Representation Hypothesis Needs a Group Action

Abstract

To make claims about representations that generalize beyond a particular trained model, we need to specify when two representations should count as equivalent. The Linear Representation Hypothesis is often discussed without making this equivalence explicit. Different notions of equivalence preserve different structures, so metrics, probes, and interventions that appear to study the same representation may in fact correspond to different hypotheses. We therefore argue that the Linear Representation Hypothesis is not one hypothesis but a family of claims distinguished by representation equivalence. We formalize this idea using group actions, specifying the representation object, the procedure that produces it, and the property ultimately asserted, while accounting for equivalences imposed by the model architecture. This framework clarifies how assumptions can change across metrics, reading points, and analysis stages, and we use it to audit common representation quantities and recent interpretability analyses.

Primary area:
interpretability and explainable AI
Keywords:
interpretability, linear representation hypothesis, representation geometry, position paper

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_vGeNOpenReview ↗
Rating
3(+2)
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Summary

This is a position paper arguing that the Linear Representation Hypothesis (LRH) is actually a family of hypotheses since it has several formulations that take place in different spaces and, depending on the space, different transformations keep the object equivalent. The authors propose that the LRH be accompanied by a representation claim, which is a four-tuple including the specification of the space in which the object lives, the group that acts on the object, the procedure that produces the object, and the predicate ultimately asserted. They further define a concept of the admissibility of a representation claim, which allows for some interesting observations about the impact of the architecture on the possible group. If the practice of stating this specification were to be adopted, it would make it easier to quickly see when analyses in different papers are comparable.

Recommendation

While the paper is well-motivated and the idea of stating a specification of a representation claim is a good one, the writing and organization need more improvement than is likely possible in a short rebuttal period and therefore I recommend the paper be rejected. My supporting arguments are the weak points listed below. Briefly, I think there is too great of a burden on the reader to work out the details of Table 1, an example of a representation claim, whether the paper is concerned with multiple formulations of "linear representation" or the linear representation hypothesis, and the relationship to existing work.

Critical strengths
  • The idea to give a specification along with a claim about a linear representation is a good one, and the authors provide helpful examples.
Critical weaknesses
  • Is the paper concerned with the ambiguity of the linear representation hypothesis or ambiguity of "linear representation"? The authors do not state the LRH. Section 2 is titled "The linear representation hypothesis is not one hypothesis". In section 2, they list different formulations of linear representation. However, is this really supporting the claim that "The linear representation hypothesis is not one hypothesis" or is it instead supporting the claim "there are multiple interpretations of linear representation". Is linear probing really treated as equivalent to the linear representation hypothesis in the literature? Maybe if the authors defined LRH it would be easier to understand their claims.
  • The related work is mixed in throughout the paper rather than in a separate section. Consequently, the answer to the following question is scattered: have other authors discussing the LRH made an effort to discuss the implicit assumptions and take care with the issues the authors bring up? For example, Park et al (2024) discuss different interpretations of linear representation. While the authors cite Park et al several times, they should add an additional citations to make the following clear: are the three common formulations the authors list in Section 2 precisely the three interpretations Park et al list in Section 1? (I don't think they are, but I think there is enough similarity to warrant a citation).
  • The paper in its current state is a collection of ideas and observations, some of which are interesting and nicely explained, but I think the authors need to spend some time thinking about what they want to emphasize in the 9-page limit and how to best organize it. Some paragraphs contain redundancies or sentences that don't seem to add much content to the paper, and some parts feel rushed. To make this critique more concrete: section 5 gives Table 1 which I think has potential to be one of the more interesting and useful parts of the paper, but currently has very little exposition. Adding definitions or citations so that the reader can find the definitions (what definition/estimator are you using for intrinsic dimension?) would help. Section 7 discusses how the Platonic Representation Hypothesis seems to suffer from some of the same issues which I think should be a lower priority than explaining Table 1. I find Section 8 confusing as written. My confusion about Section 8 is detailed in Questions For Authors/Additional feedback below.
Questions for authors

Questions whose answers could materially affect my assessment

The first question in critical weaknesses above.

Additional feedback with the aim to improve the paper

Introduction

  • The second and fourth paragraph of the introduction seem redundant.

Section 2

  • I recommend re-reading the second paragraph and rewriting to improve clarity.

Section 3

  • Giving an example of a predicate could be helpful. Even better would be an example with G, M, F, P explained.

Section 4

  • What is a reading point?
  • Corollary 4.2, as written, is lengthy, including a short proof/justification (lines 260-261) instead of just consisting of a statement. Is the first statement the Corollary, and the rest is the proof (providing an example of an architecture that induces different symmetries at different reading points?)

Section 8

This section seems overly verbose for the content it contains. First, it states a 'reporting principle' but could this be stated earlier? Then, the authors write "This requirement has a real cost" but the first sentence after that begins "Some conclusions become narrower once their geometric assumptions are made explicit..." One can argue having conclusions stated corrected is a benefit, so perhaps there is a stronger cost to start with? Then the final two paragraphs begin with 'A second objection is...' and 'A third objection is...' respectively, so I'm taking that to mean that we are discussing potential objections to the position. I think it could be made more clear in the section introduction and title that this is where the section was headed.

Minor comments

  • In the third paragraph, the authors should state what V and \mathbb{P}(V) are (lines 41 and 43)
  • In the first occurrence of 'equivariance' (line 52) the authors should define it.
  • Could help to say what you mean by: "changes of description", "nuisance group" (line 124)
3 comments

We sincerely thank the reviewer for the thorough and careful reading of our paper. We especially appreciate the detailed comments and suggestions, which reflect a level of engagement that we do not take for granted, particularly given the time constraints of academic reviewing. We have revised the manuscript accordingly, and our point-by-point replies follow below.

Is the paper concerned with the ambiguity of the linear representation hypothesis or ambiguity of "linear representation"?[...]Maybe if the authors defined LRH it would be easier to understand their claims.

We thank the reviewer for raising this important question.

We did not give a single definition of the Linear Representation Hypothesis (LRH) because there is no universally agreed-upon definition. This is precisely why we wrote this paper. Park et al. (2024) describe the LRH informally as the idea that "high-level concepts are represented linearly as directions in some representation space." This description raises two important questions. The first is how we define a concept, and the second is what it means for a concept to be represented linearly. Our paper focuses on the second question, particularly how to define linear representation through the choice of representation object and equivalence.

We also do not see the ambiguity of "linear representation" and that of the LRH as two separate issues. The LRH makes a claim about the conditions a representation must satisfy to be considered linear. Different definitions of these conditions lead to different versions of the LRH. Our goal is to understand how these conditions should be specified so that the hypothesis is well-defined. This matters because a hypothesis cannot be properly tested without a clear definition. Otherwise, different studies may claim to test the LRH while actually testing different hypotheses.

Regarding linear probing, it is often used to support the LRH. For example, Park et al. (2024) themselves link linear probing to their measurement view of linear representation. In our framework, successful linear probing shows that a concept can be read out using a linear function in the dual space V∗V*. However, it does not show that the concept is represented by a displacement vector in the primal space VV, which is relevant when using the linear probe's weights to steer. It also does not justify comparing different probe weights using cosine similarity, since this requires a choice of inner product within V∗V* instead of between VV and V∗V*. These are additional claims that require further assumptions or evidence.

The related work is mixed in throughout the paper rather than in a separate section. [...] are the three common formulations the authors list in Section 2 precisely the three interpretations Park et al list in Section 1? (I don't think they are, but I think there is enough similarity to warrant a citation). The paper in its current state is a collection of ideas and observations, some of which are interesting and nicely explained, [..] I find Section 8 confusing as written. My confusion about Section 8 is detailed in Questions For Authors/Additional feedback below.

We thank the reviewer for this suggestion. We have reorganized Section 8 into a Discussion section with a dedicated Related Work paragraph. We have also moved the Platonic Representation Hypothesis discussion to the appendix to make more room.

The three formulations in Section 2 are not the same as the three views discussed by Park et al. (2024), although they are related. Park et al. provide a specific definition of linear representation, while our goal is to provide a general framework for specifying and comparing different versions of the LRH. We have made this distinction clearer in the revised manuscript.

Other modifications regarding the presentation

We have made the following changes to improve the clarity and presentation of the paper:

  • Introduction: Shortened paragraph 4 and merged it into paragraph 3. Removed mathematical symbols that had not yet been defined.
  • Section 2: Rewrote paragraph 2 for clarity.
  • Section 3.1: Added a definition of reading point (line 141).
  • Section 3.2: Clarified the definition of equivariance (line 191).
  • Section 3.5: Added a concrete example to explain (G,M,F,P)(G,M,F,P).
  • Section 4: Reorganized Corollary 4.2 to separate its statement from the explanation.
  • Section 5: Clarified that the intrinsic dimension entry in Table 1 refers to nearest-neighbor distance-based estimator.
Area_Chair_HuxQOpenReview ↗

Engaging with the authors

Dear reviewer,

the authors have posted a reply to your review. Please have a look and interact with the authors as needed.

Thank you, Your AC

Thank you to the authors for the thoughtful response to my review and the revisions. I apologize for the delay in my response.

I am happy with the revisions. Specifically, I appreciate the edits to the introduction, which substantially improved the clarity for me, the addition of the example in section 3.5, and the revision of Sections 7 and 8 of the original paper.

Overall, I still see room for improvement in the exposition, especially around Table 1, where more detail would help the readers. But I think the ideas in this paper are important and the paper makes a valuable contribution by identifying this issue with the ambiguity of linear representation and providing a framework to address it. I therefore raise my score to a 3.

P.S. I also found this exerpt from your response very helpful:

Park et al. (2024) describe the LRH informally as the idea that "high-level concepts are represented linearly as directions in some representation space.”[…]Our paper focuses on the second question, particularly how to define linear representation through the choice of representation object and equivalence.

I think this is nicely stated and I would encourage you to add it to introduction!

Reviewer_4Q9xOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The main point is that claims about linear representations need to specify which coordinate changes they should survive. The paper gives a group-action formulation and uses it to examine existing probes, steering methods and representation metrics. Validation consists of mathematical arguments and analyses of previous work, with no new experiments.

Critical strengths

The distinction between probe covectors and steering vectors is useful and relevant to how these methods are used. The framework also gives a concrete way to check whether interpretability conclusions depend on arbitrary choices of coordinates.

Critical weaknesses

The added value over standard invariance arguments seems limited. More importantly, I am not convinced the audits show that a substantive conclusion in prior work needs to change. Identifying different geometric assumptions is useful, but does not by itself show that the resulting interpretation is misleading.

Questions for authors

Which audited example most clearly changes an existing scientific conclusion? In the refusal-direction example, why would using a vector and its span undermine the shared feature claim, given the natural equivariant map between them?

2 comments

We thank the reviewer for the thoughtful comments and for engaging with the central argument of our paper. We appreciate the questions about the contribution and practical implications of our framework. Our replies follow below.

The added value over standard invariance arguments seems limited.

Our contribution is not simply to check whether a quantity is invariant under a given group. Our central argument is that how we define equivalence between representations determines what we mean by the Linear Representation Hypothesis. Different choices of symmetry group lead to different notions of representation equivalence and therefore different versions of the hypothesis.

More importantly, I am not convinced the audits show that a substantive conclusion in prior work needs to change. Identifying different geometric assumptions is useful, but does not by itself show that the resulting interpretation is misleading. Which audited example most clearly changes an existing scientific conclusion? In the refusal-direction example, why would using a vector and its span undermine the shared feature claim, given the natural equivariant map between them?

We do not claim that the experimental results of previous studies are incorrect. Rather, our framework helps clarify what their results actually establish.

A concrete example is KeyDiff (Park et al., 2025), which uses cosine similarity between attention keys to identify redundant tokens. In Section 6.3, we show that the query and key vectors can undergo transformations that preserve all attention scores but change the cosine similarities between keys. This can lead to different eviction decisions, even though the model computes exactly the same function. This does not invalidate the reported performance of KeyDiff, but it does show that key cosine similarity is not an intrinsic property of the model's computation. In Appendix C.3, we also derive an invariant alternative for standard dot-product attention.

Regarding Arditi et al. (2024), we agree that there is a natural equivariant map from a nonzero vector to its span. Our concern is not that using both objects undermines the shared feature claim. Rather, adding a vector to activations and orthogonally projecting out its span require different geometric structures. The former is affine-covariant, while the latter requires an inner product and is not equivariant under general invertible linear transformations. Both interventions can be valid, but their geometric assumptions should be distinguished when interpreting the evidence.

Area_Chair_HuxQOpenReview ↗

Engaging with the authors

Dear reviewer,

the authors have posted a reply to your review. Please have a look and interact with the authors as needed.

Thank you, Your AC

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

This paper argues that the Linear Representation Hypothesis is not a single claim but a family of hypotheses, distinguished by the notion of representation equivalence that each one implicitly assumes. Using group actions, the authors formalize what the representation object is, how it is obtained, and which property is being asserted, while also accounting for symmetries imposed by the architecture. They then use this framework to audit common representation metrics and recent interpretability analyses, showing how methods that seem to test the same representation may actually test different hypotheses.

Critical strengths
  • Interesting formalization and rigorous study clearly connecting and discussing relevant notions such as displacement, superposition, etc...
  • I think the specification principle may appear to be something recurring in different cases and not just for LRH and PRH as discussed in the paper.
Critical weaknesses
  • AFAIU, the is no experimental analysis in the paper, so its not clear the practical advantage of the theoretical machinery introduced, nor the impact of what the authors discovered about some "naive/erroneous" practice in the literature.
  • Many terms/symbols are not defined in the main text (see some examples below).
  • Instead of discussing the Platonic Representation Hypothesis, the authors could use the space to give more details about the issues they found in the literature, provide more examples, or clarifying experiments.
  • While some objections are discussed, it is missing a proper alternative views discussion about the paper position

MINOR CONCERNS

  • In the whole paper there is never clearly explained what is the Linear Representation Hypothesis. I'd start immediately introducing this notion, possibly using examples, before describing your position.
  • While trivial, Def 3.1 is just a definition (not a proof) and it should be proved/stated why "~" is an equivalence relation.
  • In the paper is never specified what it is meant exactly by function-preserving reparameterization. While used already before, only from Eq 4, it becomes more clear.
  • "Let M denote the space of mathematical objects, (...) representation coordinates, the procedure and predicate must satisfy." Many terms from these sentence are unclear, please provide a concrete example of what you mean by mathematical object, claim, statement, in this context.
  • There isn't written that the proof of Proposition 3.2 is in the appendix. Same for other proofs.
  • V* is undefined.
  • GL(V) is undefined.
Questions for authors
  1. "Let M denote the space of mathematical objects, (..) " what does it mean clearly mathematical objects? please clarify
  2. "A metric thus supplies an identification between V ∗ and V only under isometries." Why?
4 comments

We thank the reviewer for carefully reading our paper and providing constructive feedback. We especially appreciate the attention to the clarity of our definitions and arguments. Our point-by-point replies follow below.

In the whole paper there is never clearly explained what is the Linear Representation Hypothesis. I'd start immediately introducing this notion, possibly using examples, before describing your position.

This is a good question, and we would like to address it first. We did not give a single definition of the Linear Representation Hypothesis (LRH) because there is no universally agreed-upon definition. Our position is precisely that the LRH needs to be defined more carefully.

In Section 2, we already introduce several notions of linear representation through examples and explain how they differ from one another. These differences motivate the formal framework developed in the following sections, which specifies what each version of the LRH actually claims.

AFAIU, the is no experimental analysis [...] "naive/erroneous" practice in the literature. Instead [...] provide more examples, or clarifying experiments.

Our focus is not on showing that particular experimental results are incorrect. Rather, our goal is to clarify how representation claims should be defined before conducting experiments and how experimental results should be interpreted afterward. For example, KeyDiff (Park et al., 2025) uses cosine similarity between attention keys to identify redundant tokens. In Section 6.3, we show that this similarity can change even when the model's computation remains unchanged, potentially leading to different eviction decisions. We also derive an invariant alternative in Appendix C.3.

While some objections are discussed, it is missing a proper alternative views discussion about the paper position

We did not explicitly separate objections from alternative views, as the two are closely intertwined in our discussion. For example, Section 7 (original section 8) discusses the view that representation analysis should focus on the geometry of a particular trained model. This is an alternative view to our position, which we acknowledge and respond to. We have also rewritten Section 7 (original section 8) to make the objections clearer and added a Related Work discussion.

There isn't written that the proof of Proposition 3.2 is in the appendix. Same for other proofs.

The proof of Proposition 3.2(i) is provided in Appendix B, while (ii) and (iii) follow directly from the transformation properties of the inner product, so we omitted their proofs. The derivation supporting Corollary 4.2 is provided in Appendix C.

"A metric thus supplies an identification between V ∗ and V only under isometries." Why?

Consider the standard Euclidean inner product, which identifies a covector w∈V∗w \in V^* with a vector w∈Vw \in V. Under a linear transformation AA, vectors transform as v↦Avv \mapsto Av, while covectors transform as w↦A−⊤ww \mapsto A^{-\top}w. For this identification to be equivariant, we require A−⊤w=AwA^{-\top}w = Aw for all ww, which is equivalent to A⊤A=I.A^\top A = I. This is precisely the condition that AA is orthogonal, so the identification is equivariant only under isometries.

Many terms/symbols are not defined in the main text (see some examples below). V* is undefined. GL(V) is undefined. In the paper is never specified what it is meant exactly by function-preserving reparameterization. While used already before, only from Eq 4, it becomes more clear.

We have made the following changes:

  • Clarified the meaning of function-preserving reparameterization in the Introduction (line 51), before its formal definition in Section 4.
  • Removed VV and V∗V^* from the Introduction so that they only appear after being defined.
  • Defined GL(V)\mathrm{GL}(V) at its first occurrence in Section 3.3 under proposition 3.2.
  • Added a concrete example in Section 3.5 to clarify the meanings of MM, FF, and PP.
  • Moved the discussion of the Platonic Representation Hypothesis to the appendix to make more room for the main arguments.
Area_Chair_HuxQOpenReview ↗

Engaging with the authors

Dear reviewer,

the authors have posted a reply to your review. Please have a look and interact with the authors as needed.

Thank you, Your AC

I thank the authors for the detailed rebuttal and for taking the time to revise the manuscript. I appreciate the effort put into editing the paper: according to me the presentation is noticeably clearer now, and the revisions successfully address several of my initial concerns and questions.

However, after carefully reading the response and the updated draft, I still find the overall scientific contribution to be somewhat limited and below the bar for acceptance at ICLR. While the improvements are welcome, they do not fully overcome my reservations regarding the depth of the contribution. Therefore, I will maintain my borderline score, leaning towards rejection.

We thank you for carefully reading our response and the revised manuscript. We are glad that the revisions have made the paper clearer, and we sincerely appreciate your constructive feedback throughout the review process.

We would like to briefly reiterate the motivation behind our work. The Linear Representation Hypothesis is widely discussed, but what it means for a representation to be linear depends on how representation equivalence is defined. Our goal is to understand how these conditions should be specified so that the hypothesis becomes well-defined and testable. We believe this matters because a scientific hypothesis cannot be properly tested without a clear definition. Otherwise, different studies may claim to test the LRH while actually testing different hypotheses.

We hope that our work can help pave the way for more rigorous and systematic studies of representations and mechanistic interpretability. Thank you again for your time and thoughtful feedback.

Comments on the paper

1 comment

We thank the reviewers and the Area Chair for their time and constructive feedback.

We would like to briefly reiterate the motivation behind our work. To test the Linear Representation Hypothesis, we first need to specify what it means for a concept to be represented linearly. Otherwise, different studies may claim to test the same hypothesis while actually testing different ones. We found that properly defining the hypothesis so that conclusions can generalize beyond an individual model requires accounting for the relevant symmetry transformations.

To formalize these requirements, we use group actions as a natural mathematical language. Our framework makes the assumptions behind each representation claim explicit, helping clarify what each result establishes and when findings from different studies can be meaningfully compared.

Following the reviewers’ suggestions, we have clarified definitions and notation, added a concrete example, and reorganized the discussion of related work and objections.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.