Machine Learning · arXiv:2602.14111 v2 · submitted 15 Feb 2026
Sanity Checks for Sparse Autoencoders: Do SAEs Beat Random Baselines?
The paper says
“Together, these results suggest that SAEs in their current state do not reliably decompose models' internal mechanisms.”
from the abstract · picked out as the main claim by a classifier
Reviews
Reviews by people
No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Sparse Autoencoders (SAEs) have emerged as a promising tool for interpreting neural networks by decomposing their activations into sparse sets of human-interpretable features. Recent work has introduced multiple SAE variants and successfully scaled them to frontier models. Despite much excitement, a growing number of negative results in downstream tasks casts doubt on whether SAEs recover meaningful features. To directly investigate this, we perform two complementary evaluations. On a synthetic setup with known ground-truth features, we demonstrate that SAEs recover only 9% of true features despite achieving 71% explained variance, showing that they fail at their core task even when reconstruction is strong. To evaluate SAEs on real activations, we introduce three baselines that constrain SAE feature directions or their activation patterns to random values. Through extensive experiments across multiple SAE architectures, we show that our baselines match fully-trained SAEs in interpretability (0.87 vs 0.90), sparse probing (0.69 vs 0.72), and causal editing (0.73 vs 0.72). Together, these results suggest that SAEs in their current state do not reliably decompose models' internal mechanisms.
The paper's own sentences, arranged. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-11-04 · no language-model review
How the main claim is stated: hedged possibility 0.99
What the abstract itself states, as probabilities: claims first 0.04 · states limitation 0.27 · human participants 0.04 · controlled comparison 0.98 · randomized assignment 0.03 · states quantitative result 0.99 · compares with prior results 0.96 · mentions code or data availability 0.03
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v2). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Sparse Autoencoders (SAEs) have emerged as a promising tool for interpreting neural networks by decomposing their activations into sparse sets of human-interpretable features. | background 1.00 | 0.04 | 0.02 | asserted 0.85 |
| Recent work has introduced multiple SAE variants and successfully scaled them to frontier models. | background 1.00 | 0.05 | 0.01 | asserted 1.00 |
| Despite much excitement, a growing number of negative results in downstream tasks casts doubt on whether SAEs recover meaningful features. | background 0.58 | 0.05 | 0.03 | hedged possibility 0.59 |
| To directly investigate this, we perform two complementary evaluations. | contribution 0.61 | 0.59 | 0.02 | asserted 0.53 |
| On a synthetic setup with known ground-truth features, we demonstrate that SAEs recover only $9\%$ of true features despite achieving $71\%$ explained variance, showing that they fail at their core task even when reconstruction is strong. | result 1.00 | 0.96 | 0.20 | asserted 1.00 |
| To evaluate SAEs on real activations, we introduce three baselines that constrain SAE feature directions or their activation patterns to random values. | method 0.86 | 0.88 | 0.03 | asserted 0.96 |
| Through extensive experiments across multiple SAE architectures, we show that our baselines match fully-trained SAEs in interpretability (0.87 vs 0.90), sparse probing (0.69 vs 0.72), and causal editing (0.73 vs 0.72). | result 1.00 | 0.96 | 0.14 | asserted 1.00 |
| Together, these results suggest that SAEs in their current state do not reliably decompose models' internal mechanisms. | interpretation 1.00 | 0.71 | 0.95 | hedged possibility 0.99 |
What kind of paper the classifier reports this is: empirical study 1.00 · evidence from observation or experiment 0.92
What was examined, and what was not
Sources read
- arXiv listing abstract v2
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-04: 1 request, 6,287 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Cited by 3 Pith papers
Reviewed papers on Pith that cite this one. The full citation record
- Decomposing Earth Embeddings with Sparse Autoencoders cs.LG · Nov 2026
- User Embeddings are Superpositions of Interpretable Behavioral Modes cs.LG · Nov 2026
- Reducing Logprob Without Causal Verification: Diagnosing SAE-Based PII Suppression in LLMs cs.AI · Nov 2026
How the literature uses it
9 arXiv papers cite this in 12 sentences
Papers citing it · 8 sentences without a confident label
Cites it as support 4 papers
-
For post-hoc SAEs, this is not guaranteed: SAEs trained on randomly initialized transformers produce auto-interpretability scores approaching those obtained from trained models [ref] , and random baselines that replace SAE feature directions or activation patterns with noise match fully-trained SAEs on auto-interpretability, sparse probing, and causal editing [ref] .
Scaling Interpretable Transformers with Parity Bottleneck Layers · Fine-grained causal intervention · 0.96
-
Proxy metrics such as reconstruction error and sparsity reflect optimization behavior rather than semantic quality [ref] , and SAEs that perform well on these metrics have been shown to detect only a small subset of true features [ref] .
Building Fast, Evaluating Slow: Pipeline Choices Dominate Autointerpretability Score Variance · Introduction · 0.98
-
[ref] show that seed alone substantially changes which features are learned, TopK more than ReLU ([ref]); [ref] find automated interpretability scores do not separate trained transformers from random ones; [ref] show frozen-random baselines match fully-trained SAEs across reconstruction, interpretability, sparse probing, and causal editing ($0.73$ vs $0.72$); and [ref] finds SAEBench's TPP and SCR fail three independent reliability lenses.
Where You Measure Decides What You Measure: Position Selection in Ablation-Based SAE Evaluation · Related work · 0.96
Challenges and discussion
Know something about this paper? Leave a comment.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment