A Dominant Diffuse Phase in the Sparse Autoencoder Phase Diagram
cs.LG
Quick verdict
在 MAIS-O43 网格上进行 3,500 次 SAE 拟合训练：没有恢复，没有合并，只有持续的弥散阶段；严谨的预注册工作，但存在范围限制。
Summary
这篇实证研究在 64 个维度上使用包含 256 个合成 ground-truth 特征的字典训练 ReLU+ℓ1 稀疏自编码器 (SAE)，其中可调比例的 γ 特征对是嵌套的父子对。在 MAIS-O43 协议下（全批 Adam，10 个网格单元中的 200 次拟合），从未观察到完整字典恢复和合并；每次运行都收敛到可重现的“弥散阶段”，具有近乎完美的重建，但相对于 0.95 的恢复标准，最佳余弦相似度中位数仅为 0.47–0.73，且代码密度比 ground truth 高 10–28 倍。该发现在训练预算加倍后得以保留，并在用于实践的标准 minibatch Adam 下重现，覆盖了所有 165 个网格单元（额外 3,300 次拟合）。所陈述的局限性包括：仅使用了一个字典抽取，规范协议仅覆盖 165 个单元中的 10 个，以及缺乏推断统计；γ 干预还将嵌套与子特征稀有度混淆了。
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Across 200 canonical MAIS-O43 full-batch fits (10 grid cells) plus 3,300 minibatch companion fits (all 165 cells), zero runs achieve full dictionary recovery and zero atoms satisfy the pairwise merge criterion.
- Every run converges instead to a reproducible diffuse phase: near-perfect reconstruction, matched-feature fractions ≤0.036 per cell, median best cosine 0.47–0.73 against a 0.95 criterion, and codes 9.9–27.9× denser than ground truth.
- The diffuse phase is a genuine training attractor, unchanged by doubling the update budget (Probe A) and reproduced under standard minibatch Adam ~64× faster (Probe B).
- The phase diagram of trained SAEs is a different object from the phase diagram of the exact-coding objective's minimizers, so theory calibrated on minimizers need not transfer to trained models.
- For interpretability practice, near-perfect reconstruction does not certify that an SAE has recovered true features.
Strengths
Protocol fidelity and verification infrastructure
Section 4 describes a 49-test suite: gradient agreement with an independent NumPy implementation at 1e-12, chunked-vs-single-batch gradient equality, nine adversarial classifier fixtures (greedy-defeating matching, threshold-boundary cases at 0.95/0.90/0.05), deterministic regeneration from master seed 20260805, and SHA-256 data hashes.
The headline null is stress-tested twice
Probe A doubles the 200k-update budget (median best cosine 0.727→0.729); Probe B swaps in standard minibatch Adam reaching cosine 0.739 ~64× faster; a 3,300-fit companion grid then covers all 165 cells with zero recoveries and zero merges.
Candor about confounds and scope
Genuinely novel empirical regime
Complete artifact release
Weaknesses
Single dictionary and dataset draw
Canonical protocol covers 10 of 165 cells
Descriptive statistics only
Companion-grid zero-recovery claim is partly structural
Editorial seams suggest light polish
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This paper reports the first measurements of the MAIS-O43 phase diagram (Section 3): a fully specified synthetic experiment with n=64 ambient dimensions, m=256 unit-norm Gaussian dictionary features, and ⌊128γ⌋ of 128 candidate pairs nested as parent–child (child activity A_j = B_i·B_j). The centered ReLU+ℓ1 SAE is trained on G_λ (Eq. 1) with full-batch Adam for 2×10^5 updates, 20 seeds per cell. The central empirical claims: zero full recoveries and zero merges in 200 canonical fits over 10 of 165 cells (Table 1); a converged "diffuse phase" instead — near-floor reconstruction, matched-feature fractions ≤0.036 per cell, median best cosine 0.472–0.727 against a 0.95 criterion, codes 9.9–27.9× denser than ground truth; and an interior alignment maximum at λ=3×10⁻³ along the λ-sweep. A further claim is conceptual: trained SAEs differ from the exact-coding minimizers that provably merge nested features, so minimizer-based theory need not transfer.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the MAIS-O43 open problem?
It is a fully specified experiment from the MAIS research agenda on superposition: measure how often a sparse autoencoder recovers, merges, or splits a known synthetic dictionary across an 11×5×3 grid of nesting fraction γ, sparsity penalty λ, and dictionary size M, with 20 random initializations per cell (3,300 fits total), publishing seeds and checkpoints.
What is the 'diffuse phase' this paper reports?
A converged training regime where the SAE reconstructs the data almost perfectly, yet learned atoms stay poorly aligned with the true features (median best cosine similarity 0.47–0.73 versus a 0.95 recovery criterion), only a few percent of features are matched per run, and learned codes are roughly 10–28 times denser than the ground-truth codes.
Did any sparse autoencoder run recover the true features or merge them?
No. Across 200 canonical full-batch fits and 3,300 additional minibatch fits, zero runs achieved full dictionary recovery and zero atoms satisfied the merge criterion; the only systematic structure was occasional feature splitting at dictionary size M=512.
Why does this matter for interpretability?
Because practitioners often read low reconstruction error as evidence that an SAE has found the true underlying features. This paper shows a regime where reconstruction is nearly perfect while almost no learned atom aligns with any true feature, so reconstruction quality alone must not be treated as evidence of feature recovery.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Only the paper text was analyzed; the released repository, checkpoints, and preregistration were not independently inspected, so artifact claims rest on the paper's own description.
- The provided text appears complete (through references and acknowledgments) and is internally consistent between the abstract and body; figures were not visually inspected, so reported s.d. bands could not be verified.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.