SAE Features
Section titled “SAE Features”Source. Sparse Autoencoders Find Highly Interpretable Features (Cunningham et al., 2024).
Description. Sparse autoencoders are trained on residual-stream activations to recover an overcomplete dictionary of directions, evaluated two ways: automated interpretability scoring of 150 features per method against six baselines (default basis, random directions, PCA, ICA, top- PCA, top- ICA), and activation patching on 50 IOI data points. Evaluated on: Pythia-70M and 410M at origin, GPT-2 small and Gemma 2 2B in the follow-up. Being a method-level claim, these are the systems the evidence was gathered on rather than one audited model. Description mode: representational.
Readings of the claim
Section titled “Readings of the claim”SAE features. Readings of the claim, from the version the authors state to the version the field cites.
| Reading | Verdict | Missing |
|---|---|---|
| Primary. Sparse dictionary learning recovers directions more interpretable than PCA, ICA and the neuron basis, and causally usable | Proposed | Baseline separation (M2), where the origin’s within-model controls and the later random-network controls disagree; then specificity (I4), independent convergence (C3), separation from other decompositions (I6), and calibration of the scoring instrument (M1, M3, M5, M7) |
| SAE features are atomic, canonical units of the model | Disconfirmed | Nothing — tested and failed: meta-SAEs decompose latents further (C4), and the instrument cannot distinguish a trained transformer from a random one (Heap et al., 2025) |
| SAE features localize causation better than neurons | Disconfirmed | Nothing — tested and failed: (Mueller et al., 2025) report no causal-localization advantage over the neuron basis |
Verdict
Section titled “Verdict”SAE features. Verdict: Proposed. The claim stops at the rung marked Blocked; rungs below it are Reached and rungs above it are blocked by that rung. Criteria in bold are absent or adverse at that rung; only those the rung requires can block it.
| Tier | Requires | Missing |
|---|---|---|
| Proposed | construct (C1–C2) | Reached. |
| Causally Suggestive | measurement (M2); internal (I1) | Blocked. M2 baseline separation — Separates from random directions; not from a random network when that control is run |
| Mechanistically Supported | internal (I2, I4); external (E1) | Blocked at Causally Suggestive. I4 specificity — Single-feature ablation moves 12,000 logits; left unanalyzed |
| Triangulated | construct (C3–C4); internal (I5–I7); external (E2, E4) | Blocked at Causally Suggestive. I6 double dissociation — Unattempted; no second decomposition shown intact |
| Validated | construct (C5–C6); measurement (M1–M6); internal (I3, I10–I12); external (E2–E6); interpretive (V1–V5) | Blocked at Causally Suggestive. C6 complementation validity — Feature splitting is the open question; no test of atomicity is run; E3 cross-task generalization — Unattempted; generalization stated as an expectation; I10 rescue reversibility — Unattempted, though patching is reversible by construction; I11 onset coupling — Unattempted, and cheap: Pythia ships pretraining checkpoints (Biderman et al., 2023); I12 offset coupling — Unattempted; the nearest result is at an extreme, not a trajectory (Heap et al., 2025) |
36-criterion audit
Section titled “36-criterion audit”SAE features. Full 36-criterion audit. Status: C = Confirmed, PC = Partially confirmed, U = Untested, I = Inconclusive, D = Disconfirmed, N/A = Not applicable.
| ID | Criterion | Status | Evidence |
|---|---|---|---|
| Construct Validity | |||
| C1 | Falsifiability | C | Directional prediction against four named baselines; partly fails in layer 4 |
| C2 | Structural plausibility | PC | Predicted by superposition; MLP training fails, attention not attempted |
| C3 | Convergent validity | PC | Two evidence types; the concurrent-origin pairing comes later (Leask et al., 2025) |
| C4 | Discriminant validity | PC | Scored against the origin’s operational construct, not the field’s |
| C5 | Nomological validity | PC | Fits superposition; the theory does not underwrite the method |
| C6 | Complementation validity | U | Feature splitting is the open question; no test of atomicity is run |
| Measurement Validity | |||
| M1 | Reliability | I | One run at origin; later seed studies disagree (Paulo & Belrose, 2025) |
| M2 | Baseline separation | I | Separates from random directions; not from a random network when that control is run |
| M3 | Stability | I | Sparsity sweep gives a smooth tradeoff, not a stable optimum |
| M4 | Calibration | PC | Interpretability score is a correlation with a meaningful zero |
| M5 | Sensitivity | D | Tested later on planted features and the recovery is poor (Korznikov et al., 2026) |
| M6 | Invariance | PC | All layers of Pythia-70M, five expansion ratios, two scoring regimes |
| M7 | Selection correction | U | A feature scoring fewer than 20 varying fragments is skipped; the count is unreported |
| Internal Validity | |||
| I1 | Necessity | PC | One feature ablated for behavior; a full previous-layer sweep reads out to features |
| I2 | Sufficiency | PC | Interchange intervention on IOI; counterfactual features written in |
| I3 | Minimality | PC | ACDC orders features; the next feature usually adds little |
| I4 | Specificity | I | Single-feature ablation moves 12,000 logits; left unanalyzed |
| I5 | Rival mechanism exclusion | PC | Rival methods excluded: PCA, ICA, top-, non-sparse dictionary |
| I6 | Double dissociation | U | Unattempted; no second decomposition shown intact |
| I7 | Confound control | PC | Three designed controls, including an dictionary |
| I8 | Confounding sensitivity | U | Unattempted; no E-value or sensitivity analysis |
| I9 | Epistatic interaction | PC | Assumed away explicitly; ACDC treats features as a flat graph |
| I10 | Rescue reversibility | U | Unattempted, though patching is reversible by construction |
| I11 | Onset coupling | U | Unattempted, and cheap: Pythia ships pretraining checkpoints (Biderman et al., 2023) |
| I12 | Offset coupling | U | Unattempted; the nearest result is at an extreme, not a trajectory (Heap et al., 2025) |
| External Validity | |||
| E1 | Intervention reach | PC | Two forms: activation patching and less-than-rank-one ablation |
| E2 | Prompt generalization | PC | Cross-corpus by construction: trained on the Pile, scored on OpenWebText |
| E3 | Cross-task generalization | U | Unattempted; generalization stated as an expectation |
| E4 | Cross-model recurrence | C | Transfers everywhere tried: Pythia, GPT-2 small, Gemma 2 2B (Leask et al., 2025) |
| E5 | Graded response | PC | Two graded curves: KL against features patched, edit size against KL |
| E6 | Novel prediction | PC | The IOI experiment is a prediction in the weak sense |
| Interpretive Validity | |||
| V1 | Level declaration | PC | Formal object declared precisely; “feature” carries more than it |
| V2 | Level-evidence match | PC | Evidence supports the decomposition claim the origin makes |
| V3 | Alternative level | PC | Raised at origin and not pursued: no single correct decomposition |
| V4 | Unlicensed labeling | PC | “Monosemantic” projects semantics; the origin hedges consistently |
| V5 | Scope declaration | C | Both papers declare limits and quantify them (Leask et al., 2025) |
| Total: 3 Confirmed, 20 Partially confirmed, 4 Inconclusive, 8 Untested, 1 Disconfirmed | |||
| Verdict: Proposed |
Exploratory Lens Analysis
Section titled “Exploratory Lens Analysis”An exploratory reading of this claim through the framework’s five lenses, written for this site and not part of the paper, is at SAE Features — exploratory lens analysis.