acceptodds
Under review as a conference paper at ICLR 2027

Sparse Autoencoders Reveal Temporally Stable Medical Semantics in Diffusion Models

Abstract

Sparse Autoencoders (SAEs) are useful for discovering sparse directions in generative-model activations, but diffusion activations vary strongly across denoising timesteps. An SAE trained on raw multi-timestep activations can therefore mix class-relevant medical structure with noise-stage effects, making the same sparse direction unstable over time. We introduce Piecewise DoD-SAE, a simple training recipe that reconstructs scale-normalized local Direction-of-Deviation (DoD) residuals between neighboring latent trajectory states rather than raw hidden states. The objective is motivated by residual-scale diagnostics along the denoising path and is designed to expose sparse directions whose medical responses persist across timesteps. We evaluate temporal stability with top- latent IoU, image-space steering consistency, and matched classifier scoring of guided images. Across three liver MRI benchmarks, Piecewise DoD-SAE improves cross-timestep stability and steering consistency over SAE baselines under matched protocols. The results support a calibrated claim: piecewise DoD helps SAEs reveal temporally stable medical latent responses in diffusion models.

Primary area:
interpretability and explainable AI
Keywords:
Sparse Autoencoders, Diffusion Models, Mechanistic Interpretability, Temporal Consistency, Medical Image Generation, Semantic Steering, Representation Analysis

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_1C8xOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Piecewise DoD-SAE to improve the temporal stability and interpretability of sparse features in diffusion models. The method trains SAEs on normalized differences between neighboring timestep representations instead of raw features, aiming to reduce timestep-dependent variations. Experiments on three liver MRI datasets and MVTec AD show improved feature stability, semantic steering consistency, and downstream classification performance compared with existing SAE methods.

Critical strengths
  1. The paper addresses a clear and meaningful problem with a simple and intuitive approach. The proposed method is easy to understand and does not require major architectural changes.

  2. The experiments are organized around three clear research questions, making the evaluation logical and easy to follow. The ablation studies and additional analyses provide convincing evidence for improved temporal stability and steering consistency.

Critical weaknesses

The effectiveness of the proposed DoD formulation is not fully validated. The paper lacks a direct comparison with closely related temporal SAE methods, and the steering comparisons use the same gain despite large differences in actual intervention strength. These issues make it difficult to determine whether the improvements come from the proposed method or the evaluation setup.

Questions for authors
  1. Please provide more direct comparisons with closely related temporal SAE methods and evaluate steering under matched intervention strengths to better validate the effectiveness of the proposed DoD formulation.

  2. It would be helpful to include more implementation details for downstream classification, such as the optimizer, learning rate, batch size, training epochs, and data augmentation, to improve reproducibility.

  3. The medical experiments are limited to liver MRI, while the paper makes broader claims about medical semantics. Additional evaluations on other anatomical regions (e.g., brain or cardiac MRI) or imaging modalities (e.g., CT) could further demonstrate the generalizability of the proposed method.

3 comments
AuthorsOpenReview ↗

Author Response to Reviewer

Thank you for the helpful questions on baseline coverage, matched-effect steering, reproducibility, and generalization. We address each point below.

Regarding W1 and Q1: Temporal SAE Comparison and Matched-Effect Steering

We agree that Residualized Temporal SAE [1] and TIDE [2] are closely related temporal SAE methods. However, neither method currently provides publicly available code, preventing a faithful quantitative comparison without introducing substantial reimplementation choices. Upon acceptance, we will release our implementation to support direct comparison and future temporal-SAE research.

[1] Yeung et al., “Residualized Temporal Sparse Autoencoders for Interpreting Diffusion Models,” arXiv:2605.27813, 2026.
[2] Huang et al., “TIDE: Temporal-Aware Sparse Autoencoders for Interpretable Diffusion Transformers in Image Generation,” AAAI, 2026.

Original gain calibration. Raw gain values are not directly comparable across SAE parameterizations. In Figures 4 and 7, we therefore calibrate each method's gain range using random top-KK steering, selecting ranges in which random latents already produce a comparable, non-negligible SSC change. Within these calibrated ranges, the low-entropy directions exhibit different controllability: P-DoD-SAE produces progressively stronger and more consistent semantic effects, whereas the baselines respond weakly or saturate. Thus, the original protocol compares semantic controllability after method-specific gain calibration, while matched effect-L2L_2 provides a complementary image-space magnitude control.

For the reproducible baselines, we additionally evaluate steering after matching the image-space effect magnitude across methods to verify that P-DoD-SAE's advantage persists at matched intervention scales. The five target effect-L2L_2 levels come from the P-DoD-SAE gain sweep reported in Table 12, including intermediate points from the same sweep. We tune method-specific gains so that, among reachable settings, the achieved effect-L2L_2 differs from its target by at most 2.2%. Each cell reports effect cosine / SSC LPIPS / SSC SSIM.

MethodL2≈0.0026L_2{\approx}0.0026L2≈0.0093L_2{\approx}0.0093L2≈0.0316L_2{\approx}0.0316L2≈0.1481L_2{\approx}0.1481L2≈0.3474L_2{\approx}0.3474
P-DoD-SAE.2921 / .4042 / .6660.3405 / .4038 / .6678.4215 / .3637 / .6549.5730 / .3221 / .6602.8517 / .2029 / .7565
SAeUron.0604 / .4042 / .6654.0613 / .3946 / .6637.0630 / .3970 / .6648.0674 / .4545 / .6233.0731 / .5065 / .5722
DiffLens.0045 / .4037 / .6654.0046 / .4045 / .6632.0047 / .4123 / .6648unreachableunreachable
  • Effect direction. P-DoD-SAE retains substantially higher effect-direction cosine at every matched magnitude.

  • Main SSC. The methods are similar under weak interventions, while P-DoD-SAE becomes clearly stronger at the two largest reachable magnitudes.

  • Response to gain. SAeUron requires gains several orders of magnitude larger while its effect cosine remains nearly flat, and DiffLens saturates before reaching the two strongest targets. In contrast, P-DoD-SAE reaches all five levels with increasingly aligned effects, consistent with its schedule-normalized, direction-based training target rather than an advantage caused only by larger interventions.

Regarding Q1

Please see our response to W1, where we report the requested matched-effect comparison.

Regarding Q2: Downstream Classification Details

Thank you for the suggestion. All four classifiers are initialized from public pretrained checkpoints and trained with Adam, batch size 32, and at most 110 epochs, with validation-loss early stopping. We use cross-entropy with 0.1 label smoothing. Training images are resized and randomly cropped, while validation and test images use center crops; no mixup is applied. The learning rate and weight decay are initialized at 10−410^{-4} and 10−410^{-4}, respectively, and selected on the validation split; the resulting classifier-specific settings are kept identical across Real and all SAE-guided conditions. We will include the complete per-classifier configurations in the appendix.

Regarding Q3: Generalization Beyond Liver MRI

Thank you for the suggestion. Appendix E provides an additional evaluation on the natural-image MVTec AD dataset, showing that the temporal-stability improvement extends beyond the liver MRI setting. During the rebuttal period, our available computational resources were insufficient to retrain the diffusion and SAE pipelines on additional anatomical regions or imaging modalities. We therefore leave broader medical validation, including brain or cardiac MRI and CT, as important future work and limit our current medical claims to the evaluated liver MRI benchmarks.

Thank you for the considerate and detailed rebuttal, which has addressed some of my concerns. I am maintaining my original rating of weak acceptance.

Thank you for your positive feedback and for taking the time to carefully consider our rebuttal. We sincerely appreciate your thoughtful comments and support.

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

This paper introduces Piecewise DoD-SAE, a novel training recipe for Sparse Autoencoders in diffusion models that reconstructs scale-normalized local Direction-of-Deviation (DoD) residuals along trajectory segments rather than raw multi-timestep activations or global t=0 targets. Evaluated across three liver MRI benchmarks, the proposed method successfully mitigates timestep non-stationarity to reveal temporally stable medical semantic representations and improve steering consistency.

Critical strengths

The transition from global h0h_0-anchored residuals to piecewise, scale-normalized local DoD targets effectively addresses the non-stationarity and activation drift inherent in multi-step diffusion processes.

The validation protocol is rigorous and multifaceted, moving beyond mere reconstruction to rigorously assess cross-timestep stability via top-K IoU, Semantic Steering Consistency (SSC), and downstream class-relevant signal preservation using multiple classifiers.

Grounding the interpretability framework in specialized medical imaging benchmarks demonstrates high practical utility for explaining complex generative features in high-stakes domains.

Critical weaknesses

The effectiveness of the piecewise DoD target heavily relies on the choice of interval size ss. The paper lacks a comprehensive sensitivity analysis regarding how varying segment lengths across different network depths impact feature granularity and stability.

Although the method is validated thoroughly on liver MRI datasets, its generalizability to natural image generation models (such as Stable Diffusion or Imagen on ImageNet/COCO) remains empirically unproven, leaving open whether the benefits translate to highly heterogeneous semantic spaces.

Computing local trajectory segments and requiring additional latent caching steps or secondary forward passes during training and intervention introduce extra computational complexity compared to standard SAE baselines.

While the denominator in Eq. (9) accounts for diffusion-scale factors, the paper does not theoretically analyze potential error accumulation or amplification when propagating these local normalizations recursively across long denoising trajectories.

Questions for authors

See Weakness

1 comment
AuthorsOpenReview ↗

Author Response to Reviewer

Thank you for the thoughtful review. We address the sensitivity, generalization, computational-cost, and error-propagation concerns in turn.

Regarding W1: Interval Sensitivity and Network Depth

The interval ss is the only method-specific hyperparameter. We have evaluated its sensitivity and extensibility from five aspects:

  • Interval size: Table 5 sweeps s∈{100,200,500,1000,2000}s\in\{100,200,500,1000,2000\} and reports SSC and IoU, showing that small ss weakens image-space consistency while large ss reduces latent overlap.

  • Network depth: Appendix B, Figure 9 compares residual-scale alignment across four decoder layers; Appendix D, Figure 17 reports temporal IoU at output_blocks.7, showing that deeper image-adjacent layers increasingly approach global geometry while middle and shallower features benefit more from piecewise normalization.

  • Downstream setting: Appendix D, Table 6 evaluates classifier scoring with K=100K=100, where P-DoD-SAE leads across all four classifiers and extends the K=5K=5 conclusion.

  • SAE architecture: Appendix D, Figure 18 compares ReLU and TopK parameterizations, with DoD-SAEs retaining stronger temporal overlap under both.

  • Domain and denoiser: Appendix E, Tables 7--8 evaluate MVTec AD with LDM-style and DiT denoisers; Figures 20--21 report the corresponding interval-scale diagnostics, showing that the objective and interval-selection principle transfer beyond the medical U-Net setting.

Together, these results demonstrate that P-DoD-SAE is adaptable across interval choices, hook layers, steering budgets, SAE parameterizations, data domains, and denoiser architectures, rather than being tied to a single experimental configuration.

Regarding W2: Natural-Image and Diffusion-Architecture Generalization

We have evaluated natural-image and diffusion-architecture generalization from three aspects:

  • Natural-image/LDM setting: Appendix E, Table 7 evaluates MVTec AD with an LDM-style denoiser and 10-step DDIM sampler, where P-DoD-SAE achieves the strongest IoU and SSC results despite the domain and schedule shift.

  • DiT setting: Appendix E, Table 8 reports the same evaluation with a DiT denoiser, where P-DoD-SAE outperforms both baselines on every IoU and SSC metric.

  • Qualitative case: Appendix E, Figure 19 changes the abnormal lower-left pin while preserving the remaining chip structure, demonstrating localized steering of the target concept.

Regarding W3: Computational Complexity

As shown in Algorithm 3 for training and Algorithm 1 for steering, P-DoD-SAE evaluates only the two endpoint states hbh_b and hth_t rather than unrolling the local trajectory. Let nn denote the SAE computation and mm one frozen-denoiser forward; training or steering therefore costs O(n+2m)O(n+2m), compared with O(n+m)O(n+m) for a standard SAE. The additional no-gradient forward changes only the constant computational factor, is independent of ss and TT, and introduces no additional trainable parameters.

Regarding W4: Error Accumulation Along the Denoising Trajectory

Thank you for raising this concern. The required forward-diffusion endpoints can be sampled directly in O(1)O(1) time without unrolling the intermediate timesteps. In Algorithm 1, the normalization in Eq. (9) is computed only for the current pair (hb,ht)(h_b,h_t), and the resulting update is mapped directly to hth_t; neither the normalized residual nor the inverse scale factor is recursively propagated to the next timestep. Therefore, the normalization error does not accumulate through repeated local transitions. Table 5 further verifies stability across interval lengths.

Reviewer_Dbs1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether sparse autoencoders can identify features that remain consistent across diffusion timesteps. Rather than reconstructing raw activations, the proposed Piecewise DoD-SAE reconstructs normalized differences between nearby hidden states. The experiments examine latent-selection stability, steering consistency, and downstream classification, primarily on liver MRI.

The temporal-stability results are promising, and the appendix contains substantial supporting experiments.

Critical strengths

I like the idea. The proposed change is simple and directly addresses timestep-dependent representations.

At K = 5, Table 1 reports an IoU of 88.5%, compared with 72.1% for Global DoD, 35.5% for DiffLens, and 30.1% for SAeUron. The normalization ablation also helps distinguish latent stability from image-level consistency.

The supplementary experiments are supportive to include patient-disjoint splits, validation-only latent selection, ReLU and TopK variants, another layer, and non-medical experiments with both an LDM-style model and a DiT.

Critical weaknesses

W1. Stable feature concerns.

Do the stable features represent specific medical concepts? Could you distinguish temporal consistency, class association, and medical-concept validity more explicitly? A feature can remain associated with a diagnosis across timesteps while representing acquisition characteristics, background anatomy, or class-conditioned texture.

Could you validate one or two claimed attributes, such as margin appearance or texture heterogeneity, using independent annotations or blinded assessment? Would steering change the intended attribute while preserving unrelated properties?

The existing rectangular-ROI analysis is useful, but could you explain what it establishes beyond localization? Table 11 reports ROI IoU of 37.0% for P-DoD versus 36.1% for SAeUron, which provides less separation than the temporal-stability results.

W2. Steering comparisons.

Are steering comparisons made at comparable intervention strength? Multiplying latent codes by the same gain does not guarantee comparable edits. In Appendix Table 12, at log r = 5, effect L2 is 0.0316 for P-DoD, 0.0049 for SAeUron, and 0.0014 for DiffLens. These measurements show that P-DoD is not winning by producing negligible changes, but could its directional advantage be reassessed at matched effect magnitude?

W3. What is the distinction from existing methods?

Could you compare the proposed objective more directly with [1,2].

Residualized Temporal SAEs remove predictable neighboring-timestep dynamics before sparse coding, whereas this paper uses schedule-normalized local differences. TIDE also explicitly addresses temporal representations. Could you explain which benefit comes from residualization, which from the chosen normalization, and which from using a shared dictionary?

[1] Residualized Temporal Sparse Autoencoders for Interpreting Diffusion Models [2] TIDE: Temporal-Aware Sparse Autoencoders for Interpretable Diffusion Transformers in Image Generation?

W4. paired hidden states.

How are the paired hidden states obtained during steering? Could you specify how the second state x_b is obtained in Algorithm 1, including its noise coupling with x_t? Does it follow the same construction as the paired states used for training?

For b = 0, is the method using an available original image in a reconstruction setting, an estimated clean image, or another trajectory state? Could the paper distinguish these settings and ensure that baseline methods receive equivalent information?

W1

Questions for authors

Q1. Could you apply the local-residual scale diagnostic to the main MRI layer? Figures 8–9 examine a global residual, while Equation 9 normalizes a local residual. How directly do those plots support the normalization actually used?

Q2. Could you specify decoder-column normalization or an equivalent scale constraint?

Q3. Could the main equation include the boundary rule already present in Algorithm 3, namely clamping the earlier timestep at zero?

Q4. Some passages read as revision history rather than descriptions of the final experiments. There are also duplicated citations and uncommon terminology. Eg, Appendix B.2, page 14, line 741: “introduced by Beizaee et al. Beizaee et al. (2025)”. Page 8, lines 412 and 415; Table 4: “whiten normalization” and “whiten norm” are awkward terms. Should these be “whitening” throughout?

8 comments
AuthorsOpenReview ↗

Author Response to Reviewer (W1, W2)

Thank you for the constructive review. We clarify below how the current evidence distinguishes temporal consistency from class association and medical-concept validity.

Regarding W1: Stable Feature Concerns

1. What do the stable features represent? Yes, we can explicitly distinguish them.

(a) Temporal consistency vs. class association. Top-KK IoU and SSC establish temporal consistency in latent and image space, respectively, while cross-classifier prediction (Table 3) shows that the steered signal remains label-predictive across classifier architectures rather than reflecting a classifier-specific shortcut.

(b) Medical concept vs. acquisition/anatomy. We combine lesion-ROI overlap with case studies to make this distinction. A feature reflecting only global acquisition characteristics or background anatomy would not consistently align with lesion-containing ROIs while matching specific pathological appearances, such as lesion margins in Appendices C and H. This joint evidence supports the medical-concept interpretation of the selected features.

2. Does steering alter a target attribute while preserving unrelated content? We agree that blinded assessment would add rigor, and we will include it in the final version. However, our current evidence strongly supports attribute specificity. DoD masking provides a controlled setting by keeping visible regions fixed; correspondingly, Figure 19 shows that steering modifies the target anomaly at the lower-left pin while preserving the surrounding context. This supports localized, attribute-specific steering rather than a global alteration of image texture.

3. What does the rectangular-ROI analysis establish beyond localization? We use ROI overlap as a spatial check against the background-shortcut hypothesis, not as standalone proof of medical-concept validity. The modest 37.0% versus 36.1% separation is expected because the available ROIs are coarse bounding boxes rather than precise lesion masks and is not decisive by itself. More importantly, together with the cross-classifier tests and case studies, it shows that the steering effects concentrate in lesion-containing regions and carry class-relevant information rather than reflecting only global acquisition artifacts or background anatomy. If the features relied only on global characteristics, their effects would not consistently concentrate within lesion ROIs while producing the localized attribute edits observed in the case studies.

Regarding W2: Steering at Matched Effect Magnitudes

Original gain calibration. Raw gain values are not directly comparable across SAE parameterizations. In Figures 4 and 7, we therefore calibrate each method's gain range using random top-KK steering, selecting ranges in which random latents already produce a comparable, non-negligible SSC change. Within these calibrated ranges, the low-entropy directions exhibit different controllability: P-DoD-SAE produces progressively stronger and more consistent semantic effects, whereas the baselines respond weakly or saturate. Thus, the original protocol compares semantic controllability after method-specific gain calibration, while matched effect-L2L_2 provides a complementary image-space magnitude control.

To verify that P-DoD-SAE's advantage persists at matched intervention scales, we compare methods at five matched effect-L2L_2 magnitudes in Table 12. Method-specific gains match every reachable target within 2.2%. Each cell reports effect cosine / SSC LPIPS / SSC SSIM.

MethodL2≈0.0026L_2{\approx}0.0026L2≈0.0093L_2{\approx}0.0093L2≈0.0316L_2{\approx}0.0316L2≈0.1481L_2{\approx}0.1481L2≈0.3474L_2{\approx}0.3474
P-DoD-SAE.2921 / .4042 / .6660.3405 / .4038 / .6678.4215 / .3637 / .6549.5730 / .3221 / .6602.8517 / .2029 / .7565
SAeUron.0604 / .4042 / .6654.0613 / .3946 / .6637.0630 / .3970 / .6648.0674 / .4545 / .6233.0731 / .5065 / .5722
DiffLens.0045 / .4037 / .6654.0046 / .4045 / .6632.0047 / .4123 / .6648unreachableunreachable
  • Effect direction. P-DoD-SAE retains substantially higher effect-direction cosine at every matched magnitude.

  • Main SSC. The methods are similar under weak interventions, while P-DoD-SAE becomes clearly stronger at the two largest reachable magnitudes.

  • Response to gain. SAeUron requires gains several orders of magnitude larger while its effect cosine remains nearly flat, and DiffLens saturates before reaching the two strongest targets. In contrast, P-DoD-SAE reaches all five levels with increasingly aligned effects, consistent with its schedule-normalized, direction-based training target rather than an advantage caused only by larger interventions.

AuthorsOpenReview ↗

Author Response to Reviewer (W3, W4)

Regarding W3: Distinction from Existing Temporal SAE Methods

1. Residualization alone does not remove schedule-dependent scale. Residualized Temporal SAE removes dynamics predicted by learned neighboring-timestep regressors and subsequently standardizes the resulting components using empirical statistics. However, it does not explicitly model the relationship between residual magnitude and the diffusion variance schedule. Our analysis in Appendix B shows that hidden-state residual scales vary systematically with this schedule. Without schedule-aware calibration, residual magnitude can remain an easy timestep-dependent reconstruction signal for the SAE. P-DoD-SAE corrects this by normalizing each coupled local residual using the corresponding schedule ratio before applying a shared sparse dictionary.

2. TIDE introduces temporal modeling inside the SAE. TIDE augments a standard SAE with a trainable timestep-conditioned modulation module. While this increases temporal modeling capacity, the resulting features may depend jointly on the sparse dictionary and the temporal module, making their respective contributions harder to isolate. In contrast, P-DoD-SAE introduces zero additional trainable parameters and leaves the SAE architecture unchanged. Its improvement therefore comes from reformulating the representation being decomposed, preserving the architectural simplicity and direct feature-level interpretation of a standard SAE.

3. We explicitly evaluate whether the resulting interpretation is time-consistent. Existing work primarily studies reconstruction, temporal feature profiles, and qualitative steering. Our evaluation separately tests whether the same sparse directions remain stable in latent space, exhibit consistent visual manifestations across timesteps, and preserve label-predictive information, corresponding to Questions 1–3.

Regarding W4: Construction and Use of Paired Hidden States

1. The paired states follow the same noise-coupled trajectory construction during training and steering. As specified in Appendix B, Algorithm 3, we first sample the earlier state at b=max⁡(t−s,0)b=\max(t-s,0) and then sample the later state from the corresponding transition from bb to tt. The two hidden states therefore share the same underlying image, spatial mask, and diffusion trajectory rather than being obtained from independently noised inputs. Steering uses the same construction, with hbh_b obtained through an additional forward pass of the frozen denoising model.

2. The clean state is the input being interpreted. Our setting is reconstruction-based interpretation of observed medical images, for which the clean image x0x_0 is available as the object of explanation rather than as privileged supervision. When b=0b=0, hb=h0h_b=h_0 is obtained by a frozen-denoiser forward pass on this image. This clean-state reference is intrinsic to DoD, which characterizes the recovery direction from a noised state toward the input image.

3. The comparison methods have access to the same underlying data. All SAE methods are trained on the same images, masks, and frozen denoising model; importantly, the uniformly sampled baseline SAE training also includes the t=0t=0 state. P-DoD-SAE uses hbh_b only to construct its method-specific residual-space representation, whereas the comparison SAEs operate directly on hth_t. Thus, no method receives additional labels or different training samples; the controlled difference is the representation supplied to the SAE.

AuthorsOpenReview ↗

Author Response to Reviewer (Q1, Q2, Q3, Q4)

Regarding Q1: Local-Residual Scale Diagnostic

Appendix E already provides this diagnostic on the natural-image MVTec AD setting: Figure 20 compares the normalized empirical-residual and schedule-derived local-scale curves, while Figure 21 shows their fitted-scale correspondence. Their agreement provides cross-domain evidence that the local-scale relationship is not specific to liver MRI.

We additionally applied the requested diagnostic to the main LLD-MMRI hook layer, middle_block.1. Following Appendix E, Eqs. (27)--(28), for hook-layer activation htℓ(x)h_t^\ell(x), we define the feature-dimension normalizer as mℓ=(CℓHℓWℓ)−1/2m_\ell=(C_\ell H_\ell W_\ell)^{-1/2}. We compute the empirical local residual-scale curve ρsℓ(t)=Ex∼D[mℓ∣∣htℓ(x)−has(t)ℓ(x)∣∣2]\rho_s^\ell(t)=E_{x\sim D}[m_\ell ||h_t^\ell(x)-h_{a_s(t)}^\ell(x)||_2], where as(t)=max⁡(t−s,0)a_s(t)=\max(t-s,0), and compare it with the schedule-derived factor γs(t)=(1−αˉt/αˉa)1/2\gamma_s(t)=(1-\bar{\alpha}_t/\bar{\alpha}_a)^{1/2}, where a=as(t)a=a_s(t). For each interval, we report their Pearson correlation across timesteps and the MSE between the normalized curves:

Interval ssCorrelation ↑\uparrowNormalized MSE ↓\downarrow
500.7020.0366
1000.8020.0260
2000.6750.0415
5000.7660.0375
10000.8780.0270
20000.6190.0795

The results show that the schedule-derived factor captures substantial variation in the local hidden-state residual scale, including at the main setting s=500s=500, directly supporting the normalization used in Eq. (9). The correlations and low normalized MSEs further support its empirical validity. This diagnostic pre-specifies s∈[200,1000]s\in[200,1000] as a practical candidate range, within which we use s=500s=500 as the default setting.

Regarding Q2: Decoder-Column Normalization

All SAE variants use the decoder-column normalization implemented in decoderesearch/SAELens. Each decoder direction is constrained to unit L2L_2 norm throughout training, removing the encoder–decoder scale ambiguity. We will clarify this implementation detail in the paper.

Regarding Q3–Q4: Boundary Rule and Presentation

Thank you for identifying these presentation issues. In the final revision, we will make the boundary rule explicit in Eq. (9) by defining b=max⁡(t−s,0)b=\max(t-s,0) and using hbh_b and αˉb\bar{\alpha}_b, consistent with Algorithm 3. We will also remove revision-history phrasing and duplicated citations, and standardize the relevant terminology to whitening throughout.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.