Source-linked AI summary
A Dominant Diffuse Phase in the Sparse Autoencoder Phase Diagram
Alexis D. Plascencia
TL;DR
Sparse autoencoders are intended to recover interpretable features, but co-occurrence can cause absorption or merging. This paper measures the MAIS-O43 protocol and finds a reproducible diffuse phase: reconstruction is nearly perfect while feature recovery and sparsity fail, including across the full grid under minibatch training.
Problem
The paper asks when sparse autoencoders recover a synthetic dictionary versus merge features as nesting, sparsity penalty, and dictionary size vary.
Method
The study implements the MAIS-O43 protocol with 200 canonical fits across ten cells and 3,300 additional minibatch fits covering all 165 cells.
Results
Zero full-dictionary recoveries and zero merges were observed; instead, every run entered a diffuse phase with near-perfect reconstruction and codes 10–28× denser than the generating process.
Takeaways & Limitations
The phase diagram of trained SAEs can differ from that of objective minimizers, so near-perfect reconstruction does not establish recovery of true features.
Takeaways & Limitations
The study leaves open whether the diffuse phase persists across additional dictionary and data draws, SAE variants, and global minimizers.
Abstract
from arXiv · showhide
Sparse autoencoders (SAEs) are increasingly used to recover interpretable features from neural-network activations, yet systematic feature co-occurrence can cause distinct features to be absorbed or merged. The MAIS-O43 open problem proposes a controlled experiment to characterize when recovery of a true synthetic dictionary gives way to feature merging as the nesting fraction $γ$, sparsity penalty $λ$, and dictionary size $M$ vary. We implement the specified protocol and evaluate 200 independently initialized fits across ten of the 165 grid cells. We observe zero full-dictionary recoveries and zero merges. Instead, every run converges to a reproducible diffuse phase: reconstruction is nearly perfect, but learned atoms typically remain far from the true features (median best cosine 0.5-0.7 against a 0.95 recovery criterion) and learned codes are an order of magnitude denser than the ground truth. This behavior persists under robustness checks and across the full 165-cell grid using standard minibatch Adam (3,300 additional fits). Since the global optimum of the exact sparse-coding objective is known to merge nested features in the two-feature case, these results suggest that trained SAEs need not reach the corresponding minima, and that the phase diagram of trained models may differ fundamentally from that of objective minimizers.
1 Introduction
This paper measures the MAIS-O43 phase diagram for sparse autoencoders and finds a diffuse outcome rather than the anticipated recovery-versus-merging behavior. The study reports robustness across training budgets and optimizer settings.
- Study scope: The experiment measures recovery, merging, and splitting probabilities for a known synthetic dictionary across an 11 × 5 × 3 parameter grid.The canonical measurements cover 10 of 165 cells, while a minibatch companion grid covers all 165 cells.
- Study scope: 200 fits across ten grid cells produced zero full-dictionary recoveries and zero merges under the specified protocol.The protocol varies nesting fraction γ, sparsity penalty λ, and dictionary size M, with 20 independent initializations per cell.
- Main finding: The observed outcome is a diffuse phase with near-perfect reconstruction, low matched-feature fractions, and learned codes substantially denser than the ground truth.This phase differs from the recovery-versus-merging outcomes anticipated by the open problem.
- Robustness: Doubling the update budget leaves the diffuse phase unchanged, while minibatch Adam reaches the same plateau 64× faster and shows no recovery or merge across all 165 cells.These checks test both under-training and dependence on the protocol’s full-batch optimizer.
2 Related work
Prior work frames sparse autoencoders as dictionary-learning systems that can recover, absorb, or split structured features. This paper instead measures a low-alignment diffuse regime and distinguishes trained-model behavior from theoretical minimizers.
- Hierarchical features: Hierarchical generators model child-implies-parent support structure, allowing an unstructured learner’s recovery of nested features to be tested directly.This setup differs from structured regularization, which builds hierarchy into the learner.
- Sparse autoencoders and dictionary learning: Sparse autoencoders extend dictionary learning to neural activations, where decoder atoms are interpreted as candidate features under superposition.The related literature connects sparse coding, compressed representations, and mechanistic interpretability.
- Hierarchical features: Feature absorption occurs when co-occurring features are represented through joint or transformed directions rather than recovered separately.Prior examples include representing “animal-except-dog” plus “dog” instead of the original pair.
- Position relative to prior work: Under MAIS-O43, no full-dictionary recovery or merge occurs, while at most 3.6% of true features per run have a distinct atom above cosine 0.95.The paper attributes the contrast with favorable toy-model results to the regime, vanilla architecture, and strict quantitative definitions as candidate explanations.
- Theory and trained models: The results test whether trained amortized SAEs reach theoretically characterized minima, which merge nested features in the two-feature case.Recoverability in principle therefore does not establish that a trained SAE attains the corresponding dictionary.
3 The MAIS-O43 protocol
The MAIS-O43 protocol generates sparse nonnegative data from a normalized synthetic dictionary with controllable parent–child nesting, then evaluates a centered ReLU sparse autoencoder using explicit recovery, merge, and split criteria.
- Generator: The generator uses n = 64 ambient dimensions and m = 256 normalized Gaussian features paired into 128 candidate parent–child relationships.A fraction γ of the pairs is nested, and nested child activity requires parent activity.
- Generator: The γ intervention changes both parent–child dependence and child marginal frequency, confounding nesting effects with rarity.Expected activation density decreases from 0.03125 at γ = 0 to 0.01611 at γ = 1.
- Model and training: The centered SAE uses a ReLU encoder and unit-norm decoder columns without output bias, trained by full-batch Adam for 2 × 10^5 updates.Twenty independently initialized models are trained per cell.
- Outcome definitions: Recovery requires an injective matching that assigns every true feature a decoder atom with cosine at least 0.95.Merge and split are evaluated separately using thresholded geometric criteria.
4 Methods
The implementation uses validated, reproducible code and vectorized training to run the canonical fits efficiently. Tests verify the generator, gradients, classifiers, and equivalence between batched and scalar training.
- Implementation and validation: A 49-test suite validates data-generation invariants, gradient correctness, exact chunking equivalence, and adversarial recovery, merge, and split classifiers.Validation includes threshold-boundary cases and the analytic two-feature nested optimum.
- Vectorized training: The 20 fits in each cell are trained as one batched tensor program whose trajectories match independently trained scalar runs to 10^-9 tolerance in float64 tests.Formal runs use strict float32 with TF32 disabled.
- Vectorized training: A 20-seed cell requires 10.2 hours for M = 256 or 17.5 hours for M = 512 on the specified GPUs.The batched implementation preserves per-model Adam behavior while reducing execution overhead.
5 Results
Across ten measured cells, none of 200 fits recovered the full dictionary or merged features; instead, models entered a diffuse phase with near-perfect reconstruction but weak feature alignment and denser codes.
- 0 of 200 fits recovered all 256 features, and 0 of 200 contained a live atom satisfying the pairwise merge criterion.Individual matches occurred in 44 runs, totaling 215 matched assignments; splitting appeared in 5 of 20 M = 512 runs.
- 2.7 × 10−6 to 5.6 × 10−3 reconstruction loss accompanied mean median best cosines of 0.472–0.727 and learned codes 9.9–27.9× denser than ground truth.The paper labels this combination a diffuse phase: near-floor reconstruction error, matched-feature fractions below 0.04, and substantially denser codes.
- λ response: At γ = 0.5 and M = 256, alignment peaked at λ = 3 × 10−3, reaching median best cosine 0.727 and 23 of 256 features above cosine 0.90.Alignment declined at λ = 10−2, where density reached 0.48 and activation magnitudes shrank; the best penalty still missed the recovery criterion.
- Nesting response: Along λ = 10−2, γ ≤ 0.5 cells were statistically indistinguishable, while merging never occurred, including at γ = 1.The nesting response showed only a mild alignment decline at γ ≥ 0.8.
- Dictionary size: At (γ, λ) = (0.5, 10−3), increasing M from 256 to 512 raised matched-feature fraction from 0.001 to 0.036 and median cosine from 0.572 to 0.712.This was larger than any measured λ or γ effect, and M = 512 was the only setting producing splits.
- Representative cell: Figure 1 compares training-objective trajectories with per-feature best cosine against the 0.95 recovery criterion for the γ = 0.5, λ = 10−3, M = 512 cell.The objective is averaged over 20 seeds, while converged feature similarities are pooled across those seeds.
- λ response: Figure 2 encodes the λ-response at γ = 0.5, M = 256 using median best cosine and the count of features reaching cosine 0.90.It identifies an interior peak at λ = 3 × 10−3 and collapse at λ = 10−2.
6 Robustness probes
Robustness checks show that the diffuse phase is not explained by insufficient training or full-batch optimization, and it persists across the complete minibatch grid.
- Probe A: 0/20 recoveries and 0/20 merges remain after doubling training to 400k updates, while matched-feature fraction stays near 0.004.Median best cosine changes only from 0.727 to 0.729, with about 23 features above cosine 0.90.
- Probe B: 0/20 recoveries and 0/20 merges persist under minibatch Adam, which reaches the same diffuse plateau about 64× faster.The minibatch run reaches median best cosine 0.739 by about 60k steps and has matched-feature fraction 0.001.
- Full companion grid: Across all 3,300 minibatch fits, recovery and merge fractions are zero, while matched-feature fraction never exceeds 0.046 against a criterion of 1.0.Alignment concentrates in the M=512, moderate-λ region and is essentially independent of γ.
- Scope: The complete companion grid is off-protocol and therefore does not resolve MAIS-O43, although it makes the canonical grid’s likely outcome concrete.The canonical protocol measured only selected cells under full-batch training.
7 Discussion
The experiments identify a diffuse phase distinct from both recovery and merging, and suggest that trained amortized SAEs need not realize the global minima predicted for exact sparse coding.
- A third phase: A converged diffuse phase dominates the sampled region: reconstruction reaches near-floor error while codes remain dense and learned atoms stay weakly aligned.The loss curves show reconstruction collapsing within a few thousand updates, followed by slow and incomplete rotation toward the true dictionary.
- Trained SAEs versus minimizers: Theory predicts merged global minimizers for a two-feature exact-coding objective, whereas the experiment trains an amortized ReLU objective on a finite 256-feature dataset.The difference may reflect amortization, many-feature structure, finite samples, or the optimization path.
- Practical reading: Low reconstruction error does not certify dictionary recovery when only a small minority of true features receive distinct atoms above a stringent alignment threshold.This warning concerns a regime in which feature alignment is largely absent, rather than conventional absorption alone.
- Practical reading: Whether hierarchy-aware architectures improve the low matched fractions under the MAIS-O43 generator remains an open, inexpensive question.The measured lack of alignment leaves such methods little structure to organize.
8 Conclusions
The study finds a reproducible diffuse phase rather than recovery or merging: reconstruction remains nearly perfect while learned features stay misaligned and codes are much denser. These results distinguish trained SAE behavior from theoretical objective minimizers and caution against using reconstruction quality as evidence of feature recovery.
- Across 200 canonical fits, there are zero full-dictionary recoveries and zero merges; instead, learned atoms remain misaligned while codes are 10–28× denser than generated codes.The diffuse solutions reconstruct the data essentially perfectly, despite failing both anticipated outcomes.
- The diffuse phase is reproducible across all initializations, survives doubled training, and reappears across all 165 cells under minibatch Adam.The minibatch companion comprises 3,300 additional fits and reaches the same plateau 64× faster.
- Alignment peaks at λ = 3 × 10−3, declines at λ = 10−2, improves from 0.001 to 0.036 at M=512, and changes little with γ, although γ also changes nested-child frequency.The γ intervention therefore confounds dependence with rarity, limiting interpretation of its alignment effect.
- The results separate trained SAE behavior from theoretical minimizers, which are predicted to merge nested features in the two-feature case.The paper argues that phase diagrams of trained models and objective minimizers should be treated as different objects.
- Near-perfect reconstruction is compatible with zero feature recovery, so reconstruction quality alone is not evidence that an SAE found true features.
- The released code, seeds, hashes, checkpoints, and regeneration scripts support reproducing the measured results, while broader draws, variants, and global-minimizer comparisons remain open.The complete minibatch grid is inexpensive to rerun, but completing the remaining canonical full-batch cells is computationally costly.
9 Reproducibility
The paper releases the artifacts needed to regenerate its figures, tables, and measured results. Deterministic data generation and published hashes provide checks on reproducibility.
- The repository provides code, preregistration materials, per-cell results with SHA-256 hashes, checkpoints, and scripts for regenerating figures and tables.The released artifacts cover the canonical fits and associated analyses.
- All randomness derives from master seed 20260805, and regenerated data are deterministic and verified against published hashes.
- The README documents the environment and commands used to regenerate each published figure and table.