acceptodds
Withdrawn from ICLR 2027

Diagnosing and Improving Latent Recovery in Sparse Autoencoders

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Sparse autoencoders (SAEs) are used to extract interpretable features from neural representations. Their codes assign feature amplitudes used to rank activating examples, attribute model behavior, and set intervention strengths. These applications make amplitude fidelity relevant once feature directions and scales have been fixed. However, reconstruction objectives do not directly measure coefficient accuracy, and interference between co-active features can cause sample-dependent errors that fixed feature-wise rescaling cannot generally remove. Understanding these errors is therefore important for interpreting the codes produced by SAE encoders. We study coefficient recovery in a blind sparse mixing model and characterize amplitude bias in one-pass tied encoders, even with an exact dictionary and correct support. We then compare each code with the code obtained after reconstruction and re-encoding. For tied and untied SAEs, we derive local bounds relating this self-consistency residual to coefficient error under correct and stable support, value-preserving activations, and suitable conditioning. This analysis motivates a self-consistency regularizer that requires one additional encoder evaluation during training and adds no parameters or inference cost. Controlled synthetic experiments show reduced latent recovery error, while CE-bench and SAEBench evaluations provide complementary evidence of improved feature quality.

Primary area:
interpretability and explainable AI
Keywords:
Sparse Autoencoders, Latent Code Recovery, Mechanistic Interpretability

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_MKxqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SAE codes assign feature amplitudes, and people use those amplitudes to rank activating examples, attribute model behavior, and set intervention strengths. Reconstruction objectives do not directly measure whether those amplitudes are accurate. Interference between co-active features can cause sample-dependent amplitude errors that a fixed per-feature rescaling cannot generally remove. The paper asks how to diagnose and reduce amplitude errors that remain even after the dictionary and support (which features are active) are recovered correctly. Contributions: They show that one-pass tied SAE encoders produce sample-dependent amplitude errors even when the dictionary and active support are exactly correct. This separates amplitude errors from dictionary and support errors. They prove that the change in a code after decoding and re-encoding (the self-consistency residual) controls its amplitude error. For tied and untied SAEs, they give local bounds that account for decoder mismatch, observation noise, and conditioning. They propose a self-consistency penalty that can be added to standard SAE training. It costs one extra encoder pass during training, adds no parameters or inference cost, and has a masked variant for TopK encoders. In controlled tied, untied, and fixed-decoder synthetic settings, latent recovery error goes down, with dictionary, support, and amplitude recovery measured separately.

Critical strengths
  1. Authors provide a clear way to detect amplitude errors while keeping the scope honest 4.1: The paper shows that a tied SAE can get the features right and pick the right active features, yet still report wrong strengths. Overlap between co-active features causes this, and the error changes from input to input. 4.3. Decoding a code and encoding it again changes the code by an amount tied directly to the amplitude error. It needs no ground truth, so it can be applied to SAEs trained on real models. 4.4 . The authors say plainly that the bounds hold at fixed parameters and that training effects are tested empirically.

2.The proposed method is easily implemented and tested in controlled settings 5. The regularizer adds one encoder pass during training and nothing at inference. There are no new parameters and no architecture change. 6.1 Different error types are reported on their own, so the effect on amplitudes isn’t hidden in one aggregate number. 6.2 The fixed-decoder control keeps the dictionary exact and the support correct

Critical weaknesses
  1. The paper doesn’t show the regularizer beats alternatives As i understand it, the authors claim any benefit must come from training dynamics, which they don’t explore. Baseline comparisons that would fix this weakness are not there: Least-squares refit on the SAE’s selected features. O’Neill et al. (2025) is cited in Sec. 2 but not compared against

  2. The results are weak in some cases 6.2 The synthetic setup is idealized. The data is noiseless.The SAE appears to use the true sparsity level k. The decoder starts near the true dictionary. F.1 The improvements are thinner in over complete settings It’s also unclear whether the baselines are trained to convergence. If not, the large untied gains may reflect faster optimization rather than a better solution. 6.3 Amplitude accuracy is the paper’s whole motivation , yet the GPT-2 experiments never measure it. CE-Bench measures how interpretable features are, not whether their amplitudes are accurate. So even a real gain doesn’t test the paper’s main claim.

Questions for authors
  1. Can you show that SC improves amplitude accuracy on real model activations, not only interpretability scores? Report the SC residual and E_coef on the GPT-2 SAEs, with and without SC

  2. Do you see the same gains as in the synthetic setting against trained baselines? Does the CE-Bench gain hold against a properly trained baseline?

Reviewer_4heKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies amplitude errors in SAE codes after dictionary and support recovery. It relates a decode-encode self-consistency residual to a least-squares coefficient gap under local assumptions, then uses this residual as a training penalty. Synthetic controls and GPT-2 evaluations test whether the penalty improves recovery and feature quality.

Critical strengths
  1. The distinction between dictionary, support, and amplitude recovery is useful. The ranking-reversal example explains why fixed rescaling cannot repair every amplitude error. The masked TopK identity also makes the analysis more relevant to the implemented objective.

  2. The fixed-decoder control reports a 69.62% latent-error reduction with exact support recovery across ten paired seeds. This is more direct evidence for coefficient refinement than reconstruction metrics alone. The paper also acknowledges that its identities do not guarantee improvement during training.

Critical weaknesses
  1. The untied objective has an unresolved scaling ambiguity. For signed TopK, replacing (E, b, D) with (cE, cb, D/c) for positive c preserves supports and reconstruction but multiplies the masked self-consistency loss by c². Decoder-normalized feature amplitudes remain unchanged. Sections 5 and E.3 do not specify a training constraint that prevents this. The fixed-decoder control is unaffected, but the jointly trained results need this detail before a smaller penalty can be interpreted as better recovery.

  2. The experiments do not yet establish a robust practical benefit. Appendix E.1 omits key synthetic dimensions and optimization settings, although Appendix D.3 shows that the fixed-decoder objectives share a minimizer. The largest relative improvements therefore need absolute errors and convergence evidence. The explicitly overcomplete control improves latent error by only 1.15%. On GPT-2, probing gains are 0.024 to 0.239 percentage points without paired uncertainty. The CE-bench improvement accompanies a dead-feature reduction from 0.1599 to 0.0286, leaving the contribution of amplitude refinement unclear. The AutoInterp checkpoint standard deviations likewise do not establish uncertainty in paired improvements.

Questions for authors
  1. What training constraint prevents the untied scaling transformation? Is decoder normalization enforced after each update?

  2. Can you provide the missing synthetic settings, absolute errors, and convergence diagnostics, together with paired downstream results? Do the coefficient diagnostics and CE-bench gains persist when feature usage is comparable between conditions?

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

This paper shows that, because of co-active features, sparse autoencoders can distort feature amplitudes even when the dictionary and active support are correct. It proposes a self-consistency regularizer that can be added to standard SAEs without adding to its inference cost. The experimental evaluation shows a decrease in coefficient level error in synthetic settings and modest changes on CE-bench and SAEBench.

Critical strengths
  • The topic of the paper is important and the research goal is proposed in a clear manner.
  • The theoretical results presented seem to be correct and they state assumptions clearly.
  • The proposed method doesn't add cost to the inference stage.
Critical weaknesses
  • Table 9 presents the results of AutoInterp across multiple weights and aggregated over 3 training seeds. It is stated that at a λ=10−5\lambda = 10^{-5} the mean change was of 0.00505, while the standard deviation across different seeds of the scores is about 0.02, roughly 4x larger. No standard deviation or confidence interval is reported for the changes, making it unclear whether this improvement exceeds seed variance.
  • Table 7 reports the mean percentage-point changes against unregularized checkpoints aggregated across four seeds. The datapoints don't appear to be associated with the values of λ\lambda. Presenting the standard deviation of this measure would help to defend the claim that the weights give relevant positive changes and are not caused by simple variance in the runs. Although all the values are positive in this table, this can be caused by a low scoring baseline seed.
  • Part of the motivation of this paper was that coefficient errors can even reverse activation rankings due to co-active features. However, no ranking fidelity experiments are performed even though the ground-truth is known in the synthetic setup.
  • Section 3.1 states that overcomplete dictionaries (dm>dpd_m > d_p) are allowed. Appendix F.1 shows only a reduction of about 1% in the latent error for a dm/dpd_m/d_p ratio of 2. It would be important to explore how the proposed method behaves for higher (and more realistic) dm/dpd_m/d_p ratios (e.g. 1 (as a reference), 4, 16, 64). Also, d_m and d_p can be reported for other experiments as well, such as the ones in Figures 1-3.
Questions for authors
  • What is the reason for changing in the parameters (K from 32 to 8, λ\lambda from 0.50.5 to 10−510^{-5}) used for CE-bench and SAEBench evaluations?

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

Related papers

Discussion (0)

Sign in to comment.