acceptodds
Withdrawn from ICLR 2027

PARSE: An Autonomous Algorithm for Mechanistic Interpretability Based on Dynamic, Adaptive Sparse Autoencoders

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

The resolution of semantic superposition in Large Language Models (LLMs) currently relies on sparse autoencoders (SAEs) optimized via static regularization and heuristic warm-up schedules. These static schedules often force concepts into entangled representations and induce semantic hallucination. We propose that representation collapse is not an inherent flaw of SAEs, but an artifact of static latent expansion which lacks quantitative feedback on the condition of the feature covariance matrix. In this paper we present a novel autonomous algorithm that eliminates heuristic warm-up schedules by allowing sparsity to emerge from the dynamic linear algebraic properties of the covariance matrix under the action of the optimizer. PARSE (Parabolic Adaptive Relaxation for Sparse Encoding) continuously monitors the geometric strain of the latent manifold via a matrix-free Krylov approximation of the condition number. When local semantic cross-talk exceeds the LLM's inherent noise floor, the algorithm dynamically spawns orthogonal vectors strictly aligned with the principal eigenvectors of the local residual error matrix. By temporarily substituting the penalty with a soft orthogonalization penalty, PARSE absorbs dense concepts and safely recycles dead latents without inducing representation shock. This localized expansion reduces the computational profile from quadratic to linear scaling, eliminating GPU bottlenecks and reducing overhead by over 90%. We validate PARSE on the 4096D residual stream of Llama-3(8B) across tokens and demonstrate that complex semantic superposition can be isolated with low error under the constraint of strict numerical orthogonality. We further demonstrate that feature death and representation collapse are distinct phenomena. PARSE thus establishes a scalable, autonomous foundation for mechanistic interpretability and the translation of emergent frontier model concepts.

Primary area:
interpretability and explainable AI
Keywords:
Mechanistic Interpretabilty, Sparse Autoencoders, Representation collapse, Autonomous topology, Linear algebra

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_FwQFOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This work proposes PARSE as an algorithm to eliminate heuristic warmup schedules in SAE training.

Critical strengths

None

Critical weaknesses
  1. The paper proposes an improvement to SAE training yet has almost no empirical validation of it's claims with respect to reasonable baselines. The only empirical results in Table 2 are missing the key "hallucinations" column mentioned in the main text.

  2. There is very little discussion or comparison different kinds of SAEs or engagement with the literature.

  3. Proposing PARSE as a tool that should be used when training SAEs would require much more extensive empirical results including comparing to other static SAEs on existing benchmarks like SAEBench and a closer analysis of the change in computational cost of using PARSE with respect to static baselines at different dictionary sizes.

Questions for authors

The main thing I would need to see to improve my score are the missing empirical results discussed in weaknesses 1) and 3). The clarity of the writing should also be improved.

Reviewer_QfX9OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a dynamic, adaptive training algorithm for SAEs. Expansion of the latent space is triggered by a Lanczos-estimated condition number κ together with an L2 strain threshold. On expansion, dead latents are first recycled along the principal residual directions, and new orthogonal dimensions are added only when none remain; training then enters Phase A, which applies only a local soft-orthogonality penalty. And appplies this method on Llama-3 8B, which shows higher L2 error than static SAEs but lower κ, so claim that optimizing L2 alone leads to polysemanticity, and that feature death and representation collapse are distinct phenomena.

Critical strengths

Adaptively determining dictionary capacity is a genuine problem in SAE research, and using online condition-number monitoring as an expansion trigger is a novel and computationally feasible approach. The block-local orthogonality penalty and the residual-aligned initialization of new latents are intuitive. The ablation with and without dead-latent recycling also offers some support for the mechanistic discussion.

Critical weaknesses
  1. The SAEs in the experiments (1536 and 3072 latents) are smaller than the 4096-dimensional input, so they are undercomplete and do not operate in the superposition regime. In the overcomplete setting where SAEs are normally used, κ is necessarily infinite and strict orthogonality is impossible. Moreover, κ reflects only the dictionary’s geometry and does not equate to semantic monosemanticity.
  2. The experiments only use 10⁷ tokens, which is far less than the training dataset of the LLM, and also lack standard downstream evaluations, and support the “hallucination” claim with anecdotal prompts. There is no comparison with closely related methods (Towards Monosemanticity: Decomposing Language Models With Dictionary Learning) (Scaling and evaluating sparse autoencoders) (OrtSAE: Orthogonal Sparse Autoencoders Uncover Atomic Features). The paper also misrepresents the main point of a cited work (Negative Results for Sparse Autoencoders On Downstream Tasks and Deprioritising SAE Research).
Questions for authors
  1. The residual that triggers expansion contains both dictionary capacity error and encoder amortization error (Compute Optimal Inference and Provable Amortisation Gap in Sparse Autoencoders). How do the authors show that expansion addresses a capacity deficit rather than a failure of the encoder’s inference?
  2. Dead latents appear to be counted within a single batch, whereas prior work considers a latent dead only if it has not activated over 10 million tokens (Scaling and evaluating sparse autoencoders). Do the conclusions in Sections 4.2 and 4.3 still hold under this standard definition?
Reviewer_aHLbOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes PARSE, a sparse autoencoder training procedure in which the dictionary grows during training whenever the decoder becomes ill-conditioned or the reconstruction error rises. Dead latents are re-initialised along the leading directions of the residual error, new latents are appended when none remain, and a soft-orthogonality penalty on the decoder stabilises each expansion. The method is evaluated on one layer of Llama-3-8B against a fixed-width L1 SAE (Table 2).

Critical strengths

None.

Critical weaknesses

W1. The methodological contribution reduces to a known decoder orthogonality penalty. Stripped of terminology, the change to the SAE objective (Eqs. 1 and 3) is a penalty ||WD^T WD - I||F^2 on the decoder. Penalising the Gram matrix of a dictionary, including the cross-Gram between blocks that the paper calls basis separation, is long established in dictionary learning [1] and deep learning [2]. It has recently been studied for SAEs on LLMs [3], and orthogonal decoder initialisation was even available as a configuration flag in SAELens [4]. The other components also have direct precedents. Re-initialising dead latents towards poorly reconstructed directions corresponds to neuron resampling [5] and to atom replacement in dictionary learning [6]. Growing the dictionary during training corresponds to function-preserving network growth [7]. None of these is compared against, so the paper does not show what PARSE adds.

W2. The paper's description of current SAEs does not reflect the field. The problem is framed around SAEs trained with static L1 regularisation and warm-up schedules. The most widely used sparsifiers are instead TopK [8], BatchTopK [9] and Matryoshka SAEs [10], together with JumpReLU [11]. The paper mentions TopK and JumpReLU (Sections 1 and 4.5) only to dismiss them, and it compares exclusively against an L1 SAE. The motivation therefore concerns a largely superseded baseline and does not transfer to current practice.

W3. The dead-latent problem is presented as open although it is already addressed. Dead latents under L1 training are well known, and standard remedies exist, including neuron resampling [5], the AuxK loss of Gao et al. [8], and initialising the decoder as the transpose of the encoder [12]. None of these serves as a baseline. The paper supports the problem with Smith et al. [13], a post about SAEs underperforming linear probes on downstream tasks rather than about L1 schedules. A substantial fraction of latents also remains dead under PARSE itself (Table 2).

W4. The evaluation does not use the criteria by which interpretability methods are judged. SAEs are typically evaluated on faithfulness when the reconstruction is spliced into the model, on probing [14], on steering [15] and on causal interventions [16, 17], across several models and layers [18, 19]. The paper instead reports one layer of one model with reconstruction MSE, on which PARSE is clearly worse (Table 2). It adds the condition number of the decoder Gram matrix, L0, dead latents and anecdotal prompt completions. These geometric proxies do not measure interpretability. An untrained Gaussian dictionary of the same width has a lower condition number than every model in Table 2, and the mean cosine in Section 4.4 equals the value expected for random directions in 4096 dimensions.

W5. The standard SAE metrics and baselines are missing. Table 2 reports single points at different L0 values instead of the sparsity-fidelity frontier, so the comparison cannot be interpreted, and loss recovered is absent. The only baseline is a ReLU-L1 SAE without dead-latent treatment. The paper includes no TopK, BatchTopK, JumpReLU or Matryoshka SAEs with AuxK, no resampled ReLU SAEs and no SAEBench evaluation [12]. Training is short, and SAEBench, which the paper cites as justification, in fact reports continued improvement well beyond the budget used here. All results come from a single seed, although SAE features vary substantially across seeds [20].

W6. There is no related-work section. Positioning is limited to a few paragraphs in the introduction, and the bodies of work in W1 to W3 are not discussed. The theory and both toy models are deferred to an external document with the same title, so the paper is not self-contained and the contribution cannot be located relative to prior work.

W7. Central claims are assumptions presented as findings. The abstract states that static schedules force concepts into entangled representations and induce semantic hallucination, and that representation collapse is an artefact of static latent expansion. Neither statement is supported by a citation or an experiment. Semantic hallucination is also undefined for an SAE, whose output is a reconstruction rather than generated text. These premises motivate every component of PARSE, so the method inherits their lack of support.

W8. Vague, non-standard terminology makes the method impossible to follow precisely. The abstract states that orthogonal vectors are spawned when local semantic cross-talk exceeds the LLM's inherent noise floor. Semantic cross-talk is never defined, and spawning simply means appending latents. The noise floor turns out (Section 2.4) to be a cross-entropy that is compared with a Frobenius norm of residuals, two quantities in different units. Representation shock, geometric strain, matrix stiffness and alien minima are likewise used as explanations without definitions. Quantities that drive the algorithm (κ∞, L2,max, the update rule for Λ, γ, ρ and c) are not specified. As written, the claims cannot be tested.

W9. The computational claim is undefined. The abstract claims quadratic-to-linear scaling of the computational profile and an overhead reduction of over 90%, without stating which cost scales with which quantity. Section 2.6 shows that the claim is a count of entries in the orthogonality penalty, with no measured runtime or memory. The full penalty does not require the dh × dh Gram matrix in any case, because ||WD^T WD||F = ||WD WD^T||F allows it to be evaluated at a cost linear in dh.

W10. The submission does not comply with the ICLR 2027 guidelines. The Nomenclature section precedes the references and therefore belongs to the main text. The main text consequently runs onto page 10, beyond the 9-page limit that the guidelines state will be strictly enforced. The mandatory AI-use section is also absent, although the prose strongly suggests heavy use of generative AI.

W11. The reported runs do not implement the described method or test the superposition claim. Appendix A is presented as the code behind Table 2. Its loop optimises only Lrecon plus a fixed L1 term (lines 78 to 81), with no Lortho, no Phase A, no κ trigger and no Ebase stopping rule. Recycling modifies decoder columns but not encoder rows (line 117), which is the reverse of Section 2.3, and spawning uses plain Gaussian initialisation (lines 24 to 34). All runs also use fewer latents than the 4096-dimensional residual stream, whereas superposition requires an overcomplete dictionary. For an overcomplete dictionary, ||WD^T WD - I||F^2 cannot fall below (dh - din) dh / din. The claim that superposition is isolated under strict orthogonality is therefore untested by these experiments, and in the overcomplete regime it is unattainable.

References

[1] I. Ramirez, P. Sprechmann, G. Sapiro. Classification and Clustering via Dictionary Learning with Structured Incoherence and Shared Features. CVPR 2010. [2] N. Bansal, X. Chen, Z. Wang. Can We Gain More from Orthogonality Regularizations in Training Deep Networks? NeurIPS 2018. arXiv:1810.09102. [3] A. Korznikov et al. OrtSAE: Orthogonal Sparse Autoencoders Uncover Atomic Features. ICLR 2026. arXiv:2509.22033. [4] J. Bloom, C. Tigges, A. Duong, D. Chanin. SAELens. https://github.com/decoderesearch/SAELens, 2024 (option decoder_orthogonal_init, versions 3 to 5). [5] T. Bricken et al. Towards Monosemanticity: Decomposing Language Models with Dictionary Learning. Transformer Circuits Thread, 2023. [6] J. Mairal, F. Bach, J. Ponce, G. Sapiro. Online Learning for Matrix Factorization and Sparse Coding. JMLR 11:19-60, 2010. [7] U. Evci, B. van Merriënboer, T. Unterthiner, M. Vladymyrov, F. Pedregosa. GradMax: Growing Neural Networks using Gradient Information. ICLR 2022. arXiv:2201.05125. [8] L. Gao et al. Scaling and Evaluating Sparse Autoencoders. ICLR 2025. arXiv:2406.04093. [9] B. Bussmann, P. Leask, N. Nanda. BatchTopK Sparse Autoencoders. NeurIPS 2024 Workshop. arXiv:2412.06410. [10] B. Bussmann, N. Nabeshima, A. Karvonen, N. Nanda. Learning Multi-Level Features with Matryoshka Sparse Autoencoders. ICML 2025. arXiv:2503.17547. [11] S. Rajamanoharan et al. Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoders. arXiv:2407.14435, 2024. [12] A. Karvonen et al. SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability. ICML 2025. arXiv:2503.09532. [13] L. Smith et al. Negative Results for SAEs on Downstream Tasks and Deprioritising SAE Research (GDM Mech Interp Team Progress Update #2). Alignment Forum, 2025. [14] S. Kantamneni et al. Are Sparse Autoencoders Useful? A Case Study in Sparse Probing. ICML 2025. arXiv:2502.16681. [15] Z. Wu et al. AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders. ICML 2025. arXiv:2501.17148. [16] J. Huang, Z. Wu, C. Potts, M. Geva, A. Geiger. RAVEL: Evaluating Interpretability Methods on Disentangling Language Model Representations. arXiv:2402.17700, 2024. [17] S. Marks et al. Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models. ICLR 2025 (oral). arXiv:2403.19647. [18] T. Lieberum et al. Gemma Scope: Open Sparse Autoencoders Everywhere All at Once on Gemma 2. arXiv:2408.05147, 2024. [19] Z. He et al. Llama Scope: Extracting Millions of Features from Llama-3.1-8B with Sparse Autoencoders. arXiv:2410.20526, 2024. [20] G. Paulo, N. Belrose. Sparse Autoencoders Trained on the Same Data Learn Different Features. ICLR 2026. arXiv:2501.16615.

Questions for authors

How does PARSE differ from the established literature and methods it builds on, namely decoder orthogonality and incoherence penalties [1, 2, 3], dead-latent resampling and auxiliary losses [4, 5], atom replacement in dictionary learning [6], function-preserving network growth [7], and Matryoshka SAEs for the choice of dictionary width [8]?

References

[1] I. Ramirez, P. Sprechmann, G. Sapiro. Classification and Clustering via Dictionary Learning with Structured Incoherence and Shared Features. CVPR 2010. [2] N. Bansal, X. Chen, Z. Wang. Can We Gain More from Orthogonality Regularizations in Training Deep Networks? NeurIPS 2018. arXiv:1810.09102. [3] A. Korznikov et al. OrtSAE: Orthogonal Sparse Autoencoders Uncover Atomic Features. ICLR 2026. arXiv:2509.22033. [4] T. Bricken et al. Towards Monosemanticity: Decomposing Language Models with Dictionary Learning. Transformer Circuits Thread, 2023. [5] L. Gao et al. Scaling and Evaluating Sparse Autoencoders. ICLR 2025. arXiv:2406.04093. [6] J. Mairal, F. Bach, J. Ponce, G. Sapiro. Online Learning for Matrix Factorization and Sparse Coding. JMLR 11:19-60, 2010. [7] U. Evci, B. van Merriënboer, T. Unterthiner, M. Vladymyrov, F. Pedregosa. GradMax: Growing Neural Networks using Gradient Information. ICLR 2022. arXiv:2201.05125. [8] B. Bussmann, N. Nabeshima, A. Karvonen, N. Nanda. Learning Multi-Level Features with Matryoshka Sparse Autoencoders. ICML 2025. arXiv:2503.17547.

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

Related papers

Discussion (0)

Sign in to comment.