acceptodds
Under review as a conference paper at ICLR 2027

FlowSAE: Non-Invasive Sparse Autoencoding with Lifted Invertible Flows

Abstract

Sparse autoencoders (SAEs) are used both to interpret neural network representations and to intervene on them, but a standard SAE's decoder is only an approximate inverse of its encoder, so reconstruction error contaminates any causal effect an intervention is meant to isolate. Prior work compensates for this after the fact, by carrying the error as an external residual channel or by normalising away its damage, rather than removing it. We introduce FlowSAE, to our knowledge the first nonlinear, multi-layer flow-based SAE whose decoder is the exact inverse of its encoder. FlowSAE lifts the activation into a larger space, applies an invertible flow, sparsifies, and inverts. Top- sparsification is the only lossy step. When no feature is removed, the input is recovered exactly for any learned parameters, not only after training. On ImageNet ViT features this holds to numerical precision. It also removes a problem we find in classical SAEs, where uncorrected single-feature ablations can invert the true attribution ranking. At the deployed sparsity, however, exactness does not carry over. The nonlinear flow reconstructs better than a linear exact control, but its feature effects become non-additive, so the standard error-term correction, which is exact for any linear decoder, no longer recovers them. Exact reconstruction adds no asymptotic cost per forward pass, only a constant factor. Making sparse interventions trustworthy is a separate problem.

Primary area:
interpretability and explainable AI
Keywords:
Sparse Auto Encoders, Invertible Neural Networks, Normalizing Flows

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_gPSpOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

In their paper, the authors propose FlowSAE, in which the decoder is the exact inverse of the encoder. The activation is lifted into a higher-dimensional space, passed through an invertible flow, sparsified via TopK, and then inverted back, so at k = n the reconstruction is exact for any parameters. The motivation of the paper is that the reconstruction error of classical SAEs contaminates the effects of interventions. However, both the exactness guarantee and the main positive results apply only to the k = n case, which is not interesting in practice, since the whole point of SAEs is k << n. At the realistic sparsity level (k = 32), the reconstruction offset of FlowSAE is the same as that of a classical SAE, the feature effects become non-additive, and there is a lot of overhead from computational and architectural complexity (unlike vanilla SAEs).

Critical strengths
  1. The paper is very honest, and the claims are well calibrated. The authors added a linear control themselves and openly report negative results.
  2. Good reproducibility. the code is attached, and the experimental details are described thoroughly.
Critical weaknesses
  1. The motivation of the paper only works for the k = n case, which is not interesting in practice, since the whole point of SAEs is k << n. But even at k = n, the problem is solved trivially. Exact reconstruction can be obtained without any flow, simply by taking an encoder with orthonormal columns and a transposed decoder. This is essentially the authors own linear control.

  2. At the realistic k = 32, FlowSAE does not solve the stated problem. Its reconstruction error is the same as that of a classical SAE. At the same time, the feature effects become non-additive, so even after correction the attributions are worse than for any SAE with a linear decoder. The same linear control at k = 32 is both additive and more stable. In the end, the nonlinear flow only gives a gain in FVU, and even then only up to the level of a regular SAE.

  3. In addition, the paper does not show at all what changes from an interpretability perspective. Interpretability is the main goal of SAEs, while reconstruction is essentially just a proxy and a loss. The paper has neither autointerp nor other standard metrics (SAEBench, SynthSAEBench). Perhaps the method would actually show improvements on them, but the paper only mentions that sparse probing performs roughly on par with existing SAE baselines.

I am open to discussion. If the authors can show that I have misunderstood something or that my concerns are wrong, I will be happy to respond and reconsider my evaluation.

Questions for authors

Is there a practical interpretability scenario at realistic sparsity (k << n) where FlowSAE outperforms standard SAEs? In particular, do you have results on standard interpretability metrics (e.g., autointerp, SAEBench, SynthSAEBench) at realistic budgets (e.g., k = 32 to 128), compared to classical SAE baselines? Positive results here would substantially change my assessment.

4 comments
AuthorsOpenReview ↗

Response to Reviewer gPSp: results below k = n and new interpretability evidence

Thank you for the detailed review and for your openness to discussion. We would like to correct one central point, and then answer each weakness.

"The exactness guarantee and the main positive results apply only to k = n." Respectfully, the positive results are not confined to k = n. Three are at k < n:

  1. Against the exact linear construction. At k = 32, FlowSAE reduces FVU 1.8× (0.216 vs 0.397) under identical training.
  2. Across budgets. One multi-k model's error falls monotonically to exactly zero. A single-k specialist is U-shaped.
  3. At large k. For k ≥ 256 on ImageNet, FlowSAE is below every classical SAE we trained (0.020 vs 0.090 at k = 1024, Fig. 2).

The k = n result also says something new about classical SAEs. Their uncorrected attribution rankings can be inverted (ρ = −0.83), not just noisy.

W1: exact reconstruction is achievable with a linear map. Agreed, and we built that control ourselves to isolate exactly this question. It is exact but reconstructs poorly (FVU 0.397 at k = 32). Classical SAEs reconstruct well but are not exact. FlowSAE is the only model in the comparison with both properties. That combination, not exactness alone, is the contribution, and the revised introduction says so up front.

W2: at k = 32 the offset matches a classical SAE, and effects are non-additive. No SAE is exact at k < n, and FlowSAE is not worse there: its offset is 0.585 vs 0.577. Non-additivity is inherent to any nonlinear decoder, and we quantify it directly (ρ = 0.56 after correction, Table 2). We also give a protocol that sidesteps it: select features at the deployed budget, then intervene in the exact k = n representation, where effects carry no reconstruction error. Non-additivity can also be reduced in training. The revision adds an explicit additivity penalty (App. H), which at λ = 0.1 lowers the additivity error 7× (1.41 to 0.19) at FVU 0.220, a trade-off the linear control (FVU 0.397) does not offer, while the k = n identity stays exact.

Overhead. A forward pass costs O(LNnd), plus an O(nd² + d³) setup per layer that does not grow with the number of samples (App. D). Per sample this is the order of a classical SAE times the depth L, so the overhead is a constant factor. We measure it openly (App. C, D) and do not hide it. The method trains and evaluates at d = 768 on Pythia-160M with the same pipeline as at d = 384.

W3: interpretability metrics at k = 32 to 128. The paper reports more than sparse probing. Table 5 of the submission (Table 6 in the revision) gives SAEBench core across L0 = 19 to 630. The revision adds SAEBench's published TopK SAEs at matched L0, on the same model, layer and width:

L0FlowSAE (one model)Published TopK (one model per L0)
380.9370.959
760.9540.974
1550.9680.986

FlowSAE stays within about 2 points of models that were each trained at that budget. At large budgets, most of that gap is the price of covering every budget with one model: our FlowSAE specialist trained at k = 458 reaches 0.997, against 0.984 for the multi-k model and 0.999 for the best published SAE. On concept detection, sparse probing at L0 = 32 (0.764 ± 0.013, three seeds) is on par with standard SAEs and above full-rank PCA.

To test directly whether individual features correspond to concepts, the revision adds a class-semantics evaluation on ImageNet (new App. M, Fig. 4 and Table 8). For 64 random latents per model at k = 32, FlowSAE latents are as class-selective as those of a classical SAE trained identically (purity of the top-50 images 0.31 vs 0.34, chance 0.03). The five classes a FlowSAE latent promotes in the frozen classifier agree with the labels of the images that activate it almost twice as often (0.43 vs 0.24, p ≈ 1e-5, chance 0.005), and they are more semantically coherent (WordNet similarity 0.65 vs 0.54, random sets 0.45). For example, a latent that fires on stoves promotes stove, wok, frying pan, coffeepot and microwave.

We do not report SAEBench autointerp. In its released form it scores feature explanations with gpt-4o-mini through a paid API, which is not an option for this research.

Practical scenario at k ≪ n.

  • Causal intervention studies. Identify features at k = 32, then intervene exactly at k = n. No error term needs to be modelled, and rankings are preserved.
  • Reconstruction-critical uses at moderate k. Here the classical SAEs we trained plateau and FlowSAE keeps improving.

We hope this shows that the contribution extends beyond k = n, and we would be grateful if you would reconsider your score.

AuthorsOpenReview ↗

Response to Reviewer gPSp (part 2): concept detection and autointerp across budgets

Following up on W3 (interpretability metrics at k = 32 to 128), we have extended SAEBench sparse probing for the multi-k FlowSAE from L0 = 32 to the published budgets L0 = 38, 76 and 155, with three seeds each. We compare against SAEBench's published SAEs on the same model, layer and width. The metric is the accuracy of a probe on the single best latent, averaged over SAEBench's eight datasets.

L0FlowSAE (one model, 3 seeds)Published TopKPublished BatchTopK
320.764 ± 0.013––
380.766 ± 0.0150.7340.739
760.783 ± 0.0110.7460.748
1550.787 ± 0.0130.7920.773
  • At L0 = 38 and 76, FlowSAE features detect concepts better than published TopK and BatchTopK, by 2.7 to 3.7 points, which is 2 to 3.5 seed standard deviations.
  • At L0 = 155 it is level with them, 0.5 points below TopK and 1.4 points above BatchTopK.
  • Against all seven published architectures at the nearest budget, FlowSAE is above five at L0 = 38 and 155 and above six at L0 = 76. Matryoshka-BatchTopK is the only architecture ahead of it at both larger budgets.
  • Accuracy rises monotonically with the budget, and all four points come from one model per seed. Each published point is a separate model trained at that budget.

So in the range your W3 asks about, FlowSAE's features are not only on par with classical SAEs. At the sparser budgets they are more concept-aligned than the most widely used architectures. The published points are single runs, so we report our seed spread beside every number. They are now Table 6b of the revised PDF (App. L, marked in red), next to the loss-recovered comparison in Table 6a.

Autointerp, as an additional experiment. Our first response noted that SAEBench autointerp relies on a paid API. We have since run it with a free open-weight judge (Qwen2.5-32B-Instruct). A language model writes a one-sentence explanation of each latent and then predicts where it fires. The judge reproduces the published gpt-4o-mini ranking of the TopK SAEs, so the comparison is reliable.

kFlowSAE (one model, 3 seeds)Published TopK (same judge)
200.843 ± 0.0030.850
400.816 ± 0.0040.841
800.792 ± 0.0060.827
1600.778 ± 0.0070.814

FlowSAE's latents score well above chance (0.714) at every budget, are stable across seeds, and all four points come from one model. They are close to TopK at the sparsest budget and a few points lower at larger ones.

The two experiments are best read together. Probing tests whether a latent detects a concept. Autointerp tests whether everything a latent responds to fits into one sentence. FlowSAE latents pass the first test better than TopK, and the second slightly less well as the budget grows. This is consistent with latents that track a concept reliably while also responding to a wider range of contexts. At the small budgets where features are read, FlowSAE is therefore at least as good a concept detector as TopK, and one model covers every budget. The autointerp results are Table 6c of the revised PDF.

Together with reconstruction within about 2 points of models trained for each budget, and the class-semantics evaluation of App. M, this answers the interpretability metrics your review asked for, with a clear positive result on concept detection below k = n. We would be grateful if you would take it into account in your assessment.

Thank you for the detailed responses and additional experiments. I agree that my original statement about positive results being limited to k = n was too broad. The new sparse-probing and ImageNet results address part of my concerns, and I am raising my score to 2.

My main remaining concern is the balance between complexity and demonstrated benefit. FlowSAE introduces a substantially more involved architecture, while at realistic sparsity its reconstruction is comparable to standard SAEs and the interpretability results are mixed: sparse probing improves at some budgets, but autointerp scores are lower than TopK. I appreciate the exact reconstruction guarantee and the careful analysis of its limitations, but the results do not yet establish a sufficiently clear practical advantage to justify this added complexity.

AuthorsOpenReview ↗

Response to Reviewer gPSp

Thank you for engaging with our responses and for raising your score. We agree the case should rest on concrete use, so we state it precisely.

1. Interventions without a residual pathway. Circuit and attribution analyses with classical SAEs must keep the reconstruction error as a separate node or add it back after intervening. Without it, single-feature attributions can be inverted (Spearman ρ = −0.83 and −0.62 for two of the classical SAEs we test). FlowSAE removes this by construction: features are selected at k = 32 and interventions are made in the exact k = n representation, so every effect is carried by features. The linear control is also exact, but its features at k = 32 reconstruct 1.8× worse (FVU 0.397 vs 0.216). FlowSAE is the only model in our comparison that offers both.

2. One training run for every budget. SAEBench trains a separate SAE for each L0. One FlowSAE model stays within about 2 points of loss recovered of these per-budget models from L0 = 38 to 155, detects concepts better at L0 = 38 and 76, and is exact at k = n. When the budget an analysis needs is not known in advance, this replaces a sweep of trainings with one.

On autointerp, we agree, and we report it as lower than TopK at larger budgets. At the sparsest budget, where features are read, the two are close (0.843 vs 0.850 at k = 20).

On complexity, the cost is a constant factor per layer (App. C, D). Training uses a standard SAE objective and pipeline, and the base model is not modified.

FlowSAE is not a replacement for classical SAEs in every use, and the revised paper says so. Its advantage is specific: exact, residual-free interventions and one model for all budgets, with reconstruction and concept detection comparable to the best per-budget SAEs. We hope this clarifies the practical case.

Reviewer_Uwc1OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces flowsae which uses lifted invertible flows to make the decoder the exact inverse of the encoder. this design gets rid of the reconstruction error that confounds causal interventions when the full capacity is used. sparsification via topk is the only non-invertible step. exactness holds at full capacity but authors show that at deployed sparsity levels the nonlinear nature of the flow makes feature effects non-additive. this complicates traditional attribution.

Critical strengths

Architectural choice to use the matrix exponential over the cayley transform to maintain exact invertibility during training without boundary issues is elegant. authors are honest about the downstream consequences of their design. acknowledging that the nonlinear decoder breaks the additivity of features adds nuance to the paper.

Critical weaknesses

The non-additivity of features at the deployed sparsity budget breaks the standard paradigm of using saes to extract independent concepts. if a features decoding depends heavily on all other active features it becomes difficult to assign it a stable semantic meaning.

non-additivity of features at the deployed sparsity budget breaks the standard paradigm of using saes to extract independent concepts. if a features decoding depends heavily on all other active features it becomes difficult to assign it a stable semantic meaning.

Questions for authors

Given the non-additivity of the latent space do you think these flow-based features can still be mapped to independent semantic concepts? did you explore any regularization strategies during training that might explicitly penalize non-additivity or encourage the local jacobian to be more stable across different active supports?

2 comments
AuthorsOpenReview ↗

Response to Reviewer Uwc1: stable feature meaning and an additivity regulariser

Thank you for the thoughtful review and for highlighting the exponential-map design. Your questions led to new experiments, which we report below.

On soundness. We respectfully ask you to reconsider the soundness score. Non-additivity is a property of the method that we measure and report ourselves, not an error in a claim. Each central claim in the paper is checked directly:

  • exactness on untrained models (max error 1.6e-13) and after training (logit MSE 2e-23);
  • the composition result over 6 and 3 seeds with Welch tests;
  • the k = 32 attribution over three seeds.

We would value knowing any specific claim you consider unsupported.

Stable semantic meaning. The evidence supports stable meaning at the level that matters for interpretation, the feature direction.

  • Each coordinate names a consistent direction. The local direction Δ_i = ∂ decode / ∂ z_i has mean cosine similarity 0.96 to 0.99 across different points of latent space, including codes with different active supports (App. G).
  • What depends on context is the contribution's size. At k = 32 the summed per-coordinate contributions point the right way (cosine 0.89) but are not additive. We quantify this rather than assume it away (App. G, Table 2).

So a FlowSAE feature has a fixed semantic direction whose effect is modulated by co-active features. We make this explicit in the revised Discussion (Sec. 5), and we flag that additive tools (steering, SCR, TPP) need adaptation.

Q1: do flow features map to independent semantic concepts? On Pythia-160M, sparse probing is on par with standard SAEs (0.764 ± 0.013 at L0 = 32 over three seeds, within the 0.729 to 0.781 of published SAEs at L0 ≤ 45, and above full-rank PCA). Concept information is preserved in individual features. We have also tested this directly on ImageNet, where class labels provide a ground truth (new App. M). FlowSAE latents are as class-selective as those of a classical SAE, and the classes a latent promotes agree with the images that activate it almost twice as often (0.43 vs 0.24). This holds in context as well (0.39), where non-additivity could have broken it, and the promoted classes are semantically coherent (WordNet similarity 0.65 vs 0.54, random sets 0.45). Individual FlowSAE features therefore map to recognisable concepts, while the size of their effect depends on the co-active features.

Q2: regularisation for additivity or Jacobian stability. We had explored the architectural route, and we now add the explicit penalty you suggest.

  • Architectural route. The piecewise-linear activation makes the decoder's Jacobian piecewise constant across supports, which is the source of non-additivity. A smooth activation (sinh-arcsinh) stabilises the code. Over three seeds on ImageNet, re-encoding support agreement rises from 0.889 to 0.948 and feature directions become stable. Its effect on the additivity error itself is not significant over three seeds (1.30 ± 0.53 vs 1.52 ± 0.13). On Pythia it costs reconstruction and features (loss recovered 0.843 ± 0.056 vs 0.933, 30% of latents alive vs 85%). Smoothness alone does not fix additivity, which motivates the explicit penalty.
  • Explicit penalty. We penalise the squared relative gap between the decode of the k-sparse code and the sum of its single-coordinate decodes. The penalty is exactly zero for any linear decoder, so it targets only the nonlinearity and leaves the k = n identity intact. At λ = 0.1, the additivity error falls from 1.41 to 0.19 (7×) while FVU at k = 32 changes from 0.216 to 0.220, 52% of latents stay alive, and the k = n identity is intact (FVU 6e-24 in float64). Single-feature effects therefore add up at almost no cost in reconstruction, the trade-off that the linear control (FVU 0.397) does not offer. The penalty also steadies the local Jacobian. The summed single-coordinate contributions align with the joint decode at cosine 0.98 (0.89 without it), and the worst-case cosine of a coordinate's direction across its activation range rises from 0.75 to 0.89. Re-encoding support does not become more stable (0.875 vs 0.890), so additivity and support stability are separate properties.

Revisions prompted by your review. Following your suggestions, we have added a new appendix (App. H, marked in blue in the revised PDF) that presents both the architectural and the regularisation approach to additivity. Alongside the additivity and stability results, we now also report the fraction of active latents (App. H, App. L and the Limitations section).

We hope these results address your concerns, and we would be grateful if you would reconsider your assessment.

AuthorsOpenReview ↗

Response to Reviewer Uwc1 (part 2): do features map to concepts across budgets?

Following up on Q1 (whether flow features map to independent semantic concepts), we have extended the language-model evidence. In our first response, sparse probing was reported only at L0 = 32. We now run SAEBench sparse probing for the multi-k FlowSAE at the published budgets L0 = 38, 76 and 155, with three seeds each, and compare it with SAEBench's published SAEs on the same model, layer and width. The metric is the accuracy of a probe on the single best latent, averaged over SAEBench's eight datasets, so it asks directly whether one feature on its own carries a concept.

L0FlowSAE (one model, 3 seeds)Published TopKPublished BatchTopK
320.764 ± 0.013––
380.766 ± 0.0150.7340.739
760.783 ± 0.0110.7460.748
1550.787 ± 0.0130.7920.773
  • At L0 = 38 and 76, a single FlowSAE feature detects concepts better than a single feature of published TopK or BatchTopK, by 2.7 to 3.7 points, which is 2 to 3.5 seed standard deviations.
  • At L0 = 155 it is level with them, 0.5 points below TopK and 1.4 points above BatchTopK.
  • Against all seven published architectures at the nearest budget, FlowSAE is above five at L0 = 38 and 155 and above six at L0 = 76. Matryoshka-BatchTopK is the only architecture ahead of it at both larger budgets.
  • The result is stable across budgets. Accuracy rises monotonically with L0, and all four points come from one model per seed.

This complements the picture in our first response. Probing reads one feature at a time, so it does not depend on effects being additive. Even though a feature's contribution is modulated by the co-active features, the feature itself is a concept detector at least as good as those of the most widely used linear SAEs, and better at the sparser budgets. As an additional experiment we also ran SAEBench's automated interpretability score (autointerp, Table 6c). A language model writes a one-sentence explanation of each latent and then predicts where it fires. We used an open-weight judge (Qwen2.5-32B-Instruct), which reproduces the published ranking of the TopK SAEs. FlowSAE's latents score well above chance at every budget and are stable across seeds. They are close to TopK at the sparsest budget and a few points lower at larger ones.

The two experiments are best read together. Probing tests whether a latent detects a concept. Autointerp tests whether everything a latent responds to fits into one sentence. FlowSAE latents pass the first test better than TopK, and the second slightly less well as the budget grows. This is consistent with latents that track a concept reliably while also responding to a wider range of contexts, in line with the context-dependent effects discussed in our first response. At the small budgets where features are read, FlowSAE latents are therefore at least as good concept detectors as TopK latents.

Together with the ImageNet class-semantics evaluation (App. M), this answers Q1 positively on both substrates. The new results are Tables 6b and 6c of the revised PDF (App. L, marked in red).

We hope this further addresses your question, and we would be grateful if you would take it into account in your assessment.

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

In this paper the authors try to address a problem when using SAE to interpret neural network representations and to intervene on them. The problem is that a standard SAE’s decoder is only an approximation of its encoder's inverse, so reconstruction error will contaminate any causal effect an intervention is expected to isolate. The authors propose a framework called FlowSAE, which the authors claim to be the first nonlinear multi-layer flow-based framework whose decoder is the exact inverse of its encoder. The authors first lift the features to higher dimension, then use multiple orthogonal layers with Cayley reparameterization for projections, and then do truncation and reconstruct. The authors show by experiments that the reconstruction is accurate when using full-rank and have a good performance when k=32.

Critical strengths
  • The construction is simple and elegant by using lift, invertible flows and truncation etc. The authors also verify the correctness (error at 10^-23 level) rather than just providing some architectural guarantee.
  • The method (implementation) also carefully designed. The low-rank matrix parameterization is a clever idea and achieves good trade-off between memory and accuracy.
  • The overall presentation does not oversell the results. The attribution experiments and the limitation section reported negative results.
Critical weaknesses
  • The only thing I am concerned about is that the authors show very good results when using large kk (e.g. k >= 256) but actually used k=32k=32 most of the time. Not sure why this is the case as I don't work in this field.
Questions for authors

See above.

2 comments
AuthorsOpenReview ↗

Response to Reviewer gjtV: why k = 32, and the exponential-map chart

Thank you for the positive assessment, and for recognising the design of the low-rank parametrisation and the paper's careful verification. Your question pointed to something the paper should explain earlier. We have fixed that.

Why k = 32, when results are strongest at large k? The two regimes serve different purposes, and FlowSAE is designed to serve both from one model.

  • k = 32 is where features are interpreted. An SAE is useful for interpretation because each input activates only a few features, few enough for a person to inspect them and to intervene on one at a time. Budgets of roughly 16 to 64 are standard in the SAE literature, so our attribution and composition experiments use k = 32. There FlowSAE matches classical SAEs (FVU 0.210 vs 0.211 to 0.241) and reconstructs 1.8× better than the exact linear construction.
  • Large k is where the flow's advantage is clearest. FlowSAE's error keeps falling with k, to 0.020 at k = 1024 and exactly zero at k = n. The classical SAEs we trained plateau (Fig. 2). This matters whenever a sparse code must preserve the activation faithfully.

Because FlowSAE is trained across all budgets at once, a practitioner can read features at k = 32 and use the same model at larger k, or exactly at k = n, when fidelity matters. The revised introduction (Sec. 1, in blue) states this up front.

A small clarification. The orthogonal layers use the matrix exponential, Q = exp(B). The Cayley transform appears only as an ablation (App. C). We show that it silently loses exact invertibility during training, around epoch 55, which motivates the exponential map.

Changes in response to all reviews. The revision places SAEBench's published TopK SAEs beside our language-model results (App. L), adds an additivity regulariser (App. H), and adds an ImageNet test of whether individual features correspond to concepts (App. M).

We hope the clarifications strengthen your confidence in the work.

AuthorsOpenReview ↗

Response to Reviewer gjtV (part 2): one model for every budget

Following up on your question about k = 32 versus large k, we have made the practical consequence explicit in the revised PDF (Contribution 3, Sec. 4.4 and the Conclusion, marked in red).

One FlowSAE model serves every budget. A classical SAE is trained for one sparsity budget and is used at that budget. Covering several budgets takes one model per budget. FlowSAE is trained across all budgets at once, and its error falls monotonically with k, to exactly zero at k = n. So the same model is read at k = 32 for interpretation and used at larger k, or at k = n, when fidelity matters. This is why the paper reports both regimes. They are not two settings to choose between, but two uses of one model.

New evidence at the small budgets where features are read. Since our first response we have tested whether FlowSAE's features at these budgets carry concepts, on the language model, with three seeds:

  • Concept detection (SAEBench sparse probing, Table 6b). A single FlowSAE latent detects labelled concepts better than a single latent of the published TopK and BatchTopK SAEs at L0 = 38 and 76, by about 3 points, and is level with them at L0 = 155. Each published point is a separate model trained at that budget. All FlowSAE points come from one model.
  • Automated interpretability (Table 6c). A language-model judge explains each latent in one sentence and predicts where it fires. FlowSAE's latents score well above chance at every budget. They are close to TopK at the sparsest budget and a few points lower at larger ones, which is consistent with latents that detect a concept reliably while also responding to a wider range of contexts.

Together with the ImageNet class-semantics test (App. M), this shows that the k = 32 operating point is not a compromise. At the budgets where features are read, FlowSAE's features are at least as concept-aligned as those of widely used SAEs, and the same model remains exact at k = n.

Thank you again for your positive assessment. We hope these additions further strengthen your confidence in the work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.