acceptodds
Under review as a conference paper at ICLR 2027

Spike-and-Slab Sparse Autoencoders with Adaptive Sparsity

Abstract

Sparse Autoencoders (SAEs) learn sparse representations of deep neural activations, but fixed Top- architectures impose the same feature budget on every input. This aspect limits their ability to adapt representational capacity to inputs of different complexity. In this paper, we introduce Spike-and-Slab Sparse Autoencoders (S-SAE), a Bayesian framework in which the sparsity level can vary using a global learned threshold. By casting sparse encoding as a MAP inference problem, we derive an objective that balances reconstruction fidelity and expected sparsity. We additionally introduce an optional post-hoc refinement that selects among candidate code lengths without retraining. We evaluate our framework on representations from image classification models and Large Language Models, including Pythia-160M and Gemma-2-2B. Results show that S-SAE improves the observed sparsity–reconstruction Pareto frontier across vision and language benchmarks. These findings connect Bayesian sparse coding with practical adaptive sparsity, with potential applications to concept discovery in neural representations. Code is available at https://anonymous.4open.science/r/s2sae.

Primary area:
interpretability and explainable AI
Keywords:
sparse autoencoders, mechanistic interpretability, bayesian learning

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

S²-SAE is an adaptive sparse autoencoder motivated by a spike-and-slab prior. Its encoder combines a learned shared threshold with dual updates controlling average activity; an optional refinement selects a reconstruction-dependent feature budget. Experiments assess reconstruction, feature allocation, behavior preservation and inference cost on two vision datasets and Pythia/Gemma activations. Exact MAP reasoning is restricted to the stated idealized setting; the practical overcomplete encoder is an approximation (pp. 3–5; Appendix A, pp. 13–14).

Critical strengths
  • The vision evaluation spans two datasets and three SAE seeds, with favorable reconstruction results at nearby measured activity levels (p. 6, Figure 1; p. 16, Figure 6; p. 18, Appendix D). Fast inference already obtains these benefits, making the empirical contribution independent of expensive refinement.

  • The probabilistic formulation coherently motivates a shared activation criterion, with the exact-MAP assumptions and practical approximation explicitly separated (p. 4; pp. 13–14, Appendix A). This supplies a principled design rationale and a useful connection between activation costs and average-budget control.

Critical weaknesses
  • The methodological increment is limited and incompletely attributed. Input-dependent activity is already available in JumpReLU and BatchTopK (pp. 2–3). The shared threshold follows from a common activation cost in the chosen prior (pp. 13–14); this is a coherent modeling choice, but its empirical advantage remains to be established. Complete-system comparisons do not isolate threshold sharing from budget control or other training choices, leaving the distinctive design's importance uncertain.

  • Evidence beyond vision remains mixed. In the language experiments, several fast-inference gains are small or negative, and the largest high-ℓ0 Pythia gain is interpolation-sensitive (pp. 6–7; p. 17, Tables 3–4). In Figure 4, four low-to-mid-budget Pythia points lie below the connected baseline frontier; these are not precisely budget-matched comparisons. The text gives refined loss recovery of 0.9285 versus 0.9356 at the lowest-budget example despite favorable MSE (p. 8). Thus behavioral benefits are limited. Additionally, the printed Gemma MSE values are compatible with bf16 rounding; Table 4's +0.38% gain corresponds to about 0.9 bf16 ULP. Actual metric precision is unspecified, so this is a clarification need, not established numerical error (pp. 17–18). These limitations leave robust practical benefits concentrated in vision reconstruction.; several fast-inference gains are small or negative, and the largest high-ℓ0 Pythia gain is interpolation-sensitive (pp. 6–7; p. 17, Tables 3–4). In Figure 4, four low-to-mid-budget Pythia points lie below the connected baseline frontier; these are not precisely budget-matched comparisons. The text gives refined loss recovery of 0.9285 versus 0.9356 at the lowest-budget example despite favorable MSE (p. 8). Thus behavioral benefits are limited. Additionally, the printed Gemma MSE values are compatible with bf16 rounding; Table 4's +0.38% gain corresponds to about 0.9 bf16 ULP. Actual metric precision is unspecified, so this is a clarification need, not established numerical error (pp. 17–18). These limitations leave robust practical benefits concentrated in vision reconstruction.

Questions for authors
  • Which existing result most directly identifies the benefit of threshold sharing separately from average-budget control? An available representative control or a precise account of the combination's distinct contribution would help assess significance.

  • What precision is used to accumulate and report Gemma MSE? With existing checkpoints, can you establish whether the +0.38% comparison is stable to metric precision, and identify language advantages supported by observed points and Figure 4 rather than interpolation alone?

2 comments
AuthorsOpenReview ↗

Addressing W1, W2, Q1.

W1/Q1: Does threshold sharing help independently of budget control?

The most direct comparison is C1 vs. C3 in the added Tab. 9. Both use the same dual controller and training setup, and they differ only in whether the threshold is shared. We also compare C2 vs. C4 with fixed λ\lambda. Sharing the threshold reduces MSE in all six CUB comparisons and all four Pythia comparisons. The three Pythia seeds per cell at target 140 support the C1 vs. C3 ordering in this smaller ablation.

Shared vs. per-feature thresholds with budget control held fixed.

DomainMatched ℓ₀Dual λ: C1 sharedDual λ: C3 per-featureDual λ: MSE reductionFixed λ: C2 sharedFixed λ: C4 per-featureFixed λ: MSE reduction
CUB15.60.352710.364293.18%0.355560.365392.69%
CUB32.50.272410.278692.25%0.271570.279242.75%
CUB65.70.201300.206692.61%0.202210.210113.76%
Pythia36.40.075740.076350.80%0.075430.076180.98%
Pythia141.90.043500.043730.54%0.042790.042880.20%

MSE is compared at matched ℓ₀ (lower is better). MSE reduction is 100 × (per-feature MSE − shared MSE) / per-feature MSE within each budget-control setting.

W2: Evidence beyond vision remains mixed.

In SAE-based works for language models, gains have notably been small so far (e.g., BatchTopK, JumpReLU). Despite the difficult task, our Pareto-based visualizations in Fig. 2 (where all the S^2 models lie in the Pareto frontier, regardless of interpolation) were complemented with added reports in Appendix G of sensible improvements in other key metrics, including the following AvgRank metric.

MethodPythia-160M (4k latents)Gemma-2-2B (16k latents)
BatchTopK2.673.00
TopK2.333.67
JumpReLU3.672.33
GatedSAE5.005.00
Matryoshka BatchTopK6.006.00
PAnneal7.007.00
Standard8.008.00
S²-SAE (ours)1.331.00

W2: Language runs use one seed.

The published SAEBench benchmarks follow the same single-seed setting, so we chose to follow the same procedure to ensure fair comparability and to reserve computation power to provide more model training at different sparsity levels. To reassure the reviewer about the impact of this procedure, in the new 30M-token Pythia ablation study at target ℓ0=140\ell_0=140, we ran three seeds each for C1 and C3, finding small interseed variability: the MSE standard deviation is at most 1.3×10−41.3 \times 10^{-4}, and all three shared-threshold runs have lower MSE than all three per-feature-threshold runs.

W2: Several fast-inference gains are small or negative.

We provide an AvgRank-based analysis (Appendix G.2, reported above) which confirms that fast-inference presents the best average ranking across all baselines considered.

AuthorsOpenReview ↗

Addressing W2, Q1, Q2.

W2/Q2: Gemma-2B benchmark precision (bf16).

To preserve comparability, we matched the published SAE Bench settings. Gemma uses BF16, a 16-bit format used to reduce memory during inference. The SAE Bench evaluator we used carries this limited precision into the MSE calculation and final reported score. The +0.38%+0.38\% gain in Table 4 corresponds to a gap of 0.00353 MSE, smaller than the spacing of 0.00391 between adjacent BF16 values in that range. Allowing for final rounding alone, the ordering can change, so we treat this comparison as a near tie. In Table 11, even if S2S^2-SAE and JumpReLU swap ranks at target 280, S2S^2-SAE still has the best Gemma AvgRank (1.33 vs. 2.00). Pythia uses FP32 and is unaffected by the issue. We decided to report this aspect in the paper to make other researchers aware of the issue in SAEBench evaluation; it is found in a dedicated paragraph in the Appendix C. A separate discussion on the AvgRank-based report is present in Appendix G.2.

W2: These are not precisely budget-matches comparisons.

Those mismatched points correspond to the PAnneal and GatedSAE models, which are more evident in the added Appendix G.1; we do not have access to these model weights and we cannot provide more exact reports over the baseline behavior.

Q1: Which existing result most directly identifies the benefit of threshold sharing separately from average-budget control? A representative control or a precise account of the combination distinct contribution would help assess significance.

We provide an additional ablation study which isolates the benefit of shared thresholding and training procedure. We confirm that the shared threshold has a tangible benefit, while average-budget control is more of a convenience choice. This answer summarizes the extended, common answer reported above ("Ablation study").

Can you identify language advantages supported by observed points and Figure 4 rather than interpolation alone?

In addition to the (interpolation-less) Pareto-based visualizations in Fig. 2, which highlight that our models always belong to the Pareto frontier, Appendix G.1 shows MSE and cosine similarity for each baseline separately. On Gemma at target 72, fast S2S^2-SAE has MSE 1.60156 and cosine similarity 0.92188 at measured ℓ0=84.24\ell_0=84.24, versus MSE 1.63281 and cosine similarity 0.91797 at ℓ0=85.28\ell_0=85.28 for JumpReLU. At the highest Pythia target, fast inference gives MSE 0.000256 at ℓ0=605.16\ell_0=605.16, versus 0.001399 at ℓ0=603.68\ell_0=603.68 for the nearby JumpReLU checkpoint. This shows a sizable difference already in the observed scores at nearby sparsities.

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

This paper introduces Spike-and-Slab Sparse Autoencoders, which allocate input-dependent numbers of active features using a single learned global threshold. The authors derive a joint-MAP objective combining reconstruction error, a coefficient-magnitude penaly, and an activation-count penalty. They establish an exact coordinate-wise threshold rule, which also gives the global MAP solution for orthogonalk dictionaries. The method is evaluated on ConvNeXt-Tiny features (CUB, Pet; 3 seeds; self-trained baselines) and on Pythia-160M and Gemma-2-2B. The results show consistent reconstruction improvements on vision and substantial improvements at some higher-budge Pythia settings. Refinement provides modest additional reconstruction gains at considerably higher inference cost.

Critical strengths
  1. The proposed method is straightforward to implement, and its theoretical motivation is presented with appropriate caveats. The derivation shows how a shared activation cost gives rise to a uniform coordinate-wise threshold. The paper also explicitly acknowledges that the orthogonality assumption does not hold for overcomplete dictionaries and that the learned encoder is not guaranteed to recover exact MAP codes. These qualifications establish a clear connection between the probabilistic model and the practical architecture without overstating the theoretical guarantees.

  2. The empirical evidence is presented transparently and is easy to verify. The authors provide code and openly report negative results, including the underperformance of fast inference relative to baselines at Pythia targets 18–72 (Table 3) and the lack of improvement in CE recovery (Fig. 4).

Critical weaknesses
  1. The robustness and practical significance of the modest improvements on language tasks remain uncertain. While the strong vision results and high-budget Pythia performance are encouraging, the remaining language results provide less conclusive evidence of consistent gains. In particular, fast inference underperforms the interpolated baseline at the three lowest Pythia targets (Table 3), raising questions about the method's effectiveness across different sparsity regimes.

  2. The distinctive methodological contribution is not sufficiently disentangled from existing threshold-based SAE approaches. Equation (2) can be viewed as a shared-threshold specialization of the JumpReLU activation, while BatchTopK already employs a global threshold at inference time. Thus, the practical novelty appears to lie primarily in the specific combination of shared thresholds, dual budget control, surrogate gradients, and auxiliary training for dead features. However, the theoretical derivation alone does not establish which of these components is responsible for the observed improvements.

Questions for authors
  1. Could the authors provide an ablation comparing shared and feature-specific thresholds while holding the dual controller, surrogate gradients, AuxK, initialization, and training protocol fixed? Similarly, can the contribution of dual budget control be isolated from that of threshold sharing?

  2. What interpolation scheme and endpoint-handling policy are used in Tables 1–4, and how sensitive are the reported improvements, particularly the smaller gains, to these choices? Could the authors also provide comparisons against the actual nearby baseline checkpoints, along with results across multiple random seeds for representative language settings?

2 comments
AuthorsOpenReview ↗

Addressing W1, W2, Q1.

We thank reviewer TBjT for their feedback.

W1: The modest improvements on language tasks remain uncertain.

In SAE-based works for language models, gains have notably been small so far (e.g., BatchTopK, JumpReLU). Despite the difficult task, our Pareto-based visualizations in Fig. 2 (which already show that S2S^2-SAE always belong to the Pareto frontier, with no interpolation) were complemented with added reports in Appendix G of sensible improvements in other key metrics.

W2/Q1: Separating threshold sharing and dual control. / Differences from BatchTopK

We added the requested 2x2 ablation (Appendix F.1). C1 vs. C3 compares shared and per-feature thresholds with the same dual controller, and C2 vs. C4 repeats the comparison with fixed λ\lambda. We hold the surrogate gradient, AuxK, initialization, and training protocol fixed. Shared thresholds lower reconstruction error in both comparisons at all tested CUB and Pythia targets (Tab. 9). The fixed-λ\lambda sweep is centered on the value reached by the corresponding dual run. Dual control tends to help on CUB, and the Pythia ablation favors fixed λ\lambda at matched ℓ0\ell_0 (Tab. 10). We believe dual control to be an optional training choice which makes choosing a hyperparameter easier, but does not provide definitive advantage. Our shared comment also explains how BatchTopK's global inference threshold differs from the rule used during its training.

Shared vs. per-feature thresholds with budget control held fixed.

DomainMatched ℓ₀Dual λ: C1 sharedDual λ: C3 per-featureDual λ: MSE reductionFixed λ: C2 sharedFixed λ: C4 per-featureFixed λ: MSE reduction
CUB15.60.352710.364293.18%0.355560.365392.69%
CUB32.50.272410.278692.25%0.271570.279242.75%
CUB65.70.201300.206692.61%0.202210.210113.76%
Pythia36.40.075740.076350.80%0.075430.076180.98%
Pythia141.90.043500.043730.54%0.042790.042880.20%

MSE is compared at matched ℓ₀ (lower is better). MSE reduction is 100 × (per-feature MSE − shared MSE) / per-feature MSE within each budget-control setting.

AuthorsOpenReview ↗

Addressing Q2.

Q2: Interpolation and nearby checkpoints.

For each baseline separately, we interpolate MSE linearly in log⁡ℓ0\log \ell_0 between its two checkpoints bracketing the measured ℓ0\ell_0 of S2-SAE. At an exact match, including endpoints, we use the measured MSE. We omit a baseline when the comparison requires extrapolation and select the lowest eligible MSE. Revised Tab. 1–4 also report gains from linear interpolation in ℓ0\ell_0, and Tab. 5 summarizes the choice.

Sensitivity of MSE gains to the interpolation rule

SettingMedian gain (log ℓ₀)Median gain (linear ℓ₀)Median abs. change (pp)Max abs. change (pp)
Pythia-160M, refined+2.79%+3.02%0.160.95
Pythia-160M, fast+1.07%+1.31%0.150.95
Gemma-2-2B, refined+1.63%+1.89%0.221.76
Gemma-2-2B, fast+1.04%+1.33%0.262.05
All 24 comparisons+1.55%+1.82%0.162.05

The gain keeps the same sign in all 24 comparisons, with a median absolute change of 0.16 percentage points and a maximum of 2.05. Fast Pythia still has higher interpolated MSE at the three lowest targets. Fig. 2 and Appendix Fig. 8 show the measured checkpoints. For fast Pythia at target 280, our MSE is 0.00479 at measured ℓ0=309.00\ell_0=309.00, versus 0.00515 at ℓ0=309.94\ell_0=309.94 for TopK. At target 72, BatchTopK achieves slightly lower MSE at slightly higher measured ℓ0\ell_0.

Q2: Could the authors provide comparisons against the actual nearby baseline checkpoints?

If we understood this point correctly, this request cannot be satisfied since submitted results employ aggregated SAEBench data. Those mismatched sparsity points correspond to the PAnneal and GatedSAE models, which are more evident in the added Appendix G.1; we do not have access to these model weights and we cannot provide more exact reports over the baseline behavior.

Q2: Could the authors provide results across multiple random seeds for representative language settings?

The published SAEBench benchmarks follow the same single-seed setting, so we chose to follow the same procedure to ensure fair comparability and to reserve computation power to provide more model training at different sparsity levels. To reassure the reviewer about the impact of this procedure, in the new 30M-token Pythia ablation study at target ℓ0=140\ell_0=140, we ran three seeds each for C1 and C3, finding small interseed variability: the MSE standard deviation is at most 1.3×10−41.3 \times 10^{-4}, and all three shared-threshold runs have lower MSE than all three per-feature-threshold runs.

Reviewer_2wfoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces Spike-and-Slab Sparse Auto-encoders to help representational capacity to inputs of different complexity. It cast sparse encoding as a MAP inference problem, and derive an objective that balances reconstruction fidelity and expected sparsity. In draft, authors establish theoretical connection between a shared activation cost and a common threshold rule for coordinate-wise inference. The experiments show improvements in the sparsity-reconstruction trade-off across many evaluated settings, with stronger gains on vision tasks and more mixed results on language benchmarks.

Critical strengths
  1. The paper looks at important fixed-Top-K SAEs where every input receives the same number of active features regardless of the encoder's scores or the input's reconstruction needs.
  2. The derivation is clean and correct to me and authors are honest about the limits of the theory such as orthogonal-case caveat, proxy framing, etc.
  3. Evaluation is diverse and paper identifies SAE-only latency comparisons rather than end-to-end model speedups.
Critical weaknesses
  1. The paper distinguishes itself over existing adaptive-sparsity SAEs by using a common threshold, a spike-and-slab motivation, and a dual average-sparsity constraint. But the practical implementation does not estimate the Bayesian prior parameters or solve the MAP problem. Instead, it trains an encoder with a straight-through estimator and a sparsity-control objective.
  2. How much of the reported improvement comes from the spike-and-slab formulation, rather than from the choice of thresholded activations, the training procedure, and implementation details?
  3. What are authors thoughts about how well the learned scores reproduce the residual-based coordinate decisions?
  4. The paper compares its language results against published SAE checkpoints and uses interpolation to estimate baseline performance. Interpolation is useful for exploration, but it does not establish performance against a model trained and evaluated at exactly the same sparsity. I feel stronger evidence is required to ensure that the gains are attributable to the proposed sparsity mechanism.
Questions for authors
  1. What is the precise algorithmic novelty over JumpReLU? Can you isolate benefit of the shared threshold and spike-and-slab motivation from the effects of the straight-through estimator, auxiliary loss, and training setup?
  2. Can you empirically validate the relationship between learned encoder scores and residual-dependent evidence, or show a concrete benefit from a Bayesian-derived design choice?
2 comments
AuthorsOpenReview ↗

Addressing W1, W2, W3 comments.

We thank reviewer 2wfo for their time and observations.

W1/W2: What comes from the spike-and-slab formulation?

The formulation gives us a reason to tie the thresholds of a JumpReLU-style encoder. The common activation cost yields the same residual-evidence threshold for each coordinate (Appendices A and E). For a learned overcomplete dictionary, we train an amortized encoder and learn its shared threshold end-to-end using an STE. In Table 9, C1 vs. C3 isolates threshold sharing with the same dual controller, STE, AuxK, initialization, and training protocol. C2 vs. C4 tests the same choice with fixed λ\lambda. The improvement in both comparisons is a concrete benefit of the design choice motivated by the Bayesian model. Appendix F.2 also reports a separate STE bandwidth sweep. Regarding the reviewer’s concern about the misalignment of the MAP problem formulation and its solution: in the general case, exact ℓ0\ell_0 optimization requires exponential time to solve exactly; thus, approximations should be investigated to find good solutions with acceptable runtimes.

Shared vs. per-feature thresholds with budget control held fixed.

DomainMatched ℓ₀Dual λ: C1 sharedDual λ: C3 per-featureDual λ: MSE reductionFixed λ: C2 sharedFixed λ: C4 per-featureFixed λ: MSE reduction
CUB15.60.352710.364293.18%0.355560.365392.69%
CUB32.50.272410.278692.25%0.271570.279242.75%
CUB65.70.201300.206692.61%0.202210.210113.76%
Pythia36.40.075740.076350.80%0.075430.076180.98%
Pythia141.90.043500.043730.54%0.042790.042880.20%

MSE is compared at matched ℓ₀ (lower is better). MSE reduction is 100 × (per-feature MSE − shared MSE) / per-feature MSE within each budget-control setting.

W3: How closely do learned activations match residual evidence?

In Appendix E, we hold the other coefficients fixed and compare each active coefficient zjz_j with the residual-optimal magnitude cj∗=(dj⊤r−j)+c_j^{\ast} = (d_j^{\top} r_{-j})_+. On held-out CUB patches at matched mean ℓ0=15\ell_0=15, the shared encoder has 5.5% lower coefficient MSE than the per-feature threshold control (3.061 vs. 3.237). The reductions at ℓ0=30\ell_0=30 and 59 are 2.1% and 2.9%, averaged over three paired seeds (Tab. 7). This tells us how well the encoder estimates the magnitude of the features it activates, and generally that sharing a threshold is a good strategy to match residual evidence more closely.

Matched mean ℓ₀Shared thresholdNon-shared thresholdsMSE reduction
153.0613.2375.5%
301.9101.9512.1%
590.7870.8102.9%
AuthorsOpenReview ↗

Addressing W4, Q1, Q2.

W4: I feel stronger evidence is required to ensure that the gains are attributable to the proposed sparsity mechanism.

Fair evaluation in SAE research is challenging due to the MSE-ℓ0\ell_0 tradeoff. To make our analysis as fair as possible, in addition to reporting individual zooms (Fig. 2) over different sparsity ranges in language tasks, we also report the results of different interpolation strategies (Appendix, updated Tab. 1–4), confirming existing trends. Tab. 5 shows an aggregated comparison.

Sensitivity of MSE gains to the interpolation rule

SettingMedian gain (log ℓ₀)Median gain (linear ℓ₀)Median abs. change (pp)Max abs. change (pp)
Pythia-160M, refined+2.79%+3.02%0.160.95
Pythia-160M, fast+1.07%+1.31%0.150.95
Gemma-2-2B, refined+1.63%+1.89%0.221.76
Gemma-2-2B, fast+1.04%+1.33%0.262.05
All 24 comparisons+1.55%+1.82%0.162.05

Finally, we also report a AvgRank-based analysis in which aggregates the observed available baselines, highlighting the significance of our results.

MethodPythia-160M (4k latents)Gemma-2-2B (16k latents)
BatchTopK2.673.00
TopK2.333.67
JumpReLU3.672.33
GatedSAE5.005.00
Matryoshka BatchTopK6.006.00
PAnneal7.007.00
Standard8.008.00
S²-SAE (ours)1.331.00

Q1: Can you show a concrete benefit from a Bayesian-derived design choice? / What is the precise algorithmic novelty over JumpReLU?

It is a shared, learned threshold. JumpReLU does not share its threshold, and BatchTopK uses a different mechanism to learn a global threshold. The Bayesian-derived design choice motivates the usage of a shared threshold (added Appendix E). This answer summarizes the extended common answer reported above ("Relationship to JumpReLU / BatchTopK").

Q2: Can you isolate benefit of the shared threshold and spike-and-slab motivation from the effects of the straight-through estimator, auxiliary loss, and training setup?

We provide an additional ablation study that isolates the benefit of these aspects. We confirm that the shared threshold has a tangible benefit. This answer summarizes the extended common answer reported above ("Ablation study").

Comments on the paper

1 comment
AuthorsOpenReview ↗

Common answer to main doubts (JumpReLU, BatchTopK, ablation study)

We thank all reviewers for their constructive comments. We believe we addressed the main reviewers' concerns, and this gave us the opportunity to show the value of our contribution in a more direct way; in particular, addressing the common ablation-based requests gave our paper better insights on the shared-threshold mechanism we proposed; therefore, we are grateful to the reviewers contributions that helped us building a sound and more supported argument.

We understand that the main point shared by the reviews is the relationship between our approach and existing methods such as JumpReLU and BatchTopK. We added Appendix E and F to address these issues, and we summarize our observations as follows.

Relationship to JumpReLU

JumpReLU learns a separate threshold for each SAE latent, while S2S^2-SAE learns a single shared threshold. This change has a meaningful property. In our broad-slab spike-and-slab model, activating any latent has a common sparsity cost qq. With the other coefficients held fixed and a unit-norm decoder vector, the best nonnegative coefficient is cj∗=(dj⊤r−j)+c_j^{\ast} = (d_j^{\top} r_{-j})_{+}, which reduces MSE by (cj∗)2/dinput(c_j^{\ast})^2/d_{\mathrm{input}}. Thus the optimal coordinate update activates jj if cj∗>q,dinputc_j^{\ast} > \sqrt{q\\,d_{\mathrm{input}}}. In the orthogonal case, this is optimal for all coordinates at once (check new Appendix E!). In practice, we learn an encoder whose scores approximate this residual-based criterion. The Bayesian model motivates the shared threshold used by this encoder.

Relationship to BatchTopK

BatchTopK has a training rule and an inference rule, which are different. S2S^2 SAE uses the same rule in training and at fast inference, making the training process more aligned with how the model is used. In particular:

  1. Training rule of BatchTopK: Let A∈RB×mA \in \mathbb{R}^{B \times m}, where BB is the batch size and mm is the dimensionality of the SAE latent vector. Then, τ\tau is the B⋅KtargetB \cdot K_{\text{target}}-th largest latent activation, where KtargetK_{\text{target}} is the desired level of sparsity. This is an estimator of a global threshold for the current batch, whose fluctuation during training is tracked to estimate the inference threshold.

  2. Inference rule of BatchTopK: τBatchTopK\tau_{\text{BatchTopK}} is a fixed value, computed aggregating the existing thresholds found (e.g., average or EMA).

BatchTopK gives them the same budget, so a feature can be kept or dropped for the same input depending on the other inputs in its batch, even with the model weights held fixed. Our Bayesian framing motivates the shared threshold with a common activation cost, and explicitly encodes a single shared parameter updated with backpropagation and learned in an end-to-end fashion.

Ablation study

To isolate these choices, we added a 2x2 ablation with shared or per-feature thresholds and dual or fixed budget control (Appendix F.1, Tables 8 – 10). We hold the straight-through estimator, AuxK, initialization, and training protocol fixed; in particular, STE and AuxK are components which are already employed by many existing SAEs. Sharing the threshold lowers MSE with either budget controller at every tested target, by 2.2 – 3.8% on CUB and 0.2 – 1.0% on Pythia, adding an empirical motivation to the usage of a shared, learned threshold. Dual control lowers the MSE point estimate in five of six CUB comparisons, with most differences within seed variation. Fixed λ\lambda gives lower interpolated MSE in the Pythia ablation. We therefore present dual control as an optional, convenient training choice.

Shared vs. per-feature thresholds with budget control held fixed.

DomainMatched ℓ₀Dual λ: C1 sharedDual λ: C3 per-featureDual λ: MSE reductionFixed λ: C2 sharedFixed λ: C4 per-featureFixed λ: MSE reduction
CUB15.60.352710.364293.18%0.355560.365392.69%
CUB32.50.272410.278692.25%0.271570.279242.75%
CUB65.70.201300.206692.61%0.202210.210113.76%
Pythia36.40.075740.076350.80%0.075430.076180.98%
Pythia141.90.043500.043730.54%0.042790.042880.20%

MSE is compared at matched ℓ₀ (lower is better). MSE reduction is 100 × (per-feature MSE − shared MSE) / per-feature MSE within each budget-control setting.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.