acceptodds
Under review as a conference paper at ICLR 2027

Does the Route Matter? Functional Validation of Shared–Private SAEs

Abstract

Sparse autoencoders (SAEs) are commonly evaluated by reconstruction and feature descriptions, neither of which establishes that a learned decomposition matters to downstream computation. We introduce an operator-matched test of functional routing validity: correct and wrong private routes replace only the component represented by a training-only low-dimensional transform while preserving the activation's orthogonal residual. We instantiate the test with a shared–private SAE estimator based on principal-vector contrasts and group-level horseshoe shrinkage. On repository-derived prose and code, correct routing lowers held-out next-token negative log likelihood at three GPT-2 and three Pythia-70M layers. On independent text, the effect replicates at all three Pythia layers but GPT-2 intervals include zero. A 10-seed Pythia experiment removes the algebraically forced shared-subspace intersection and retains positive route effects on repository ( ) and independent text ( ); 50%-randomized routing is intermediate. The same test transfers to RxRx1 cellular embeddings, where correct HUVEC/RPE routing lowers perturbation-label loss by and improves top-1 accuracy by percentage points. Calibrated simulations return to zero under known nulls. Functional routing validity is thus measurable across modalities while remaining model- and corpus-dependent.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Sparse autoencoders, Mechanistic interpretability, Causal representation learning, Functional validation, Shared–private representations, Activation interventions, Feature routing, Grouped horseshoe priors, Bayesian sparsity, Representation disentanglement

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_9gbwOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper tests whether choosing the domain-matched private expert in a shared-private SAE preserves downstream predictions better than choosing the other expert. It estimates shared structure from principal-vector contrasts and edits only the PCA-modeled component. Experiments cover GPT-2, Pythia-70M, and cell embeddings, with additional simulations and an independent text evaluation.

Critical strengths
  1. The intervention preserves the unmodeled residual exactly. Paired routing comparisons and shared-only and randomized-route controls make the downstream effect of switching fitted experts easy to interpret.

  2. The paper reports important qualifications. The Pythia experiment without a forced subspace intersection retains positive effects, while the independent GPT-2 replication does not. The RxRx1 results provide evidence beyond language models.

Critical weaknesses
  1. The routing comparison does not separate functional ownership from ordinary reconstruction quality. Each private expert is fitted to its own domain, so the wrong expert may simply reconstruct that domain less accurately. Preserving the orthogonal residual does not control this difference inside the edited span. Sections 4 and 6 promise edit magnitudes and output divergence, but the corresponding numerical comparisons are absent. Correct routing itself adds about 2.47 NLL in the independent GPT-2 index-3 example. The routing effect supports relative downstream preservation, but the missing diagnostics leave the introduction's distinction from better fitting domain statistics unresolved.

  2. The paper explicitly disclaims estimator superiority and semantic identifiability in Section 6, but the practical value of its ownership cutoff still has limited validation. Appendix A.3 shows that horseshoe ordering is identical to raw contrast-energy ordering, leaving the rank cutoff as the consequential choice. The fixed noise and slab variances lack sensitivity evidence, and the primary design's shared rank is largely constrained by the forced intersection. The additional experiment establishes a routing effect without that intersection, but does not establish accurate ownership recovery. Appendix F mainly reports downstream responses rather than recovery of known ownership by the full estimator.

Questions for authors
  1. Can you report reconstruction error, edit RMS, KL divergence, and absolute NLL for the original, correct-route, and wrong-route conditions? Does the routing advantage persist when reconstruction damage is matched?

  2. How sensitive are selected ranks and route effects to the noise and slab variances? How does the estimator compare with a simple principal-angle cutoff or matched-rank domain PCA on the same data?

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

The paper introduces* (see first question for authors) the shared-private SAE architecture, which is intended to isolate domain-specific vs domain-common features learned by foundation models. Empirical analysis investigates how swapping private experts impacts language modeling loss.

Critical strengths
  • The architecture and evaluation procedure introduced* (see first question for authors) in this work seem to be very carefully thought out, with a variety of normalization procedures, edge cases, and potential statistical artifacts accounted for. This stands in contrast to much of today's SAE literature, where key underlying theoretical assumptions (e.g., linear representation hypothesis, superposition hypothesis, bag-of-features vs structured representations, etc) are often simply accepted at face value and not interrogated in most papers. I appreciate the authors' attention to detail in defining both the proposed architecture* (see first question to authors) and their evaluation procedure.
Critical weaknesses
  • The paper's underlying motivations, core contributions, and relationship to the rest of the SAE literature are almost completely opaque to me. That is -- my understanding of shared-private SAEs, as potentially* introduced by the authors (see first question for authors), is that they allow for comparison between two or more domain-specific SAEs in order to, e.g., study cross-domain differences in SAE activations, such as which features are shared across domains and which are not (as noted in the intro). But if I am correct that studying such questions is the motivating idea, then why would we introduce all the architectural complexity of this shared-private setup rather than simply, e.g., training a wider SAE that captures both domains? Or training two separate domain-specific SAEs, then comparing feature co-occurrence rates between them when they are fed the same samples from either domain? In the current paper, it seems that a huge amount of effort has been poured into defining and evaluating a very specific SAE architecture, without any clear indication regarding the intended utility or purpose of this architecture.
  • Beyond the unclear motivation and intended contribution, the lack of comparison with any more conventional SAEs that are not shared-private means there is no way to assess the utility of this architecture, or the associated analysis/evaluation, relative to existing methods. This further exacerbates the opaqueness problem: if it is not clear what the goals of this work are, and there is also no comparison with other methods in the literature, there is neither the intuition nor the evidence to make any substantive claims about the shared-private SAE architecture.
Questions for authors
  • The paper focuses overwhelmingly on evaluating shared-private SAEs; but this architecture is not cited or explained in the intro or related work, nor is the underlying motivation behind this architecture clarified. I am not personally familiar with any such SAE architecture, among the sea of proposed SAE architectures in today's literature; so given the lack of citations, it is not clear to me: are shared-private SAEs intended to be a contribution of this work? If so, what is their core motivation, what are the most related architectures they should be compared against (conceptually and empirically), and where is the analysis for shared-private SAEs vs such baselines? If not, then why is the source of shared-private SAEs never cited nor explained?
  • As noted in the weaknesses section: what is the core problem/research question that shared-private SAEs are intended to solve/answer? What are the implications of the empirical study toward solving/answering this problem/RQ?
  • What is "s" in equation 1? It is never defined, and it is not clear to me why it is included at all, how it is computed, or what role it is intended to serve.
    • Likewise for the "anchoring (strength)" and "incoherence penalty" of page 5 -- how are these values selected, and how specifically are they used to "anchor" or "penalize"? (In this case, a simple equation indicating the specific computational role of both constants, and the authors' process for selecting these specific values, would suffice.)
1 comment
Reviewer_BKUnOpenReview ↗

Response to Authors

I would like to thank the authors for their detailed response. I have reviewed their responses to all three reviews, as well as the general response.

It is helpful to know that "shared-private SAEs" are indeed intended to be a contribution of this work, and I appreciate the results of some of the suggested followup experiments. However, the main problem remains: what is the purpose of this study? The authors responded that:

The question is whether, when a decomposition assigns a direction to one domain, the model's computation uses that assignment beyond the better fit of a domain-matched reconstruction.

And in responding to reviewer mVfc, they state that:

We rewrote the introduction around one question: when a decomposition assigns a direction to one domain, does the model's computation use that assignment beyond the better fit of a domain-matched reconstruction?

Why does this matter? It seems that the question is inherently restricted to a new class of SAEs that the authors introduce in this work (shared-private SAEs) -- in effect, the goal seems to be to define a new architecture and answer a specific question about it; but neither the architecture nor the question has clear motivation. The architecture is designed around the question about private vs shared vectors, and the question of private vs shared vectors is explored with respect to the only architecture that (to my knowledge) has this specific feature (decomposition into shared vs private domains). The motivation is circular: the architecture only matters because of the question, and the question only matters because of the architecture.

While I appreciate the authors' effort to address some technical questions and run followup experiments, the core weakness remains, so I don't see any reason to change my rating.

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

This paper investigates whether the route in a shared–private sparse autoencoder, i.e., which domain-specific private expert is applied, actually affects downstream computation rather than merely improving reconstruction. The authors propose an estimator that decides which directions are shared versus private from the principal angles between two domain-specific SAE spans, and a functional test that swaps only the modeled component of an activation while preserving the residual, comparing next-token NLL under correct versus wrong routing. On GPT-2 and Pythia-70M (prose versus code) and RxRx1 cell embeddings, correct routing yields lower loss than wrong routing, though the GPT-2 result does not replicate on an independent corpus.

Critical strengths
  • The paper presents a reasonable configuration in which a shared dictionary and two domain specific private experts can either cooperate or interfere with each other.
Critical weaknesses
  • The presentation is a major concern. The argument is hard to follow and the contribution is difficult to locate.
  • Generalization is untested. Experiments are confined to very small models (GPT-2 and Pythia-70M), and the “SAE” is a width 20 dictionary on 32 PCA coordinates rather than an overcomplete dictionary.
  • The motivation is underdeveloped. The paper argues that reconstruction scores do not guarantee functional relevance, but never establishes why shared–private SAEs with observed domain labels matter or what follows once a route is validated.
Questions for authors
  • Why is the prose vs. code split on GPT-2 and Pythia-70M representative? GPT-2 saw essentially no code in pretraining and Pythia-70M only a small amount. Notably, Pythia replicates while GPT-2 does not, suggesting the result may hinge on whether the model has any code representation to route to.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.