acceptodds
Under review as a conference paper at ICLR 2027

Beyond SAEs: Consistent Concept Discovery with Sparse Coding

Abstract

Sparse autoencoders (SAEs) explain foundation model representations by decomposing activations into sparse combinations of learned dictionary atoms. However, the learned dictionaries can vary across training runs, limiting the reliability and reproducibility of SAE-based analyses. We propose Consistent Sparse Dictionary Learning (ConSDL), which replaces the standard linear encoder with per-sample sparse coding. This is motivated by our synthetic analysis showing that coefficients from a single linear encoder pass are suboptimal compared with those obtained by per-sample optimization, suggesting that sparse coding is a reliable mechanism for dictionary learning. Across 16 synthetic settings, ConSDL achieves near-perfect dictionary and coefficient recovery. These gains translate to state-of-the-art cross-seed consistency across four vision and language settings, where ConSDL improves dictionary and coefficient consistency by 17.6% and 35.9%, respectively, over the average of six SAE baselines.

Primary area:
interpretability and explainable AI
Keywords:
Interpretability, Sparse Autoencoders, Dictionary Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_BTzPOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a novel SAE architecture for finding stable concepts across runs. The introduce a sparse coding iterative objective, where they switch from optimizing refitting the coefficients of the dictionary and the encoder weights, and also add a covariance regularization. The demonstrate their new architecture on stability and performance measures on a large set of toy as well as language and vision models against several other SAE baselines.

Critical strengths

The paper is well written and generally easy to follow. I appreciate the domain of research aiming to ensure that interpretability methods are stable. The analysis on toy + real models is great and gives a good flow of evidence to the paper. Analysis was done on language and vision models which strengthens the paper significantly.

I also thought the related work, particularly to older sparse coding and dictionary literature, was well done and interesting to read about!

Critical weaknesses

Both excluded baselines DL-FISTA (https://github.com/shrutij01/compositional-sparse-ood) and Archtetypal SAEs have public code releases (https://github.com/KempnerInstitute/overcomplete) and in my knowledge would be very string baselines. I don't think there is any need to 'reproduce' them as stated in Line 1039; the authors should either use the publicly released code on their own datasets, or also run their method (ConSDL) on the datasets/benchmarks released in the above repositories. Additionally, a publicly available Archetypal SAE for DINOv2 can be found at https://huggingface.co/matybohacek/RA-SAE-DINOv2-32k (to at least compare concept quality at inference time). If the baselines perform poorly during training, please simply report what you tried with the hyperparameters and put it in the paper, rather than exclude the evidence and reproducibility details altogether. Overall, it is a bit strange to me that the best performing baselines I was expecting to see (and I personally have experimented with) were omitted from comparison and this needs to be resolved in my opinion with any of the above options.

Questions for authors

What precisely did you try w.r.t DL-FISTA and Archetypal SAE baselines?

In Table 1, the results show that cross-seed consistency does not necessarily correlate with 'quality' of the learned dictionary; for example, when adding similarity reg., the number of duplicates lowers but consistency does as well. How can we know whether the concistency we are getting with ConSDL is 'good' consistency and improving the concepts learned, or actually hurting the dictionary in terms of the concepts learned?

7 comments

Summary

Thank you for the constructive feedback. We hope our responses address the reviewer's concerns.

(W1, Q1) Missing baselines: We add direct comparisons with DL-FISTA and Archetypal SAE (RA-SAE). ConSDL outperforms DL-FISTA across 16 synthetic settings and achieves higher consistency than RA-SAE at matched training epochs across all four backbones.

(Q2) Consistency and concept quality: We provide additional interpretability evaluations and qualitative analyses, alongside existing sparse probing results, showing that ConSDL maintains competitive concept quality while improving consistency. We further show that redundant atoms can inflate consistency scores, explaining why removing redundancy can reduce measured consistency without degrading dictionary quality.

In the revised manuscript, blue text indicates newly added content, while green text indicates revisions to existing content.

We would be happy to provide further clarification if needed.

W1, Q1 - Missing baselines

Summary: We thank the reviewer for their constructive suggestions on how to compare ConSDL with DL-FISTA and Archetypal SAE (RA-SAE). Following these suggestions, we conducted additional experiments and added the results and implementation details to Appendix E.3. ConSDL outperforms DL-FISTA across all 16 synthetic settings (Table 7). For RA-SAE, we investigated preprocessing and relaxation bounds and extended training to 10 epochs for both methods (Table 8). Although RA-SAE improves with longer training, ConSDL achieves higher DCS at matched epochs on all four backbones, while one-epoch ConSDL outperforms 10-epoch RA-SAE on three of four backbones.

DL-FISTA. Our initial submission included a variant of DL-FISTA as the sparse-coding-only ablation (Table 1), retaining its FISTA-based coefficient inference while replacing closed-form dictionary updates with mini-batch gradient-based optimization. This adaptation allowed us to evaluate sparse coding at the scale of our foundation-model activations under the same training protocol as other SAE baselines. However, this variant exhibited substantial dictionary duplication on CLIP (9,270 atoms) and latent inactivity on Pythia (9,767 dead atoms).

Following the reviewer's suggestion, we directly evaluated DL-FISTA on our synthetic benchmark using the same 5M-sample training budget. ConSDL achieved higher reconstruction quality across all 16 settings, improved ground-truth dictionary and coefficient recovery, and achieved higher cross-seed consistency. It also exhibited substantially fewer duplicate atoms (0 vs. 391) and dead atoms (18 vs. 2,125) on average. We have added the full results in Table 7 and implementation details to Appendix E.3.

MethodR² ↑GT DCS ↑GT MCC ↑Cross-seed DCS ↑Cross-seed MCC ↑# Dup ↓# Dead ↓
ConSDL0.9480.9740.9670.9690.965018
DL-FISTA0.6460.7510.7040.7230.6903912,125

Archetypal SAE. Our initial experiments with our reimplementation of Archetypal SAE yielded poor reconstruction or substantial latent inactivity and duplication. Following the reviewer's feedback, we revisited the official implementation and evaluated several preprocessing strategies and relaxation bound (δ), as the original paper specified element-wise standardization but did not fully detail the preprocessing procedure. We found that mean centering followed by per-token L2 normalization gave the best overall results. The table below reports these results on CLIP with 4× dictionary expansion and ℓ₀ ≈ 32.

PreprocessingδR² ↑#Dead ↓#Dup ↓
Z-score10.31801,949
Mean centering10.57501,462
Mean centering50.6610374
Mean centering100.8192300
Global scaling10.4352,6177
Centering + token L20.10.8391751,311
Centering + token L20.50.889105339
Centering + token L210.9000429

Since the publicly available DINOv2 checkpoint has only one seed, we instead trained independent models using the above setup to evaluate cross-seed consistency.

In our original experiments, ConSDL and all SAE baselines were trained for one epoch. Since the original RA-SAE study used substantially longer training (50 epochs), we extended both RA-SAE and ConSDL training to 10 epochs across four backbones (4× expansion, ℓ₀ ≈ 32). Although RA-SAE improved with longer training, one-epoch ConSDL still achieved higher DCS and MCC than 10-epoch RA-SAE on three of four backbones. At matched 10 epochs, ConSDL also achieved higher DCS on all backbones, with slightly lower MCC on DINOv2. Full results are provided in Appendix E.3 (Table 8).

BackboneMethod1 ep. (DCS / MCC)10 ep. (DCS / MCC)
CLIPRA-SAE0.644 / 0.3540.806 / 0.792
ConSDL0.852 / 0.8140.908 / 0.875
DINOv2RA-SAE0.730 / 0.7010.922 / 0.905
ConSDL0.905 / 0.8660.935 / 0.896
PythiaRA-SAE0.763 / 0.7120.821 / 0.791
ConSDL0.881 / 0.8130.887 / 0.821
GPT-2RA-SAE0.807 / 0.7580.850 / 0.807
ConSDL0.887 / 0.8350.898 / 0.850

Q2 Cross-seed consistency and concept quality

Summary: We agree that higher consistency alone does not guarantee better concepts. To address this concern, we provide three complementary analyses:

  1. Quantitative evaluations show that ConSDL maintains competitive concept detection and atom interpretability (Sec. 5.3, Appendix E.5).
  2. Qualitative analyses suggest that highly consistent atoms tend to capture clearer concepts, while ConSDL discovers diverse, fine-grained concepts (Sec. 6, Appendix E.6).
  3. Similarity regularization analysis shows that redundant atoms can inflate consistency scores, explaining why removing duplicates reduces measured consistency (Sec. 5.4, Appendix E.4).

Quantitative evaluation. We evaluate concept quality using two complementary SAEBench benchmarks [1]: sparse probing [2], which measures concept detection using a small number of latents, and AutoInterp [3], which evaluates individual atom interpretability. ConSDL achieves performance comparable to the strongest SAE baselines on both evaluations. At 16× expansion, Top-1 probing scores are 0.52/0.44 on CLIP/DINOv2 and 0.74/0.73 on GPT-2/Pythia (best baselines: 0.49/0.45 and 0.75/0.73). AutoInterp scores are 0.88/0.87 on GPT-2/Pythia (best baselines: 0.90/0.87). These results suggest that improved consistency does not compromise concept detection or atom interpretability (Tables 11–13; Sec. 5.3, Appendix E.5).

EvaluationBackboneBaseline Avg.Best BaselineConSDL
Sparse Probing (Top-1)CLIP0.420.490.52
Sparse Probing (Top-1)DINOv20.400.450.44
Sparse Probing (Top-1)GPT-20.730.750.74
Sparse Probing (Top-1)Pythia0.710.730.73
AutoInterpGPT-20.880.900.88
AutoInterpPythia0.860.870.87

Qualitative analysis. Our qualitative inspection suggests that low-consistency atoms often exhibit noisy or difficult-to-interpret activation patterns, while highly consistent atoms tend to capture clearer concepts (Fig. 9). We also observe that ConSDL learns distinct, fine-grained concepts, such as bottle necks, bottle shapes, and beer in glasses (Fig. 4), or baseball team names, pitching, and field positions (Fig. 10). To examine whether this pattern extends beyond individual examples, we analyze cross-method atom matching: 38% and 21% of ReLU atoms on CLIP and Pythia match multiple ConSDL atoms, compared with only 4.2% and 11% in the reverse direction (Figs. 11–12). Additional examples and analyses are provided in Sec. 6 and Appendix E.6.

Similarity regularization. Table 1 shows that similarity regularization substantially reduces duplicate atoms but also lowers cross-seed consistency. To investigate this, we analyzed the activation patterns of highly similar atoms in +Refitting (without regularization) and found substantial redundancy (Table 9). For +Refitting, removing these duplicates reduces DCS from 0.85 to 0.78 on CLIP and from 0.84 to 0.79 on Pythia, while Full ConSDL remains largely unaffected (Fig. 8). This suggests that redundant atoms can inflate consistency scores, motivating our joint evaluation of consistency and dictionary utilization rather than consistency alone (Appendix E.4).

[1] Karvonen et al. (2025). SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability. ICML 2025.

[2] Kantamneni et al. (2025). Are Sparse Autoencoders Useful? A Case Study in Sparse Probing. ICML 2025.

[3] Paulo et al. (2025). Automatically Interpreting Millions of Features in Large Language Models. ICML 2025.

Reviewer_n3TpOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Consistent Sparse Dictionary Learning (ConSDL) to improve the consistency of sparse autoencoder features across training runs. The method replaces the learned encoder with per-sample sparse coding and combines it with coefficient refitting and co-active similarity regularization. The authors evaluate ConSDL on 16 synthetic settings and foundation-model activations from CLIP, DINOv2, GPT-2, and Pythia. The results show improved dictionary recovery and cross-seed consistency compared with existing SAE methods.

Critical strengths
  1. The paper addresses an important problem with sparse autoencoders. Different training runs can produce different dictionaries from the same data, which makes it difficult to use the learned features for reliable model interpretation.
  2. The experimental evaluation is fairly extensive. The authors test different dictionary sizes and atom-selection distributions on synthetic data, and evaluate the method on both vision and language models with different sparsity levels.
  3. The ablation experiments are useful for understanding the method. They show that coefficient refitting improves reconstruction and reduces dead atoms, while similarity regularization helps reduce duplicated dictionary atoms.
Critical weaknesses
  1. Per-sample sparse coding and alternating dictionary optimization have been studied extensively, and several recent methods already use iterative coefficient estimation. The paper does not clearly establish what is fundamentally new beyond combining these existing methods.
  2. The improvement over the strongest baselines is relatively small for dictionary consistency. Although the paper reports a 17.6% improvement over the average of six SAE baselines, the improvement over the strongest baseline is only 1.9%. The main results would be more convincing if the comparisons focused on the strongest existing methods.
  3. The ablation results show that combining all components does not always improve consistency. In Table 1, the full method achieves lower dictionary and coefficient consistency than the refitting-only variant on both CLIP and Pythia, despite improving dictionary utilization. This trade-off deserves more discussion.
  4. The computational overhead is substantial. At 16x expansion, ConSDL requires roughly ten times the training latency of a standard ReLU SAE in the reported measurements. The cost analysis also uses fixed iteration counts that differ from some experimental configurations, making it difficult to assess the full computational cost of the reported results.
Questions for authors

Please see weakness.

5 comments

Summary

Thank you for the constructive feedback. We hope our responses address the reviewer's concerns.

(W1) Novelty: We clarify our methodological contribution to consistent dictionary learning and provide additional experimental evidence demonstrating its benefits over existing sparse coding methods.

(W2) Small improvement over the strongest baseline: Although the DCS improvement over MP-SAE is modest (1.9%), MP-SAE has substantially lower MCC (0.49 vs. 0.78) and dictionary utilization (0.76 vs. 1.00). ConSDL also improves MCC by 13.3% over the strongest baseline while maintaining nearly complete dictionary utilization.

(W3) Consistency–redundancy trade-off: We show that redundant atoms can inflate consistency scores, explaining the observed trade-off, while ConSDL maintains high consistency with minimal redundancy.

(W4) Computational cost: We update the cost analysis and demonstrate the practicality of ConSDL through measured training runtimes.

In the revised manuscript, blue text indicates newly added content, while green text indicates revisions to existing content.

We would be happy to provide further clarification if needed.

W1 - Novelty

We agree that per-sample sparse coding and alternating dictionary optimization are well-established techniques in other contexts. However, applying sparse coding as a practical alternative to SAEs remains challenging. To the best of our knowledge, ConSDL is the first sparse coding-based dictionary learning method to demonstrate improved cross-seed dictionary and coefficient consistency across both vision and language models while maintaining competitive reconstruction quality. We achieve this by introducing coefficient refitting and co-active similarity regularization to address coefficient shrinkage, dead atoms, and dictionary redundancy.

Our sparse-coding-only ablation (Table 1) demonstrates the limitations of applying sparse coding alone, producing 9,270 duplicate atoms on CLIP and 9,767 dead atoms on Pythia. We further compare ConSDL with DL-FISTA [1], a sparse dictionary learning baseline, across 16 synthetic settings (new Table 7). ConSDL achieves higher reconstruction quality across all 16 settings, improves ground-truth dictionary and coefficient recovery, and achieves higher cross-seed consistency. It also exhibits substantially fewer duplicate atoms (0 vs. 391) and dead atoms (18 vs. 2,125) on average.

MethodR² ↑GT DCS ↑GT MCC ↑Cross-seed DCS ↑Cross-seed MCC ↑# Dup ↓# Dead ↓
ConSDL0.9480.9740.9670.9690.965018
DL-FISTA0.6460.7510.7040.7230.6903912,125

[1] Pacela et al. (2026). Stop probing, start coding: Why linear probes and sparse autoencoders fail at compositional generalisation.

W2 - Small improvements compared to strongest baseline

We agree that comparisons with the strongest baselines are important. While ConSDL's DCS improvement over MP-SAE, the strongest baseline in DCS, is modest (1.9%), MP-SAE has substantially lower coefficient consistency (MCC: 0.49 vs. 0.78) and dictionary utilization (0.76 vs. 1.00). TopK achieves the highest MCC among baselines (0.69 vs. ConSDL's 0.78) but lower DCS (0.73 vs. 0.84). These results show that ConSDL achieves stronger consistency in both dictionaries and coefficients while maintaining nearly complete dictionary utilization. We already report comparisons with the strongest baselines in the main results (Table 6) and have further revised the Abstract to highlight them alongside the baseline-average improvements.

W3 - Consistency–redundancy trade-off

The reviewer questions why Full ConSDL achieves lower cross-seed consistency than the variant without similarity regularization (+Refitting) despite substantially reducing duplicate atoms (Table 1). Our additional analyses show that highly similar atoms in +Refitting exhibit redundant activation patterns (Table 9). For +Refitting, removing these duplicates reduces DCS from 0.85 to 0.78 on CLIP and from 0.84 to 0.79 on Pythia, while Full ConSDL remains largely unaffected (Fig. 8). These results suggest that redundant atoms can inflate consistency scores, motivating our joint evaluation of consistency and dictionary utilization rather than consistency alone. We have added these analyses to the revised Section 5.4 and Appendix E.4.

W4 - Computational cost

Summary: We updated the cost analysis in Appendix Tables 2 and 3 using the actual experimental hyperparameters and additionally measured total training-loop time across all evaluated models (revised Fig. 7). ConSDL requires 0.85–2.77 hours per seed on a single H100 GPU, corresponding to 1.7–4.2× the baseline average across expansion factors. Although SAE-specific computation incurs substantial overhead, its relative impact on total training time is reduced by the shared cost of backbone forward passes and other operations, making ConSDL computationally feasible at the evaluated scale.

ConSDL remains computationally feasible despite higher overhead. We acknowledge that ConSDL incurs computational overhead, as discussed in our Limitations section. However, SAE-specific computation represents only part of the overall training cost, which also includes frozen backbone forward passes, data loading, and other operations. To assess practical training costs, we measured training-loop times (h:mm) across all evaluated models and baselines on a single H100 80GB GPU, averaged across four backbones and three sparsity levels (ℓ₀ ≈ 32/64/128).

ExpansionReLUBaseline Avg.ConSDL
4×0:250:310:51
8×0:240:311:23
16×0:290:402:46

The relative overhead in total training time is smaller than that in SAE-specific computation because backbone forward passes and other operations contribute to the overall cost. For inference, ConSDL takes 112.8 vs. 50.3 seconds on entire ImageNet validation 50K images and 211.7 vs. 78.0 seconds on FineWeb 20.48M tokens, compared with the baseline average. Despite the computational trade-off, these absolute runtimes remain practical for the evaluated settings. Improving sparse coding efficiency remains an important direction for future work.

Fixed iteration counts We initially used fixed iteration counts (30 FISTA and 5 refitting iterations) to compare computational scaling across dictionary sizes and backbones under a common solver budget. We agree that using the actual experimental configurations provides a more accurate cost assessment. We have therefore updated Appendix Tables 2 and 3 using the hyperparameters of the reported checkpoints. As shown in the end-to-end training analysis above, ConSDL remains computationally feasible under these configurations.

Thanks for your response.

The results address several of my concerns, however, ConSDL mainly combines with coefficient refitting and similarity regularization. It would be useful to see whether improvements hold on foundation-model activations.

Another concern is the computation overhead. The provided ConSDL requires over 4x time cost at 16x expansion. What if comparing ConSDL with other methods in the same time cost?

Reviewer_Q5QhOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses the consistency issue in existing sparse autoencoders (SAEs) and proposes ConSDL, which replaces the shared linear encoder with a per-sample sparse coding pipeline. Specifically, ConSDL introduces coefficient refitting to mitigate shrinkage bias and co-active similarity regularization to reduce redundancy among dictionary atoms. These designs improve the consistency of learned dictionaries and sparse coefficients across independent training runs. Experiments on vision and language foundation models demonstrate the effectiveness of the proposed method. The authors also provide a detailed computational cost analysis in the appendix.

Critical strengths
  1. The most interesting aspect of this paper is the introduction of per-sample optimization, which enables the model to obtain more accurate, sample-specific sparse codes under the current dictionary, potentially improving the quality of SAE learning.

  2. The combination of sparse coding, coefficient refitting, and co-active similarity regularization addresses three complementary challenges: coefficient estimation, shrinkage bias, and dictionary redundancy. Each component serves a distinct purpose and contributes to the overall effectiveness of the proposed method.

  3. Extensive experiments across both vision and language settings demonstrate the effectiveness of ConSDL, particularly in improving dictionary and coefficient consistency across independent training runs.

Critical weaknesses
  1. The biggest concern comes from the substantial computational overhead introduced by per-sample FISTA optimization. Replacing a single encoder forward pass with iterative sparse coding significantly increases both training and inference costs, potentially limiting the practicality of ConSDL for large-scale SAE training and deployment.

  2. The co-active similarity regularizer assumes that highly similar atoms activated by the same sample are likely to be redundant. However, semantically related or complementary features may naturally exhibit similar directions and co-activate within the same sample. Penalizing such atoms could inadvertently suppress meaningful feature distinctions and limit the representational capacity of the learned dictionary.

  3. The paper primarily focuses on dictionary-level metrics, such as dictionary consistency and coefficient consistency. While these metrics are relevant to the paper's central objective, they may not fully reflect the practical interpretability and downstream utility of the learned features. Incorporating additional evaluations from SAEBench [1] would provide a more comprehensive assessment of ConSDL and strengthen the empirical evidence for its effectiveness.

Reference

[1] SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability. https://arxiv.org/abs/2503.09532

Questions for authors

Please refer to the Weakness section.

4 comments

Summary

Thank you for the constructive feedback. We hope our responses address the reviewer's concerns.

(W1) Computational cost: We clarify the computational overhead and demonstrate the practicality of ConSDL through measured training runtimes.

(W2) Risk of meaningful concept suppression by the regularizer: We show that the regularizer suppresses near-duplicate atoms while preserving the learned representations, rather than discarding them.

(W3) Interpretability and utility: We provide additional automated interpretability evaluations and qualitative examples, alongside existing sparse probing results, showing that ConSDL maintains competitive concept detection and interpretability while improving consistency.

In the revised manuscript, blue text indicates newly added content, while green text indicates revisions to existing content.

We would be happy to provide further clarification if needed.

W1 - Computational cost

Summary: Despite higher SAE-specific costs, ConSDL requires 0.85–2.77 hours per seed on a single H100 GPU, corresponding to 1.7–4.2× the baseline average across expansion factors. These results demonstrate that ConSDL remains computationally feasible on the evaluated vision and language models. We have added these results in Appendix C and Fig.7.

We acknowledge that ConSDL incurs computational overhead, as discussed in our Discussion and Limitations section. However, SAE-specific computation (Table 2) represents only part of the overall training cost, which also includes frozen backbone forward passes, data loading, and other operations. To assess practical training costs, we measured training-loop times (h:mm) across all evaluated models and baselines on a single H100 80GB GPU, averaged across four backbones and three sparsity levels (ℓ₀ ≈ 32/64/128).

ExpansionReLUBaseline Avg.ConSDL
4×0:250:310:51
8×0:240:311:23
16×0:290:402:46

The relative overhead in total training time is smaller than that in SAE-specific computation because backbone forward passes and other operations contribute to the overall cost. For inference, ConSDL takes 112.8 vs. 50.3 seconds on entire ImageNet validation 50K images and 211.7 vs. 78.0 seconds on FineWeb 20.48M tokens, compared with the baseline average. Despite the computational trade-off, these absolute runtimes remain practical for vision and language foundation models. Improving sparse coding efficiency remains an important direction for future work.

W2 - Risk of meaningful concept suppression by the regularizer

Summary: To address concerns about suppressing meaningful feature distinctions, we provide additional analyses showing that highly similar atoms exhibit redundant activation patterns, which are effectively suppressed by regularization (Table 9), and their underlying representations are largely preserved despite the reduction in duplicate atoms (Table 10). These analyses are included in the revised Sec. 5.4 and Appendix E.4.

Our co-active similarity regularizer penalizes the most similar co-active atoms when their cosine similarity exceeds a margin of 0.5, as now clarified in Appendix B. To examine whether highly similar atoms are actually redundant, we compare token-level activation overlap (Jaccard similarity, J) and coefficient correlation (r) between +Refitting (without regularization) and Full ConSDL, using 16× expansion and ℓ₀ ≈ 128 for both CLIP and Pythia.

Decoder cosineCLIP +Refitting (# Pairs / J / r)CLIP Full (# Pairs / J / r)Pythia +Refitting (# Pairs / J / r)Pythia Full (# Pairs / J / r)
0.9–1.04.2K / 83% / 0.930623 / 64% / 0.850
0.5–0.611.9K / 21% / 0.1828.8K / 29% / 0.3212.2K / 17% / 0.0412.6K / 16% / 0.16
0.0–0.136.3M / 0% / −0.2936.0M / 1% / −0.2733.3M / 0% / −0.3233.1M / 0% / −0.24

Without regularization, highly similar atoms exhibit strongly redundant activation patterns. For pairs with cosine ≥ 0.9, Jaccard overlap and coefficient correlation reach 83% and 0.93 on CLIP, and 64% and 0.85 on Pythia, respectively. Both metrics decrease consistently as cosine similarity decreases. In contrast, Full ConSDL has no atom pairs with cosine ≥ 0.6. The remaining moderately similar pairs (cosine 0.5–0.6) exhibit substantially lower activation overlap and correlation (29% / 0.32 on CLIP and 16% / 0.16 on Pythia). Full results are provided in Appendix E.4 (Table 9).

Importantly, the concepts represented by these duplicate atoms are largely preserved after regularization. Our matching analysis shows that near-duplicate groups in +Refitting can be matched to individual Full ConSDL atoms with high decoder similarity and activation correlation (Appendix E.4, Table 10). Redundancy is thus reduced by representing a concept once rather than many times, not by discarding it. Furthermore, our concept-quality evaluations in W3 show that this reduction in redundancy does not substantially compromise concept detection or interpretability.

W3 - Interpretability and utility

Summary: We agree that consistency alone does not establish concept quality or downstream utility. To address this, we highlight our existing sparse probing results for concept detection and add interpretability evaluations (Sec. 5.3, Appendix E.5 Table 11-13). ConSDL achieves performance comparable to the strongest SAE baselines on both evaluations, suggesting that improved consistency does not compromise concept detection or atom interpretability. Qualitative analyses further show that ConSDL learns diverse, fine-grained concepts (Sec. 6, Appendix E.6.2).

Sparse probing. Our original submission already included Top-1/Top-5 sparse probing following the SAEBench [1] evaluation framework, using 113 language tasks from SAE-Probes [2] and extending the evaluation to six vision datasets (Tables 11-12). This evaluation measures how effectively a small number of learned latents can predict task labels. At 16× expansion, ConSDL achieves Top-1 scores of 0.52 and 0.44 on CLIP and DINOv2 (best baselines: 0.49 and 0.45), and 0.74 and 0.73 on GPT-2 and Pythia-160M (best baselines: 0.75 and 0.73). We moved these results earlier in the revised Sec. 5.3 to better highlight the downstream utility of learned concepts.

AutoInterp. We additionally evaluated atom interpretability using SAEBench AutoInterp [3] (Table 13), which measures how accurately natural-language descriptions generated from activating examples predict atom activations on held-out examples. We followed the SAEBench evaluation protocol, randomly sampling 1,000 active atoms per SAE on GPT-2 Small and Pythia-160M, using Claude Sonnet 5.5 for explanation generation and scoring. ConSDL achieves scores of 0.88 and 0.87, respectively, close to the strongest baselines (0.90 and 0.87).

EvaluationBackboneBaseline Avg.Best BaselineConSDL
Sparse Probing (Top-1)CLIP0.420.490.52
Sparse Probing (Top-1)DINOv20.400.450.44
Sparse Probing (Top-1)GPT-20.730.750.74
Sparse Probing (Top-1)Pythia0.710.730.73
AutoInterpGPT-20.880.900.88
AutoInterpPythia0.860.870.87

Qualitative evidence. Our qualitative inspection suggests that low-consistency atoms often exhibit noisy or difficult-to-interpret activation patterns, while highly consistent atoms tend to capture clearer concepts (Fig. 9). We also provide examples showing that ConSDL discovers more fine-grained concepts than baselines such as bottle necks, bottle shape, and beer in glasses (Fig. 4), or baseball team names, pitching, and field positions (Fig. 10). To examine whether this pattern extends beyond individual examples, we analyze cross-method atom matching: 38% and 21% of ReLU atoms on CLIP and Pythia match multiple ConSDL atoms, compared with only 4.2% and 11% in the reverse direction (Figs. 11–12).

The sparse probing, AutoInterp, and qualitative results suggest that ConSDL improves consistency while maintaining competitive concept detection and interpretability, and supporting diverse, fine-grained representations. Additional details are provided in Sec. 6 and Appendix E.6.2.

[1] Karvonen et al. (2025). SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability. ICML 2025.

[2] Kantamneni et al. (2025). Are Sparse Autoencoders Useful? A Case Study in Sparse Probing. ICML 2025.

[3] Paulo et al. (2025). Automatically Interpreting Millions of Features in Large Language Models. ICML 2025.

AuthorsOpenReview ↗

Gentle reminder for discussion

Dear Reviewer,

Thank you again for your thoughtful review. We have addressed your concerns in our rebuttal and would appreciate it if you could let us know whether our responses have resolved them. If you have any remaining questions or concerns, we would be happy to address them during the discussion period.

Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.