acceptodds
Under review as a conference paper at ICLR 2027

Rational Sparse Autoencoder

Abstract

Sparse autoencoders (SAEs) are standard tools for mechanistic interpretability, but current SAE families are constrained by fixed encoder nonlinearities such as ReLU, JumpReLU, and TopK. This hard-codes a particular sparsity mechanism into the model and can distort the reconstruction-versus-sparsity trade-off. We introduce the *Rational Sparse Autoencoder* (RSAE), which replaces the fixed encoder activation with a trainable rational function. Our idea is developed based on a theoretical reconstruction advantage for RSAE at matched expected sparsity, with guaranteed non-inferiority to ReLU and, under stated assumptions, provable improvements over JumpReLU and TopK. The model is then implemented through a two-stage pipeline: an initialisation procedure that copies the pre-trained baseline SAE weights, and calibrates the scale parameters along with the rational coefficients; followed by a fine-tuning step under a sparsity-regularised reconstruction objective that holds at the teacher's value. Consistent with this theory, on residual-stream activations of three open-weight language models and across all three baseline activation families, the RSAE *improves* reconstruction fidelity after the fine-tuning step at the same level of sparsity. These gains are consistent across host language models, across baseline activation families, and across the full range of baseline sparsity we tested, while the upgrade itself adds only 20 scalar parameters per autoencoder and one training run takes between about 4 and 25 minutes on a single consumer GPU.

Primary area:
interpretability and explainable AI
Keywords:
Sparse Autoencoder, Mechanistic Interpretability, Rational Neural Network

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_9Gh3OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose the residual SAE (RSAE), a new activation function for sparse autoencoders that aims to outperform the activation functions currently used in SAEs.

Critical strengths

I find the idea interesting and promising. I am not familiar with every recently published SAE variant, so I will not make claims about its novelty. The theoretical results provide some support for the approach.

Critical weaknesses
  • Clarity. I found the paper hard to follow. For example (and this is not the only case), Section 3 opens directly with the exact equation for RSAE, without explaining why it was constructed this way or what each part is meant to do. Theorem 1 then states that “for p ≥ 1 and q ≥ 0, the infimum of RSAE reconstruction risk is no greater than that of ReLU at any feasible expected...”. It took me several minutes to find where p and q are defined, and more time to understand why this result matters. It may well be important for the paper, but the text does not explain its significance. Why should the reader care about this infimum, and why is a theoretical comparison with ReLU the right one to make? What does this result tell us in practice? Theorem 2 relies on material from the appendix, so the main text is not self-contained, and it introduces terms that have not been defined before. Similar issues appear throughout the rest of the paper. I would encourage the authors to give some motivation before each formal statement and to define all notation in the main text.
  • Experiments. This is my main concern and the one that determines my score. The evaluation is limited to the SAEBench core metrics, such as cross-entropy loss and the number of live features. From these results alone, I cannot tell whether the method is useful in practice. SAEs are a practical tool for interpretability, and practitioners use them for downstream tasks such as probing. SAEBench already includes benchmarks for these tasks, but they are not reported. As it stands, the results show a gain on one metric, and it is unclear whether this gain leads to meaningful improvements on the tasks that actually matter.
  • Significance. Following from the previous point, the paper currently reads as one more SAE variant among many. A large number of SAE variants have been proposed, and only a small fraction of them hold up under practical evaluation or are worth adopting. A better cross-entropy loss at a given sparsity level is not enough, on its own, to convince me that this variant will have a significant impact. Based on the current results, I am not convinced that it will.
  • Inaccurate references. Some citations appear to misdescribe the cited work. For example, the paper states that “data-free SAEs (Laptev et al., 2025) fit dictionaries directly from model weights without streaming activations.” This is not what the cited paper does. It applies more sophisticated methods to standard SAEs trained on activations. I suspect there may be other similar cases in the paper.
  • Some figures are not very informative. For example, Figure 2 shows that there are configurations where RSAE gives large improvements. I do not doubt this, but such configurations can likely be found for almost any method, and selected comparisons can show nearly any size of improvement, even “+100%” in some scenarios. Without a broader view, such as results across all configurations, it is hard to draw conclusions from this figure.
Questions for authors

See weaknesses

4 comments
AuthorsOpenReview ↗

Response to Reviewer 9Gh3 (part 1)

We thank the reviewer for the time devoted to our submission and for the candid feedback. We address each of the reviewer's concerns below.

Clarity of Section 3. We respectfully disagree that Section 3 leaves its construction unexplained. Its layout is deliberate, and the items the reviewer found missing are in the submitted text.

  • Where the motivation is. The paper follows the usual order of motivation, background, method and guarantees, algorithm, experiments. The reason for replacing a fixed activation by a trainable one is given in Section 1 (lines 36–48: the shrinkage and dead latents of ℓ1\ell_1-regularised ReLU, the hard selection of TopK, the surrogate gradient of JumpReLU). The rational family and the reason for its pole-free denominator are introduced in Section 2 (lines 127–143). Section 3 can therefore open with the definition.
  • What each part of Eq. (3) does. This is explained in the sentences that directly follow the equation (lines 155–164): the clamp [⋅]+[\cdot]_{+}; the denominator, which satisfies Q≥1Q\ge 1 so that the rational has no poles and can be trained without constraints; the trainable scales Cin,CoutC_{\rm in}, C_{\rm out}, shared across all features; and the fact that the support of the code is decided by the sign of PP, with no threshold or gating parameter.
  • Where pp and qq are defined. In the sentence immediately after Eq. (3) (line 155), fourteen lines before Theorem 1.
  • Why the infimum, and why ReLU. The two sentences before Theorem 1 (lines 166–167) state what is compared, the reconstruction risk at the same expected number of active latents, and why ReLU comes first: the RSAE family contains every ReLU SAE. The practical meaning is given where it is used: with the identity initialisation the RSAE coincides with its teacher at the start of fine-tuning (Section 4, lines 214–217), and Section 5 (line 375) relates the theorem to the measured gains.
  • Theorem 2 and the appendix. Both assumptions are stated in words in the main text (lines 175–184), and the paragraph after the theorem (lines 196–198) explains its conclusion and the regime it identifies. Only the quantitative constants and the proofs are in the appendix, which is the common practice for a theoretical result under the page limit.

We nevertheless want the section to be easy to read for someone who goes straight to it, so we will adopt the following clarification:

  • Notation repeated where it is used. The definition of pp and qq (numerator and denominator degrees) will be repeated in the sentence that precedes Theorem 1.
  • Why the infimum, and why ReLU. We will add a short paragraph saying the following: an SAE practitioner chooses an activation family and then trains; the best reconstruction reachable inside the family at a given sparsity budget kk (the infimum of the risk) is therefore the right object to compare families. ReLU is the natural first comparison because it is the most widely released SAE family and because the comparison is exact: the RSAE family contains every ReLU SAE (identity coefficients), so nothing can be lost by the replacement and fine-tuning starts exactly at the released model. In practice this is what makes the upgrade safe: step 0 of our fine-tune reproduces the teacher's codes and reconstruction exactly.
  • Self-contained Theorem 2. The main text will state the two assumptions in plain words together with what they mean for an SAE (bounded feature energy; the residual of the teacher still carries a component predictable from the pre-activation, i.e. the teacher's fixed activation leaves systematic error), and will explain the conclusion in one sentence: the calibration gain grows linearly in the number of active latents kk, while the price of replacing a top-kk support by a threshold support grows only like k\sqrt{k}, so the rational wins once kk is large enough. This is exactly the trend of Figure 2.

Correction of references. We thank the reviewer for the careful reading of the related work. Laptev et al. (2025) analyse how features of SAEs trained on activations evolve across layers, using a data-free (decoder cosine-similarity) matching between layers. Our sentence describes this work as fitting dictionaries directly from model weights; we will reword it to describe an analysis and steering method built on standard SAEs. The sentence is part of a short list of work orthogonal to ours (Section 2, lines 108–109), and no claim or result of the paper depends on it. We are also re-checking the other citations in Section 2 against the cited papers and will correct any description we find imprecise.

AuthorsOpenReview ↗

Response to Reviewer 9Gh3 (part 2)

Probing results beyond the core metrics. We would like to point out that a probing result is in the submission: Section 5 ("Sanity check on feature quality", lines 461–464) and Appendix B.3 / Table 4 report the 8-task SAEBench sparse-probing panel for the RSAE and its teacher (93.71% vs. 93.69%). For this response we extended it, following the other reviewers' suggestions, to SAEBench's default kk-sparse probing (top-kk latents per class, k∈{1,2,5,10,20,50}k\in\{1,2,5,10,20,50\}), for all three teacher families, using the already-trained RSAEs (Pythia-160m, ℓ0≈160\ell_0\approx160 level).

SAEk=1k=1k=2k=2k=5k=5k=10k=10k=20k=20k=50k=50all latents
ReLU teacher (dsae=212d_{\rm sae}=2^{12})76.982.186.088.590.391.293.2
RSAE (ReLU-init)76.380.385.588.390.191.593.3
Δ\Delta (RSAE − teacher)-0.6-1.8-0.5-0.2-0.2+0.3+0.1
JumpReLU teacher (dsae=212d_{\rm sae}=2^{12})75.881.386.388.389.791.293.4
RSAE (JumpReLU-init)79.082.486.388.690.091.493.6
Δ\Delta (RSAE − teacher)+3.2+1.00.0+0.3+0.3+0.1+0.2
TopK teacher (dsae=212d_{\rm sae}=2^{12})79.282.785.888.189.791.393.5
RSAE (TopK-init)79.682.685.888.089.991.493.5
Δ\Delta (RSAE − teacher)+0.4-0.10.0-0.2+0.2+0.10.0
ReLU teacher (dsae=214d_{\rm sae}=2^{14}, the pair of Table 4)76.081.284.987.389.291.193.7
RSAE (ReLU-init)77.980.585.288.390.191.293.7
Δ\Delta (RSAE − teacher)+1.9-0.6+0.2+1.0+0.9+0.1+0.1
residual stream (no SAE), top-kk dimensions64.973.181.884.987.390.093.2

Mean test accuracy (%) over the 8 SAEBench datasets; the last row is the SAEBench baseline that probes the top-kk dimensions of the residual stream itself. Across the four pairs and the six values of kk, the RSAE is within 2 points of its teacher in 23 of 24 cells, and the one larger difference is in its favour (+3.2 at k=1k=1 for the JumpReLU teacher). There is no systematic direction: the RSAE is slightly below the 2122^{12} ReLU teacher for k≤20k\le20 (−0.2 to −1.8), above the JumpReLU teacher for k≤2k\le2, indistinguishable from the TopK teacher, and level with or slightly above the 2142^{14} ReLU teacher. For scale: probes on one or two latents are noisy (on single datasets the RSAE–teacher difference reaches 10–20 points at k≤2k\le2 and stays below about 4 points for k≥5k\ge5; re-running the probe on the same teacher moves its mean accuracy by up to 0.4 points), and every SAE, teacher or RSAE, is 11–15 points above the top-1 residual-stream dimension.

We want to be precise about what this shows and what it does not. We have no theoretical guarantee on the interpretability of the learned codes, and we do not claim one. Sparse-probing accuracy is also not a direct measurement of how interpretable a dictionary is: it measures whether a small number of latents is linearly predictive of a handful of labelled concepts. We report it as the standard protocol, to show that the RSAE produces sparse features that score as well as its teacher's under that protocol. The claim of the paper is the one that our theory can support and that the experiments verify: better fidelity at matched sparsity.

Completeness of Figure 2. Figure 2 is not a selection of favourable configurations. It contains every SAEBench teacher that exists at dsae=212d_{\rm sae}=2^{12} on Pythia-160m: all six sparsity levels for JumpReLU and TopK and all five for ReLU, 17 teachers in total, including the ones where the RSAE gains nothing (the JumpReLU curve sits at 0–1% for ℓ0≤80\ell_0 \le 80, and one Δ\DeltaCE point is slightly negative and is shown as such). Table 5 gives the same comparison for all 51 (sparsity, width, family) settings, again with no omission. Since submission we went further: with better fine-tuned RSAEs we now have results for all 54 SAEBench teachers on Pythia-160m and all 54 on Gemma-2-2B (3 widths ×\times 6 sparsity levels ×\times 3 families). The RSAE lowers MSE for all 36 ReLU teachers, for 34 of 36 TopK teachers and for 29 of 36 JumpReLU teachers. The first is what Theorem 1 leads one to expect, and for the gated teachers the reduction grows with the number of active latents, which is the regime Theorem 2 identifies.

AuthorsOpenReview ↗

Response to Reviewer 9Gh3

Significance of the contribution. We respectfully disagree with the criterion implied here, namely that an SAE contribution is only worth publishing if it wins by a large margin on every existing benchmark at once. We see three problems with it.

(i) The existing metrics are partial, and none of them is "the" target. No current benchmark measures interpretability directly. Sparse probing (above) tests linear decodability of a few labelled concepts; automated interpretability scores depend on the explainer model and the prompt; absorption, SCR/TPP and RAVEL each probe one specific failure mode on specific datasets. SAEBench itself reports that gains on proxy metrics do not reliably translate into better practical performance, and that architectures rank differently on different metrics. A method can be "better" on one and "worse" on another without either verdict being wrong. Requiring a uniform win on all of them selects for benchmark coverage, not for understanding. By contrast, the quantities we target, reconstruction error and cross-entropy degradation at matched ℓ0\ell_0, are well-defined, architecture-independent, and are the ones for which we can prove statements (Theorems 1 and 2). We consider it a strength, not a weakness, that the paper claims exactly what it can prove and verify.

(ii) Which metric matters depends on the downstream use, and for a growing class of uses fidelity is the binding constraint. SAEs are no longer used only to browse features of a chat model. They are now applied to scientific foundation models in chemistry, materials science, biology and medicine, physics and astrophysics [1-3]. In natural-language settings one can often tolerate a small loss of fidelity: after splicing in the reconstruction or ablating a latent, the model still produces fluent text, and the experiment remains interpretable. In chemistry and materials models this is not the case. The model's inputs and outputs are constrained strings (e.g. polymer SMILES, pSMILES); a reconstruction error in the hidden state that would be harmless for text can turn the output into a string that is chemically invalid, i.e. not a molecule at all. An intervention study (ablating or steering a sparse feature) on top of such a reconstruction is then physically meaningless, whatever the probing accuracy of the dictionary. For these applications the first requirement on an SAE is that it reproduces the model's computation at the sparsity one can afford to inspect, and an RSAE, which lowers reconstruction error and Δ\DeltaCE at the same ℓ0\ell_0, is better suited than the SAE it starts from.

[1] Cohen, Jaron, Alexander G. Hasson, and Sara Tanovic. "Unveiling latent knowledge in chemistry language models through sparse autoencoders." arXiv preprint arXiv:2512.08077 (2025).

[2] Varadi, Kristof, Mark Marosi, and Peter Antal. "Circuits, features, and heuristics in molecular transformers." arXiv preprint arXiv:2512.09757 (2025).

[3] Simon, Elana, and James Zou. "InterPLM: discovering interpretable features in protein language models via sparse autoencoders." Nature methods 22.10 (2025): 2107-2117.

(iii) What we propose is an upgrade of SAEs people already have, and we say plainly what it does and does not buy. The RSAE starts from a released ReLU, JumpReLU or TopK SAE, adds 20 scalars, and is trained for minutes on one consumer GPU. For the most widely released family (ReLU) the gain over the released SAE is not marginal: reconstruction error drops by 27–98% and Δ\DeltaCE by 34–97% at matched ℓ0\ell_0 on the completed Pythia-160m grid (25–87% and 34–93% in Table 5 of the submission), by 14–69% and 43–75% on Gemma-2-2B (all three widths and all sparsity levels, completed since submission; see the general response), and both improve at every one of the 12 layers of GPT-2 small. For JumpReLU and TopK teachers the gains are small, and the paper says so ((C2), lines 399–416). The new controls requested by the other reviewers (see the general response) also show that an equally fine-tuned JumpReLU reaches the same fidelity or slightly better, so we do not claim that the RSAE beats the best gated SAE; we claim that one smooth, 20-parameter, gate-free activation reaches (or, from a ReLU teacher, comes within about 10% of) that level from any of the three starting families, with a proof that it can never be worse than ReLU. We think a method with a proof of non-inferiority, a clearly identified regime where it helps a lot, and an honest statement of where it only matches existing designs is a useful contribution to a field that the reviewer describes as crowded with variants of unclear value.

We believe that several of the concerns raised stem from material that is already in the submission but was overlooked, and from a narrower view of the purpose of SAEs than the one the paper takes. In light of the clarifications above and the additional results, we would appreciate it if the reviewer would re-evaluate the submission and reconsider the score.

Reviewer_UkezOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the Rational Sparse Autoencoder (RSAE), which replaces the fixed activation functions in SAEs (ReLU, JumpReLU, and TopK) with a trainable rational function. This function is shared across all latents and adds only 20 parameters. In practice, RSAE is not trained from scratch. Instead, it uses a pre-trained SAE as the teacher. It first copies the teacher's weights and makes the rational function mimic the teacher's activation function. Then it fine-tunes all parameters while keeping L0 unchanged. On GPT-2 small, Pythia-160m, and Gemma-2-2B, RSAE reduces both the reconstruction error and ΔCE at the same level of sparsity.

Critical strengths
  1. The method is simple and cheap. It adds only 20 parameters and can upgrade existing SAEs.
  2. The experiments are thorough, the frozen-rational control clearly shows how much of the gain comes from the trainable activation itself.
Critical weaknesses

The paper uses too few models, and they are all too small. The largest one is only 2B. Please use more and more advanced LLMs, especially with additional experiments on LLMs larger than 10B. In addition, only one layer is tested for each model, and each type of experiment seems to be done on only one LLM. This scale is far from enough.

SAEBench also provides interpretability metrics such as auto-interp, feature absorption, SCR/TPP, RAVEL, and unlearning, but the paper does not use any of them. For an SAE paper, showing only “better reconstruction” without checking whether the features are easier to interpret leaves a very large gap.

The feature quality check in this paper is quite limited. It is done on only one teacher, the ReLU SAE on Pythia-160m with L0 ≈ 160, so it is unclear whether the results hold for JumpReLU and TopK SAEs. The check also uses full-dictionary probing. It trains a linear classifier on all 16384 features instead of just a few. With this many features, almost any representation can reach high accuracy. For example, both models get 99.96% on europarl. As a result, this test does not really show whether the features are meaningful. The sparse probing in SAEBench is designed to use only the top-k features for this reason, but the authors do not use this setting.

Questions for authors
  1. Figure 1(d) shows when training starts from ReLU, the rational function learns a jump near 0 by itself. The authors also say that this is exactly the shape JumpReLU uses to fix shrinkage. So for ReLU, RSAE is rediscovering JumpReLU to a large extent. The paper seems missed a control: if the ReLU teacher is directly fine-tuned into a JumpReLU SAE, how much worse would the results be?
  2. Could the authors report sparse probing with only the top-k features, as in the default SAEBench setting, and include the JumpReLU and TopK teachers?
  3. How similar are the RSAE features to the teacher features after fine-tuning, for example in terms of decoder cosine similarity?
  4. Theorem 2 assumes the encoder is unchanged for TopK, but the actual algorithm fine-tunes all parameters. Does the theory still apply to the trained model?
  5. Do the results hold on larger models (>10B) and on other layers?
4 comments
AuthorsOpenReview ↗

Response to Reviewer 9Gh3 (part 1)

We thank the reviewer for the careful reading and for the concrete, checkable questions. We ran the requested experiments and address each of the reviewer's concerns below.

Larger model, more layers and more settings (W1 & Q5). A larger model. SAEBench itself only provides SAEs for Pythia-160m and Gemma-2-2B, so for a larger model we used the closest equivalent: the BatchTopK SAEs released for the Qwen3 family, trained with the same dictionary_learning code (adamkarvonen/qwen3-8b-saes). We ran the pipeline on Qwen3-8B (8.2B parameters, 36 layers, din=4096d_{\rm in}=4096) at the middle layer (residual stream after layer 18), on the teacher with ℓ0≈160\ell_0\approx160 and dsae=214d_{\rm sae}=2^{14}. BatchTopK is a family the submission listed as future work; following its release notes it is run at inference with its single learned threshold, which is exactly the gated form our initialisation handles (threshold re-gauging, knee initialisation). Nothing in the method was changed. The fine-tune uses 10k steps (20.5M tokens, 4% of the teacher's training budget) with the low learning rate we also use for Gemma-2-2B JumpReLU teachers, selected on held-out training-distribution text, and takes about 70 minutes on one 32 GB GPU, dominated by the language-model forward pass.

Qwen3-8B, layer 18, dsae=214d_{\rm sae}=2^{14}ℓ0\ell_0MSEexplained variancealiveΔ\DeltaCELRKL
BatchTopK teacher170.41394.50.7579100%0.053699.54%0.0984
RSAE (BatchTopK-init)169.31373.6 (−1.5%)0.7616100%0.0524 (−2.2%)99.55%0.0988 (+0.4%)

Evaluation follows the SAEBench core definitions on OpenWebText (128-token windows; 1.02M tokens for MSE and ℓ0\ell_0, 406k for cross-entropy), with one necessary change: Qwen3 has attention-sink positions whose residual norm is 100–1000$\times$ the median, the teacher was trained with positions above 10$\times$ the median norm removed, and its authors ask that they be excluded whenever the SAE is used. We apply that filter (0.8% of positions) identically to the teacher and the RSAE. MSE is in raw residual units (median token norm ≈\approx 98).

The result follows the pattern of the gated teachers on the smaller models: before any fine-tuning (Step 1 only) the gate-free RSAE reproduces the teacher to within 0.4% in MSE and 0.1% in ℓ0\ell_0 on held-out text, and fine-tuning gives an improvement at matched ℓ0\ell_0 (MSE −1.5%, Δ\DeltaCE −2.2%, KL unchanged within 0.5%, ℓ0\ell_0 0.6% lower).

Other layers. Since submission we ran the pipeline on all 12 layers of GPT-2 small (Bloom's ReLU SAEs, resid_pre, dsae=24,576d_{\rm sae}=24{,}576), each RSAE calibrated to its teacher's ℓ0\ell_0 within 1%:

layer01234567891011
ℓ0\ell_0131519263444525967725560
MSE change−21.8%−18.4%−15.0%−4.1%−1.8%−5.0%−3.0%−7.4%−10.3%−12.8%−16.0%−16.8%
Δ\DeltaCE change−46.3%−16.6%−45.7%−63.8%−41.1%−33.5%−28.9%−32.2%−34.6%−34.4%−32.3%−30.8%

The RSAE improves both metrics at every layer.

More settings per model. The submission's Gemma-2-2B results covered one width and one sparsity level. We have since completed Gemma-2-2B (layer 12) on the full SAEBench grid: 3 widths (212,214,2162^{12}, 2^{14}, 2^{16}) ×\times all sparsity levels ×\times 3 families (54 teachers), so the sparsity/width study is no longer Pythia-only. The Pythia-160m grid is completed in the same way, and both grids use better fine-tuned RSAEs than the submission: since then we kept improving the fine-tuning step, mainly with longer schedules and lower learning rates for the gated teachers. Summary for Gemma-2-2B at matched ℓ0\ell_0 (within 1% for all 54 pairs):

Gemma-2-2B teacher familyMSE change (range over 18 settings)
ReLU−14% to −69% (all 18 improve)
TopK−4.9% to +0.1% (lower in 16 of 18)
JumpReLU−1.0% to +0.2% (lower in 13, tie in 5)
The pattern is the one reported in the paper for Pythia: a large gain for ReLU teachers, parity to a modest gain for JumpReLU/TopK teachers.

We agree that models above 10B parameters would strengthen the paper further. Qwen3-8B is the largest model we can run end to end on our single 32 GB GPU; SAEs for the 14B and 32B models of the same suite exist, but those models do not fit in memory together with the SAE and the activation buffer. We will state this limit in the paper.

AuthorsOpenReview ↗

Response to Reviewer 9Gh3 (part 2)

Scope of the claims on interpretability (W2). We ran the SAEBench kk-sparse probing benchmark in its default top-kk setting, for all three teacher families (next item).

We would like to emphasize that our contribution is a statement about fidelity at matched sparsity: it is what Theorems 1 and 2 are about and what the experiments verify. We have no theoretical guarantee about the interpretability of the learned codes, and the paper does not claim one (Limitations, lines 481–485). We also do not think the available benchmarks would settle the question either way: each of them is a proxy for one aspect of feature quality on specific datasets, they do not agree with one another (SAEBench itself reports that gains on proxy metrics do not reliably translate into better practical performance), and none of them is a direct measurement of interpretability. Sparse probing, which we do report, measures whether a few latents are linearly predictive of labelled concepts; we use it as the standard protocol to show that the RSAE's sparse features score as well as its teacher's, not as evidence of improved interpretability.

We do think fidelity is a property worth improving on its own. In several current uses of SAEs, notably on scientific foundation models whose outputs must satisfy hard validity constraints (e.g. chemical strings), an inaccurate reconstruction invalidates any downstream intervention study, whatever the proxy scores of the dictionary. We will make this scope explicit in the introduction.

Top-kk sparse probing for all teacher families (W3 & Q2). We appreciate this suggestion; the reviewer is right that full-dictionary probing saturates. We re-ran the evaluation with SAEBench's default kk-sparse setting (top-kk latents selected per class, 8-dataset panel, default splits) for the ReLU, JumpReLU and TopK teachers and the RSAEs fine-tuned from them (Pythia-160m, ℓ0≈160\ell_0\approx160 level: the three Table 1 pairs at dsae=212d_{\rm sae}=2^{12}, whose RSAEs are the checkpoints in the supplementary material, and the 2142^{14} ReLU pair of Table 4):

SAEk=1k=1k=2k=2k=5k=5k=10k=10k=20k=20k=50k=50all latents
ReLU teacher (dsae=212d_{\rm sae}=2^{12})76.982.186.088.590.391.293.2
RSAE (ReLU-init)76.380.385.588.390.191.593.3
Δ\Delta (RSAE − teacher)-0.6-1.8-0.5-0.2-0.2+0.3+0.1
JumpReLU teacher (dsae=212d_{\rm sae}=2^{12})75.881.386.388.389.791.293.4
RSAE (JumpReLU-init)79.082.486.388.690.091.493.6
Δ\Delta (RSAE − teacher)+3.2+1.00.0+0.3+0.3+0.1+0.2
TopK teacher (dsae=212d_{\rm sae}=2^{12})79.282.785.888.189.791.393.5
RSAE (TopK-init)79.682.685.888.089.991.493.5
Δ\Delta (RSAE − teacher)+0.4-0.10.0-0.2+0.2+0.10.0
ReLU teacher (dsae=214d_{\rm sae}=2^{14}, the pair of Table 4)76.081.284.987.389.291.193.7
RSAE (ReLU-init)77.980.585.288.390.191.293.7
Δ\Delta (RSAE − teacher)+1.9-0.6+0.2+1.0+0.9+0.1+0.1
residual stream (no SAE), top-kk dimensions64.973.181.884.987.390.093.2

Mean test accuracy (%) over the 8 SAEBench datasets; the last row is the SAEBench baseline that probes the top-kk dimensions of the residual stream itself. Across the four pairs and the six values of kk, the RSAE is within 2 points of its teacher in 23 of 24 cells, and the one larger difference is in its favour (+3.2 at k=1k=1 for the JumpReLU teacher). There is no systematic direction: the RSAE is slightly below the 2122^{12} ReLU teacher for k≤20k\le20 (−0.2 to −1.8), above the JumpReLU teacher for k≤2k\le2, indistinguishable from the TopK teacher, and level with or slightly above the 2142^{14} ReLU teacher. For scale: probes on one or two latents are noisy (on single datasets the RSAE–teacher difference reaches 10–20 points at k≤2k\le2 and stays below about 4 points for k≥5k\ge5; re-running the probe on the same teacher moves its mean accuracy by up to 0.4 points), and every SAE, teacher or RSAE, is 11–15 points above the top-1 residual-stream dimension.

As stated above, we read this as "the fidelity gain is not bought at the expense of sparse-probe accuracy", not as a claim about interpretability.

AuthorsOpenReview ↗

Response to Reviewer 9Gh3 (part 3)

ReLU-to-JumpReLU fine-tuning control (Q1). We appreciate the reviewer's suggestion to add controlled experiments to isolate the contribution of the rational activation functions. At the reviewer's request, we performed the additional experiments of continuing training of native JumpReLU and TopK models, with a simple learned-activation baseline.

ReLU teacher

modelℓ0\ell_0MSE$\times100$Δ\DeltaCEalive %
teacher121.82.600.40476.2
Frozen rational, fine-tuned with our objective121.72.540.35879.7
JumpReLU fine-tuned from the ReLU teacher122.91.900.22176.1
learned polynomial, degree 9 (no denominator)121.51.930.22182.4
RSAE, type (9,8)121.91.6980.18799.9

JumpReLU teacher

modelℓ0\ell_0MSE$\times100$Δ\DeltaCEalive %
teacher155.11.340.11681.0
JumpReLU fine-tuned with its own activation154.81.330.11996.3
rational frozen at the distilled gate155.01.330.11899.9
RSAE, type (9,8)154.01.3120.1153100.0

TopK teacher

modelℓ0\ell_0MSE$\times100$Δ\DeltaCEalive %
teacher160.01.300.12599.5
TopK fine-tuned with its own activation (exact top-kk)160.01.310.12799.9
rational frozen at the distilled gate160.91.270.11099.9
RSAE, type (9,8)160.11.2400.105099.9

From the tables above, we can infer that the continuing training of the JumpReLU and TopK SAEs does not guarantee a better performance in terms of fidelity at matching sparsity than the publicly released ones, which also verifies that the contribution of our RSAE does not come from the continuing training of the finetuned procedure. We also add the simple polynomial activation function with degree of 9, with a more flexible and expressive activation than ReLU, polynomial activation function can achieve some improvements in terms of fidelity, however, it is still worse than the RSAE. Combined with the new experiments, we are more confident in saying, the reason why the RSAE works comes from the rational activation function, which remains the current best choice.

Feature similarity to the teacher (Q3). Latent jj of the RSAE is initialised from latent jj of the teacher, so we compare index-to-index (Pythia-160m, 2122^{12}, ℓ0≈160\ell_0\approx160; latents that are alive in the teacher; 65k held-out Pile tokens):

Teacherdecoder cosine, median (mean)latents with cosine > 0.9> 0.7encoder cosine, medianteacher firings also active in the RSAEactivation correlation, median
ReLU0.85 (0.80)23%85%0.9155%0.86
JumpReLU0.99 (0.99)99%99.7%0.9993%0.98
TopK0.99 (0.97)97%98%0.9989%0.97

For JumpReLU and TopK teachers the RSAE keeps the teacher's dictionary almost unchanged (median decoder cosine 0.99) and fires on essentially the same (token, latent) pairs (89–93% of the teacher's firings). For the ReLU teacher the dictionary moves more (median decoder cosine 0.85, 85% of latents above 0.7, 55% of the teacher's firings kept). This is the setting where reconstruction error drops by more than a third at the same ℓ0\ell_0: removing the ℓ1\ell_1 shrinkage changes which latents fire as well as how strongly, so the RSAE is a related but not identical dictionary.

AuthorsOpenReview ↗

Response to Reviewer Ukez (part 4)

Applicability of Theorem 2 to the trained model (Q4). Theorem 2 is an existence result, and it is used as a lower bound on what the family can reach, not as a description of the trained model. It says: holding the teacher's decoder and (for TopK) encoder fixed, there is a shared rational with the same expected ℓ0\ell_0 and strictly lower risk. The fine-tuning procedure optimises over a superset of that feasible set (it may also move the encoder and decoder), so the best achievable risk can only be lower; the theorem does not, and cannot, certify what Adam finds. We will say this explicitly after Theorem 2.

We hope the additional model, layers, controls and the top-kk probing results address the reviewer's concerns, and we would be grateful if the reviewer would reconsider the score.

Reviewer_QfpxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces RSAEs, which replace conventional encoder activations with a shared, trainable rational function followed by a non-negativity clamp. The approach retains the linear dictionary decoder and adapts existing ReLU, JumpReLU and TopK SAEs. Its two-stage procedure first inherits the teacher’s weights and calibrates the activation, then jointly fine-tunes the encoder, decoder and activation parameters using a reconstruction objective that targets the teacher’s average sparsity.

The study combines theoretical reconstruction comparisons at matched expected sparsity with experiments on residual-stream activations from GPT-2 small, Pythia-160m and Gemma-2-2B. Evaluation examines reconstruction error and the effect of substituting SAE reconstructions on language-model loss, with additional analyses of activation learning, dictionary width and sparsity.

The method offers a practical upgrade path with few additional parameters and promising reconstruction gains. Its principal weakness is that implementation ambiguities and gaps in theoretical and experimental support leave the specific benefit of rational activations insufficiently established.

Critical strengths

here is my view of the critical strengths:

  1. RSAE reduces Pythia reconstruction error by 35% at nearly unchanged average sparsity—122 versus 121 active latents—demonstrating a meaningful improvement in the reconstruction–sparsity trade-off.

  2. The matched-pipeline ablation indicates that activation learning accounts for approximately 93% of this gain, supporting its contribution beyond additional training alone.

  3. The activation adds only 20 shared parameters and uses a denominator bounded below by one, providing a compact extension of pretrained SAEs without real-axis poles.

  4. Experiments span three language models, three baseline activation families and 51 Pythia teacher configurations, supporting applicability across different dictionary sizes and sparsity regimes.

  5. Exact ReLU containment preserves the teacher’s codes and reconstructions at initialisation and establishes non-inferiority of the best achievable objective under the same constraints.

Critical weaknesses
  1. The denominator coefficients may not learn under the stated initialisation. With standard automatic differentiation, zero-initialised absolute-valued coefficients receive zero gradients, leaving ReLU-initialised models with a constant denominator. Without implementation details resolving this issue, the attribution of improvements to learning rational functions remains unsubstantiated.

  2. Table 5 contains inconsistent fidelity metrics. For Pythia at width 65,536 and nominal sparsity 40, cross-entropy degradation worsens from 0.170 to 0.171 while loss recovered improves from 97.7% to 97.9%. These changes are incompatible with the stated formulas under shared evaluation baselines, weakening confidence in the aggregate results.

  3. The theoretical guarantees do not directly cover the implemented model. The stronger results allow unrestricted finite rational degree and use a different JumpReLU reparameterisation, whereas experiments fix the degree at (9, 8). ReLU containment also provides no guarantee for finite optimisation, limiting the theory’s support for the practical claims.

  4. The controls do not establish superiority over equally trained alternatives. Freezing a converted rational activation does not substitute for continued training of native JumpReLU or TopK models. Without these comparisons and a simple learned-activation baseline, the specific advantage of rational functions remains unresolved.

  5. Sparsity calibration and uncertainty reporting are insufficiently specified. The separation between calibration and final evaluation data is unclear, while Tables 2 and 5 lack variability estimates. These omissions make the small improvements over gated teachers difficult to assess reliably.

  6. The feature-quality conclusion exceeds the probing evidence. The experiment uses all dictionary features in one Pythia/ReLU configuration, despite being labelled sparse probing. It demonstrates retained classification information but cannot establish that concepts remain concentrated in individual interpretable features.

Questions for authors

my questino are related to the above-mention comments:

  1. How do the zero-initialised denominator coefficients receive nonzero gradients under the absolute-value parameterisation?

  2. How do you reconcile the inconsistent cross-entropy degradation and loss-recovered values in Table 5?

  3. Which theoretical guarantees apply to the fixed degree-(9, 8) model and the implemented reparameterisation?

  4. Do the gains persist against equally fine-tuned native JumpReLU/TopK models and a learned polynomial baseline?

  5. Is sparsity calibration separate from final testing, and what is the variability across runs in Tables 2 and 5?

  6. Do the feature-quality conclusions hold with kk-sparse probes across ReLU, JumpReLU and TopK teachers?

5 comments
AuthorsOpenReview ↗

Response to Reviewer Qfpx (part 1)

We thank the reviewer for a precise and constructive review. We address the questions raised by the reviewer in the following paragraphs.

Clarification on the zero-initialised denominator coefficients (Q1 & W1). We agree that a standard abs is not sufficient for our proposed algorithm, since torch.abs has subgradient 0 at 0 and QQ would stay identically 1, reducing the activation of a ReLU-initialised RSAE to a polynomial function. However, we want to point out specifically that our implementation is designed to avoid exactly this situation: ∣bj∣|b_j| is evaluated as

def _abs_pos0(b):            # |b| with subgradient +1 at b == 0
    return torch.where(b >= 0, b, -b)

(rational_sae/rational_activation.py in the supplementary code, lines 45–52; it is applied to every bjb_j in the denominator at lines 306–318). The forward value is identical to ∣bj∣\lvert b_j\rvert; the only difference is the choice of subgradient at the single point bj=0b_j=0, where any value in [−1,1][-1,1] is a valid subgradient and we take +1+1. At the identity initialisation (b=0b=0, hence Q≡1Q\equiv 1) the gradient of the loss with respect to bjb_j is therefore ∂L∂bj=−∑t∂L∂r(t)P(t)∣t∣j,\frac{\partial \mathcal L}{\partial b_j} = -\sum_{t}\frac{\partial \mathcal L}{\partial r(t)}P(t)\lvert t\rvert^{j} , which is non-zero in general, and the denominator trains from the first step. This is easy to check with the supplementary code: at the identity initialisation all eight ∂L/∂bj\partial\mathcal L/\partial b_j are non-zero, and they become exactly zero if _abs_pos0 is replaced by torch.abs.

Two further pieces of evidence that the denominator is trained:

  • Learned coefficients. For the ReLU-init RSAE in the supplementary material (checkpoints/rsae_pythia160m_L8_2pow12_relu-init_t3.pt; identity init, b=0b=0 exactly), the fine-tuned denominator has ∣b∣=(1.12, 0.74, 0.12, 0.02, 0.09, 0.09, 0.03, 0.03)\lvert b\rvert=(1.12,\ 0.74,\ 0.12,\ 0.02,\ 0.09,\ 0.09,\ 0.03,\ 0.03). These are the entries of act.B in the checkpoint; the stored values are signed and only ∣bj∣\lvert b_j\rvert enters QQ. The numerator (act.A) is a=(−0.05, 1.51, −0.80, −0.10, 0.27, 0.19, … )a=(-0.05,\ 1.51,\ -0.80,\ -0.10,\ 0.27,\ 0.19,\ \dots). QQ is far from constant. The learned shape in Figure 1(d), a steep rise at the origin that bends into a linear tail, is the signature of a rational: a polynomial cannot bend into a linear tail.
  • Polynomial control (new). Removing the denominator altogether (degree-9 numerator only, Q≡1Q\equiv 1, same pipeline) is not a viable model: training diverges within 1000 steps (see Q4). The fact that the RSAEs in the supplementary material train stably from the same initialisation is itself evidence that QQ is active.

Correction of Table 5 (Q2 & W2). We thank the reviewer for raising this point. We agree that Δ\DeltaCE 0.170 → 0.171 cannot coincide with LR 97.7% → 97.9%. After careful investigation, we found that this was caused by a transcription error made when the table was assembled by hand from several result files. The raw evaluation output for that pair (Pythia, 2162^{16}, ℓ0≈40\ell_0\approx 40, TopK) is

ℓ0\ell_0MSE$\times 100$Δ\DeltaCELR
TopK teacher40.01.8200.169897.71%
RSAE (TopK-init)40.01.7750.167997.74%

i.e. a small improvement on both metrics, and the two values are consistent with each other.

Prompted by this, we audited all 102 cells of Table 5 mechanically against the identity and regenerated the entire table, which is now in the revised manuscript. In the regenerated table, the RSAE lowers MSE in all 54 (sparsity, width, family) settings and lowers Δ\DeltaCE / raises LR in 52 of 54. The conclusions of Section 5 are unchanged.

AuthorsOpenReview ↗

Response to Reviewer Qfpx (part 2)

Scope of the theoretical guarantees (Q3 & W3). We thank the reviewer for the suggestion to separate more clearly what is proved from what is observed. We clarify this below.

  • Theorem 1 (ReLU) applies to the implemented model as is. It holds for every p≥1p\ge1, q≥0q\ge0, hence for (9,8)(9,8), with the implemented clamp, scales and pole-free denominator. It guarantees two things that we use in practice: the (9,8)(9,8) family contains the teacher, and the identity initialisation is the teacher (step 0 of fine-tuning reproduces its codes and reconstruction exactly). It is a statement about the best achievable objective, not a guarantee for finite optimisation. Empirically, fine-tuning from that exact starting point improved MSE and Δ\DeltaCE in all 18 Pythia ReLU settings, all 18 Gemma-2-2B ReLU settings (3 widths ×\times 6 sparsity levels each; see the general response) and all 12 GPT-2 layers.
  • Theorem 2 (JumpReLU / TopK) is an existence result for some finite degree, obtained by a density argument; it does not bound that degree, so the degree (9,8)(9,8) is an empirical choice. Its role is to show that the advantage is not an artefact of a particular construction and to identify the regime that favours the RSAE (gain linear in kk against a k\sqrt{k} support-replacement cost), which is the trend shown in Figure 2. The completed grids (summary table in the general response) show the same trend on both models and all three widths. We present this as agreement in sign and trend, not as a quantitative test of the bound.
  • Reparameterisation. The construction in Theorem A.7 uses a per-feature affine map of the encoder, S^j=cj(Sj−θj)\hat S_j = c_j (S_j-\theta_j) with free cj>0c_j>0, and leaves the decoder untouched. The implementation uses the special case cj=1/θjc_j = 1/\theta_j and compensates by scaling decoder row jj by θj\theta_j. Both preserve each feature's support Sj>θjS_j>\theta_j and the teacher's function; the implemented choice additionally makes every feature's teacher response the same unit jump, so that one low-degree shared rational suffices where the proof spreads features over disjoint intervals of a higher-degree one. We will state this correspondence in the revised paper, together with the fact that it is a practical variant rather than the proof's construction.
AuthorsOpenReview ↗

Response to Reviewer Qfpx (part 3)

Controls with equally fine-tuned baselines (Q4 & W4). We appreciate the reviewer's suggestion to add controlled experiments to isolate the contribution of the rational activation functions. At the reviewer's request, we performed the additional experiments of continuing training of native JumpReLU and TopK models, with a simple learned-activation baseline.

ReLU teacher

modelℓ0\ell_0MSE$\times100$Δ\DeltaCEalive %
teacher121.82.600.40476.2
Frozen rational, fine-tuned with our objective121.72.540.35879.7
learned polynomial, degree 9 (no denominator)121.51.930.22182.4
RSAE, type (9,8)121.91.690.18799.9

JumpReLU teacher

modelℓ0\ell_0MSE$\times100$Δ\DeltaCEalive %
teacher155.11.340.11681.0
JumpReLU fine-tuned with its own activation154.81.330.11996.3
rational frozen at the distilled gate155.01.330.11899.9
RSAE, type (9,8)154.01.310.115100.0

TopK teacher

modelℓ0\ell_0MSE$\times100$Δ\DeltaCEalive %
teacher160.01.300.12599.5
TopK fine-tuned with its own activation (exact top-kk)160.01.310.12799.9
rational frozen at the distilled gate160.91.270.11099.9
RSAE, type (9,8)160.11.240.10599.9

From the tables above, we can infer that the continuing training of the JumpReLU and TopK SAEs does not guarantee a better performance in terms of fidelity at matching sparsity than the publicly released ones, which also verifies that the contribution of our RSAE does not come from the continuing training of the finetuned procedure. We also added a simple polynomial activation function. With degree 9, training diverges within 1,000 steps. With degree 3, the polynomial activation is more flexible and expressive than ReLU and achieves some improvement in fidelity; however, it is still worse than the RSAE. Combined with the new experiments, we are more confident in saying, the reason why the RSAE works comes from the rational activation function, which remains the current best choice.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the reviews and responses

We thank the three reviewers for their time and for feedback that was specific enough to act on. This comment summarises what we added; the individual responses give the details.

Additional experiments

All new results use the recipe and the evaluation protocol of the paper. Feature-quality results use existing RSAE checkpoints; nothing was retrained for them.

  1. kk-sparse probing for all three teacher families (Qfpx Q6, Ukez Q2, 9Gh3). SAEBench's default top-kk probing, k∈1,2,5,10,20,50k\in1,2,5,10,20,50, for the ReLU, JumpReLU and TopK teachers and their RSAEs. The RSAE is on par with its teacher at every kk.
  2. Controls with equally fine-tuned baselines (Qfpx Q4, Ukez Q1): the ReLU teacher fine-tuned directly into a JumpReLU; JumpReLU and TopK teachers fine-tuned with their own activation; a learned polynomial activation; and the setting of Theorem 2 (encoder and decoder frozen, only the 20 activation parameters trained).
  3. A larger model (Ukez Q5): Qwen3-8B (8.2B parameters, layer 18), starting from a public BatchTopK SAE with ℓ0≈160\ell_0\approx160. The RSAE matches the teacher before fine-tuning and improves it slightly after (MSE −1.5%, Δ\DeltaCE −2.2% at matched ℓ0\ell_0), the same pattern as for the other gated teachers.
  4. More layers and settings (Ukez W1/Q5): all 12 layers of GPT-2 small, and the full SAEBench grid on both Pythia-160m and Gemma-2-2B (3 widths ×\times 6 sparsity levels ×\times 3 families, 54 teachers each), with better fine-tuned RSAEs. See "Updated and extended main results" below.
  5. Feature similarity to the teacher (Ukez Q3), a held-out re-evaluation on tokens that played no role in calibration (Qfpx Q5), and seed variability for Table 2 (Qfpx Q5).

Updated and extended main results

Table 5. Reviewer Qfpx spotted a Δ\DeltaCE / loss-recovered pair in Table 5 that cannot both be right. It was a transcription error made when the table was assembled by hand from several result files. We audited all 102 cells of Table 5 mechanically and regenerated the entire table, which is now in the revised manuscript. In the regenerated table, the RSAE lowers MSE in all 54 (sparsity, width, family) settings and lowers Δ\DeltaCE in 52 of 54. The conclusions of Section 5 are unchanged.

More comprehensive and better fine-tuned results. Since submission we kept improving the fine-tuning step and completed the grid. We now have results for all 54 SAEBench teachers on Pythia-160m and all 54 on Gemma-2-2B: 3 widths (212,214,2162^{12}, 2^{14}, 2^{16}) ×\times 6 sparsity levels ×\times 3 families. The submission had 51 Pythia pairs (Table 5) and one Gemma-2-2B setting (Table 1). ℓ0\ell_0 is matched within 1% in 101 of the 108 pairs; in six of the other seven the RSAE is the sparser model.

Agreement with the theory. The two grids follow the qualitative predictions of the two theorems.

  • Theorem 1 (an RSAE is never worse than ReLU at the same budget). The RSAE lowers both MSE and Δ\DeltaCE for all 36 ReLU teachers, at every width and sparsity level of both models.
  • Theorem 2 for TopK (bound gk−ckgk-c\sqrt k). The bound allows no gain at small kk and predicts one once kk is large. The relative MSE reduction increases with ℓ0\ell_0 in all six (model, width) series, monotonically in four of them.
  • Theorem 2 for JumpReLU (bound gkgk). MSE is lower in 29 of 36 pairs, and the reduction again grows with ℓ0\ell_0.

Theorem 2 is an existence result under Assumption 2, so we present this as agreement in sign and trend, not as a quantitative test.

What the new controls show

Reviewers Qfpx and Ukez asked for equally fine-tuned baselines. The results indicate that the continuing training of the JumpReLU and TopK SAEs does not guarantee a better performance in terms of fidelity at matching sparsity than the publicly released ones, which also verifies that the contribution of our RSAE does not come from the continuing training of the finetuned procedure. We also add the simple polynomial activation function with degree of 9, with a more flexible and expressive activation than ReLU, polynomial activation function can achieve some improvements in terms of fidelity, however, it is still worse than the RSAE. Combined with the new experiments, we are more confident in saying, the reason why the RSAE works comes from the rational activation function, which remains the current best choice.

Scope of the claims

Several comments concern interpretability metrics. We want to be clear about what the paper does and does not claim. Our theory and our experiments are about reconstruction fidelity at matched sparsity. We have no theoretical guarantee about the interpretability of the learned codes and we do not claim to improve it; the probing results are reported as a standard check that the fidelity gain is not bought at the expense of sparse-probe accuracy. We will make this scope explicit in the introduction and the limitations.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.