Sanity Checks for Sparse Autoencoders: Do SAEs Beat Random Baselines?
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 of true features despite achieving 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.
Figures & tables
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
| Claimed Benefits | Documented Limitations |
|---|---|
| Core Purpose: Decompose dense activations into sparse, interpretable, monosemantic features [ Bricken et al., 2023 , Huben et al., 2024 ] | Spurious Features: Find features even in randomly initialized transformers with no learned structure [ Heap et al., 2025 ] |
| Monosemanticity: Address polysemanticity by learning features that correspond to single coherent concepts [ Bricken et al., 2023 , Huben et al., 2024 ] | Unfaithfulness: Reconstructed features do not faithfully represent true model computations [ Leask et al., 2025 , Menon et al., 2025 ] |
| Scalability: Successfully scaled to frontier models including Claude 3 Sonnet, GPT-4, and Gemma 2 [ Templeton et al., 2024 , Gao et al., 2025 , Lieberum et al., 2024 ] | Poor Generalization: Fail to generalize across tasks, settings, and perturbations [ Heindrich et al., 2025 , Kantamneni et al., 2025 , Li et al., 2026 ] |
| AI Safety Applications: Enable mechanistic interpretability for understanding safety mechanisms, reasoning, and misalignment [ Wang et al., 2023 , Galichin et al., 2025 ] | Downstream Performance: Underperform on downstream tasks; other approaches often outperform SAEs [ Wu et al., 2025 , Smith et al., 2025 ] |
| Reconstruction Quality: Achieve high reconstruction fidelity [ Bricken et al., 2023 , Gao et al., 2025 ] | Training Fragility: Highly sensitive to hyperparameters and initialization [ Chanin and Garriga-Alonso, 2025 , Paulo and Belrose, 2025 , Minegishi et al., 2025 ] |
| Method | Explained Variance | Sparse Probing (Top-1) | Causal Editing |
|---|---|---|---|
| Fully-Trained SAE | 0.852 | 0.721 | 0.720 |
| Soft-Frozen ( ) | 0.785 | 0.659 | 0.728 |
| Soft-Frozen ( ) | 0.734 | 0.639 | 0.712 |
| Soft-Frozen ( ) | 0.691 | 0.637 | 0.681 |
| Variant | Explained Variance | Sparse Probing (top-1) | Sparse Probing (top-5) | Causal Editing |
|---|---|---|---|---|
| BatchTopK: | ||||
| Fully-Trained (iso) | 0.852 0.001 | 0.721 0.026 | 0.849 0.025 | 0.720 0.049 |
| Frozen Decoder (iso) | 0.430 0.000 ∗∗∗ | 0.669 0.010 ∗ | 0.768 0.013 | 0.562 0.027 ∗∗ |
| Frozen Decoder (cov) | 0.570 0.004 ∗∗∗ | 0.692 0.014 | 0.797 0.014 | 0.554 0.028 ∗∗ |
| Soft-Frozen Decoder (iso) | 0.785 0.001 ∗∗∗ | 0.659 0.035 ∗ | 0.723 0.031 | 0.728 0.038 |
| Frozen Encoder (iso) | 0.475 0.004 ∗∗∗ | 0.638 0.013 ∗∗ | 0.731 0.013 | 0.641 0.027 ∗ |
| Variant | Explained Variance | Sparse Probing (top-1) | Sparse Probing (top-5) | Causal Editing |
|---|---|---|---|---|
| BatchTopK: | ||||
| Fully-Trained (iso) | 0.887 0.000 | 0.806 0.032 | 0.900 0.016 | 0.512 0.011 |
| Frozen Decoder (iso) | 0.404 0.001 ∗∗∗ | 0.709 0.019 ∗∗ | 0.829 0.020 | 0.492 0.002 ∗ |
| Frozen Decoder (cov) | 0.555 0.001 ∗∗∗ | 0.730 0.020 ∗ | 0.818 0.020 | 0.493 0.002 ∗ |
| Soft-Frozen Decoder (iso) | 0.812 0.001 ∗∗∗ | 0.758 0.037 | 0.823 0.031 | 0.509 0.010 |
| Frozen Encoder (iso) | 0.562 0.000 ∗∗∗ | 0.722 0.026 ∗ | 0.812 0.025 | 0.492 0.002 ∗ |
| Variant | Explained Variance | Sparse Probing (top-1) | Sparse Probing (top-5) | Causal Editing |
|---|---|---|---|---|
| BatchTopK: | ||||
| Fully-Trained (iso) | 0.805 0.000 | 0.766 0.039 | 0.874 0.022 | 0.533 0.015 |
| Frozen Decoder (iso) | 0.268 0.001 ∗∗∗ | 0.709 0.012 | 0.779 0.019 | 0.486 0.004 ∗ |
| Frozen Decoder (cov) | 0.426 0.001 ∗∗∗ | 0.689 0.014 ∗ | 0.771 0.016 | 0.491 0.005 ∗ |
| Soft-Frozen Decoder (iso) | 0.699 0.000 ∗∗∗ | 0.752 0.046 | 0.848 0.030 | 0.519 0.009 |
| Frozen Encoder (iso) | 0.424 0.001 ∗∗∗ | 0.713 0.010 | 0.798 0.012 | 0.492 0.005 ∗ |
| Architecture | EV | MMCS | |
|---|---|---|---|
| TopK | 20 | 0.897 | 0.644 |
| TopK | 100 | 0.947 | 0.825 |
| BatchTopK | 20 | 0.799 | 0.322 |
| BatchTopK | 100 | 0.891 | 0.489 |
| JumpReLU | 20 | 0.844 | 0.441 |
| JumpReLU | 100 | 0.889 | 0.499 |
Sparse Autoencoders
Sparse autoencoders decompose neural network activations into sparse features to study, interpret, and control learned representations.
See all 242 papers