Sparse Autoencoders
83 beliefs (83 IN, 0 OUT)
Sparse Autoencoders (SAEs) are a central tool in mechanistic interpretability, decomposing a model's dense d-dimensional residual stream activations into a typically 10x to 200x larger set of mostly-sparse, near-orthogonal latent features (sae-expansion-ratio-10-to-200x, sae-expansion-ratio-definition). The standard architecture uses a ReLU nonlinear encoder paired with a purely linear decoder (sae-architecture-relu-encoder-linear-decoder), trained to minimize reconstruction error plus an L1 sparsity penalty weighted by decoder weight norm (sae-training-objective-mse-plus-l1, sae-loss-function-decoder-norm-penalty). Sparsity is enforced either through this soft L1 penalty or, in Anthropic's scaling monosemanticity work, through a harder top-k constraint that retains only the top-K activations per token (sae-topk-architecture-used, eh2022-sae-top-k-sparsity-requirement). The SAE is applied specifically to middle-layer residual stream activations to reduce dimensionality versus MLP outputs and mitigate cross-layer superposition (sae-middle-layer-residual-stream-target). A key algorithmic improvement, proposed by Bricken, multiplies the sparsity penalty by the decoder norm and removes the unit-norm constraint on decoder vectors (sae-algorithmic-improvement-bricken). Losses across sizes (1M, 4M, 34M features) follow an approximate power law in compute, and all three achieve at least 65% variance reconstruction with fewer than 300 active features per token (sae-scaling-law-power-law-compute, sae-reconstruction-variance-and-sparsity). The Bricken et al. paper explicitly rejects the "sparse architecture" training approach as insufficient to eliminate polysemanticity (saey-paper-rejects-sparse-architecture-approach), and known pathologies include the shrinkage problem (sae-shrinkage-problem) and feature splitting, where a single coarse feature decomposes into finer sub-features as width increases (sae-jcr-prevents-feature-splitting, saey-feature-splitting-base64-three-subfeatures, feature-splitting-sae-width).
The interpretability payoff is substantial: autointerpretability analysis showed SAE-identified features are substantially more interpretable than individual transformer neurons (sae-features-more-interpretable-than-neurons), and feature granularity scales with SAE size, with larger runs yielding specific entities like "Golden Gate Bridge" while smaller runs yield broad categories like "Monuments and tourist attractions" (sae-feature-granularity-scales-with-run-size). Feature neighborhoods in decoder space, defined by cosine similarity, map onto semantic relatedness (sae-cosine-similarity-defines-neighborhoods) and exhibit a general-to-specific gradient: the Golden Gate Bridge feature is flanked by San Francisco locations, California landmarks, and further-out tourist attractions (sae-golden-gate-bridge-neighborhood-structure), while a transit-infrastructure feature activates on trains, ferries, and even wormholes, encoding a higher-level transport abstraction (sae-transit-feature-abstract-generalization, feature-hierarchy-empirical-validation-v2). This neighborhood structure has been interpreted as an operational navigation of the categorical polytope geometry identified by Park et al., with cross-model feature similarity suggesting the geometric structure reflects shared data properties rather than a model-specific artifact (sae-neighborhood-as-polytope-navigation-v2, feature-neighborhood-as-geometric-theorem-instantiation-v2). More broadly, cross-model geometric convergence in SAE features and Park orthogonality is strongly predicted by the superposition framework's over-complete representation requirement (m > d) in a shared substrate, with empirical validation on Gemma-2B and LLaMA-3-8B arguing against a pure contingent coincidence, though the evidence establishes a well-motivated expectation rather than a proven logical necessity (convergence-is-necessary-not-contingent-v2). Evaluation results reinforce a distinction between output-space task-specificity and internal representational convergence: no single embedding model dominates all MTEB task categories, yet SAE-extracted features tend to be more similar across architectures than to their own model's neurons (task-specificity-vs-feature-universality-v2). SAEs trained purely on text also generalize zero-shot to image activations, indicating shared cross-modal representations (text-sae-generalizes-to-image-activations, sae-cross-modal-text-trained-image-activation), and days of the week and months appear as circular structures in SAE-reconstructed activations across GPT-2, Mistral 7B, and Llama 3 8B (three-models-circular-pca-structure). The Engels et al. ICLR 2025 paper extends this line by adding a clustering step over dictionary elements to recover multi-dimensional irreducible subspaces, with toy experiments showing that a 2-feature SAE must activate both features on every point of a circular structure while a 10-feature SAE spreads around the circle with no canonical direction, and a lower bound on the number of pairwise orthogonal subspaces (engels-2025-sae-clustering-extension, sae-circular-feature-ambiguity, engels-2024-theorem1-lower-bound, engels-2024-sae-m2-circle-both-features-fire, engels-2024-sae-m10-sparse-subset).
Causal analysis of SAE features relies on ablation by clamping a feature's activation to zero at a specific token position (sae-ablation-clamping-zero-method) and logit-difference attribution computed as the dot product of the gradient with the feature vector times activation (sae-attribution-linear-formula), with the two methods showing 0.8 correlation in worked examples (sae-attribution-ablation-0-8-correlation). Crucially, ranking features by raw activation magnitude does not reliably identify causally important features (sae-raw-activation-not-causal-importance), and steering experiments demonstrate that safety-relevant SAE features causally influence generated output rather than merely correlating with unsafe content (safety-feature-steering-causal-not-correlational). SAE-based features outperformed few-shot probe steering vectors in 5 of 7 test cases (sae-vs-fewshot-steering-comparison). Anthropic's feature search employs four methods—single-prompt activation, multi-prompt filtering, cosine-similarity nearest-neighbor search, and logit-difference attribution (sae-feature-search-four-methods)—and uses Claude 3 Opus as an interpretability judge distinct from the analyzed model to reduce self-reference bias (sae-claude-opus-as-interpretability-judge). Specific training configurations vary across models: Mistral 7B SAEs use 65,536 dictionary elements at 16x expansion, an L_{1/2} half-norm penalty, AdamW with dead-feature resampling leaving roughly 1000 dead features out of 65,536, and are trained on layers 8, 16, and 24 (mistral-sae-65536-dict-elements, mistral-sae-l1-2-sparse-penalty, mistral-sae-optimizer-dead-features, mistral-sae-layers-8-16-24); the Bricken et al. A/1 run produces 4,096 features of which 168 are dead and 292 ultralow-density (saey-a1-autoencoder-4096-features-dead-ultralow); and GPT-2-small layer 7 has approximately 25,000 features (gpt2-small-sae-25k-features-layer7).
All beliefs in this topic are currently IN (active); none have been retracted. The knowledge base distinguishes between foundational premises—architectural specifications, training hyperparameters, and observed empirical facts such as dead-feature proportions and reconstruction rates—and higher-level derived claims about geometric structure, cross-model convergence, and the interpretability superiority of SAE features over neurons. The derived geometric and convergence claims (convergence-is-necessary-not-contingent-v2, sae-neighborhood-as-directsum-navigation-v2, feature-neighborhood-as-geometric-theorem-instantiation-v2) are explicitly framed as well-motivated expectations or parallel evidence rather than proven theorems, and the neighborhood-as-polytope-navigation claim notes that the antecedents confirm categorical-polytope navigation and imply structural decomposition separately without establishing that the same traversal navigates both subspaces simultaneously (sae-neighborhood-as-directsum-navigation-v2). The Engels et al. results, validated across GPT-2, Mistral 7B, and Llama 3 8B with references to GPT-4 and Claude 3 indicating intended generality (engels-2025-models-validated), add a formal lower-bound theorem and a clustering-based extension that moves the discussion from "SAE features look interpretable" toward "SAE features recover irreducible multi-dimensional subspaces whose combinatorial geometry is bounded from below" (engels-2024-theorem1-lower-bound, engels-2025-sae-clustering-extension). Together, these beliefs paint a picture in which SAEs serve as a bridge between low-level linear features and high-level semantic geometry, with their value resting on the conjunction of sparsity-driven interpretability, causal (not merely correlational) feature effects, and a shared geometric structure that transcends individual architectures.
-
IN
anthropic-sae-expansion-factor-range
SAEs in the Bricken et al. paper are trained with expansion factors from 1x (512 features) to 256x (131,072 features), with the A/1 run focusing on 8x (4,096 features) -
IN
convergence-is-necessary-not-contingent-v2
Cross-model geometric convergence (SAE feature similarity, Park orthogonality) is strongly predicted by the superposition framework: over-complete representation (m > d) in a shared substrate motivates a common covariance structure, providing a unified explanation for why geometric invariants appear across architectures. Empirical convergence (validated on Gemma-2B, LLaMA-3-8B, with SAE features 'mostly similar' across models) is consistent with this theoretical account and argues against a pure contingent coincidence, though the evidence establishes a well-motivated expectation rather than a proven logical necessity for all systems that encode concepts in an over-complete basis. -
IN
eh2022-sae-top-k-sparsity-requirement
A Sparse Autoencoder (SAE) uses top-k sparsity constraint; without it, the autoencoder becomes a dense linear projection and loses interpretability -
IN
engels-2024-sae-m10-sparse-subset
When a 10-feature SAE is trained on a 2D unit circle, dictionary elements spread around the circle and only a sparse subset activates per input (Engels et al., Appendix D). -
IN
engels-2024-sae-m2-circle-both-features-fire
When a 2-feature SAE is trained to reconstruct a 2D unit circle, both dictionary features must activate simultaneously on every input, violating sparsity (Engels et al., Appendix D). -
IN
engels-2024-sae-train-hyperparams
The SAE toy experiments in Engels et al. (Appendix D) use Adam optimizer, learning rate 10⁻³, sparsity penalty λ = 0.1, 20,000 training steps, and 1000-step warmup. -
IN
engels-2024-separability-index-1000-rotations
The Separability Index S(f) for a 2D feature is computed by iterating over a grid of 1000 rotation angles θ ∈ [0, 2π], binning into a 40×40 grid, and taking the minimum mutual information I(a;b) over all angles. -
IN
engels-2024-theorem1-lower-bound
In a d-dimensional space, the number of pairwise δ-orthogonal d′-dimensional subspaces is bounded below by Ω(e^{C(d/d′²)δ²}/d) (Engels et al., ICLR 2025, Appendix A). -
IN
engels-2025-funding-nsf-iaifi
The Engels et al. (ICLR 2025) research is supported by NSF GRFP grant 2141064 and IAIFI grant PHY-2019786. -
IN
engels-2025-iclr-publisher
The paper 'Not All Language Model Features Are Linear' (Engels et al.) was published at ICLR 2025. -
IN
engels-2025-models-validated
The interpretability findings are validated across GPT-2, Mistral 7B, and Llama 3 8B model families, with references also to GPT-4 and Claude 3 indicating intended generality. -
IN
engels-2025-sae-clustering-extension
The paper extends SAE dictionary learning from Bricken et al. (2023) and Cunningham et al. (2023) by adding a clustering step over dictionary elements to recover multi-dimensional irreducible subspaces. -
IN
feature-hierarchy-empirical-validation-v2
The SAE decoder-space neighborhood structure exhibits a general-to-specific gradient—broader California landmarks (Lake Tahoe, Yosemite) appear near the Golden Gate Bridge feature, and the transit-infrastructure feature activates across diverse modalities (trains, ferries, wormholes), suggesting it encodes a higher-level transport abstraction rather than a single literal entity. This hierarchical, abstraction-to-concreteness pattern is consistent with the geometric property Park characterizes in LLaMA-3's representation space, where the (child − parent) vector for WordNet taxonomic relations is approximately orthogonal to the parent vector. The two observations arise in different representational spaces (SAE decoder neighborhoods vs. continuous LLM embeddings) and are best read as parallel evidence for a shared hierarchical organization rather than a direct instantiation of one by the other. -
IN
feature-neighborhood-as-geometric-theorem-instantiation-v2
SAE decoder-space feature neighborhoods (e.g., concrete features such as the Golden Gate Bridge grouped under more abstract categories like transit infrastructure) provide an empirical instantiation of the covariance-geometric semantic space at the interpretable level: decoder-space proximity mirrors subordination relations of the kind validated by Park's orthogonality theorem across Gemma-2B and LLaMA-3-8B, linking interpretability with the geometric account of semantic space. -
IN
feature-splitting-sae-width
As SAE width increases, a single coarse feature splits into multiple finer-grained interpretable sub-features (e.g., one base64 feature becomes three more specific base64 sub-features), indicating the smaller SAE was under-resolved -
IN
gpt2-small-sae-25k-features-layer7
GPT-2-small clustering uses approximately 25k SAE features from layer 7 (sourced from Bloom 2024) with spectral clustering at n_clusters=1000. -
IN
mistral-sae-1b-tokens-datasets
Mistral 7B SAEs are trained on more than 1 billion tokens drawn from the Pile and Alpaca datasets. -
IN
mistral-sae-65536-dict-elements
Mistral 7B SAEs use a 16× expansion factor over the model dimension, yielding 65,536 dictionary elements per layer. -
IN
mistral-sae-activation-normalization
Mistral 7B SAE activations are normalized to ‖x‖ = √d_model = 64, and no pre-encoder bias is used. -
IN
mistral-sae-l1-2-sparse-penalty
Mistral 7B SAEs use an L_{1/2} (half-norm) sparsity penalty with λ = 0.012, which produces sparser codes than L1. -
IN
mistral-sae-layers-8-16-24
Mistral 7B SAEs are trained on layers 8, 16, and 24 of the 32-layer model, targeting different abstraction levels. -
IN
mistral-sae-optimizer-dead-features
Mistral 7B SAEs use AdamW with weight decay 10⁻³, learning rate 2×10⁻⁴, and linear warmup; after 5× dead-feature resampling, approximately 1000 dead features remain out of 65,536. -
IN
nanda-decoder-weight-sparsity-distribution
Nanda's SAE replication found that 4% of decoder weights are well explained by a single neuron, 4% by 2-10 neurons, and 92% are dense across many neurons in the neuron basis -
IN
nanda-replication-codebase
Neel Nanda's open-source SAE replication codebase is hosted at https://github.com/neelnanda-io/1L-Sparse-Autoencoder and uses a 1-layer GELU language model -
IN
sae-ablation-clamping-zero-method
Feature ablation is performed by clamping a feature's activation to zero at a specific token position during a forward pass, measuring the full nonlinear causal effect on logit difference. -
IN
sae-activation-normalization-rule
Activations are scalar-normalized so their average squared L2 norm equals the residual stream dimension D before SAE training. -
IN
sae-algorithmic-improvement-bricken
A key algorithmic improvement in the scaling monosemanticity paper is multiplying the sparsity penalty by the decoder norm and removing the unit-norm constraint on decoder vectors, proposed by Bricken and verified by Conerly and Templeton. -
IN
sae-architecture-relu-encoder-linear-decoder
The SAE encoder uses a ReLU nonlinearity (f_i(x) = ReLU(W^{enc}_{i,·} · x + b^{enc}_i)) while the decoder is purely linear (x̂ = b^{dec} + Σᵢ f_i(x) · W^{dec}_{·,i}). -
IN
sae-attribution-ablation-0-8-correlation
In the worked example (John says 'I want to be alone' → John feels ___), ablating every active feature yielded a 0.8 correlation with attribution scores, validating attribution as a reasonable proxy. -
IN
sae-attribution-linear-formula
Attribution is computed as the dot product of the gradient of logit difference w.r.t. residual stream with (feature_vector × feature_activation), using a baseline of 0 rather than a second-prompt baseline. -
IN
sae-circular-feature-ambiguity
An SAE with m=2 dictionary features must activate both on every point of a circular feature (degenerate), while m=10 features spread around the circle with no canonical feature direction, meaning SAE feature directions for irreducible multi-dimensional features are not unique. -
IN
sae-claude-opus-as-interpretability-judge
Claude 3 Opus is used as the interpretability judge to rate SAE feature activations, distinct from Claude 3 Sonnet which is the model being analyzed, to reduce self-reference bias. -
IN
sae-cosine-similarity-defines-neighborhoods
Feature neighborhoods are defined by cosine similarity of feature vectors in SAE decoder space, and distance in decoder space approximately maps onto semantic relatedness in concept space. -
IN
sae-cross-modal-text-trained-image-activation
SAE features trained on text-only data also activate on relevant image inputs, indicating shared latent cross-modal representations in Claude 3 Sonnet. -
IN
sae-dead-feature-proportions-by-size
Dead feature proportions (zero activation across 10⁷ tokens) are approximately 2% for the 1M SAE, 35% for the 4M SAE, and 65% for the 34M SAE. -
IN
sae-dictionary-size-gpt2-mistral
GPT-2-small layer 7 has approximately 25,000 SAE dictionary elements (Bloom, 2024); Mistral 7B layer 8 has 216,000 SAE dictionary elements. -
IN
sae-encoder-nonnegative-steering-negative
The SAE encoder output is always nonnegative, but the steering procedure can clamp a feature's activity to negative values, which is equivalent to subtracting a multiple of the feature vector from the residual stream. -
IN
sae-expansion-ratio-10-to-200x
Sparse Autoencoders decompose dense d-dimensional activations into a typically 10× to 200× larger set of mostly-sparse, near-orthogonal latent features. -
IN
sae-expansion-ratio-definition
The SAE expansion ratio is defined as SAE latent dimension divided by model hidden dimension; a 32× expansion on a 4096-dimensional model yields a 131,072-dimensional sparse code. -
IN
sae-feature-granularity-scales-with-run-size
Larger SAE runs yield more specific features (e.g., 34M run: 'Golden Gate Bridge') while smaller runs yield broader categories (e.g., 1M run: 'Monuments and tourist attractions'). -
IN
sae-feature-id-notation-format
SAE features are identified by the format [SAE_size/feature_id], e.g., 34M/31164353 denotes feature index 31164353 in the 34M-feature training run. -
IN
sae-feature-search-four-methods
Anthropic's SAE feature search employs four distinct methods: single-prompt activation, multi-prompt positive/negative filtering, geometric cosine-similarity nearest-neighbor search, and logit-difference attribution. -
IN
sae-features-more-interpretable-than-neurons
Autointerpretability analysis (Bills et al. method applied by Cunningham) showed SAE-identified features are substantially more interpretable than individual transformer neurons -
IN
sae-golden-gate-bridge-neighborhood-structure
Features near the Golden Gate Bridge feature (34M/31164353) include San Francisco locations (Alcatraz, Presidio), California landmarks (Lake Tahoe, Yosemite, Solano County), and further-out tourist attractions (Médoc wine region, Isle of Skye). -
IN
sae-interpretability-rubric-0-3-scale
The automated interpretability rubric uses a 0–3 scale where 0 means feature is irrelevant to context and 3 means the feature cleanly identifies the activating text. -
IN
sae-jcr-prevents-feature-splitting
Jump Consistency Regularization (JCR) is a regularization term that discourages two nearby features from representing the same concept, preventing the 'feature splitting' pathology. -
IN
sae-loss-function-decoder-norm-penalty
The SAE loss includes an L1 penalty on feature activations weighted by decoder weight norm ‖W^{dec}_{·,i}‖₂ with coefficient λ=5, preventing the SAE from decoupling small activations from large weights. -
IN
sae-middle-layer-residual-stream-target
The SAE is applied to residual stream activations at the middle layer specifically to reduce dimensionality versus MLP output and to mitigate cross-layer superposition. -
IN
sae-neighborhood-as-directsum-navigation-v2
The SAE feature neighborhood (Golden Gate Bridge → Alcatraz → San Francisco → California) serves as an operational navigation algorithm for the categorical polytope geometry, with cross-model universality indicating a shared semantic object rather than a model-specific artifact. Combined with hierarchical orthogonality (Theorem 8), this suggests the full representation space may decompose as a direct sum of orthogonal subspaces, one per hierarchical level. However, the antecedents do not explicitly establish that the same neighborhood traversal simultaneously navigates both the categorical and hierarchical subspaces of such a decomposition; they confirm the categorical-polytope navigation and imply the structural decomposition separately. -
IN
sae-neighborhood-as-polytope-navigation-v2
SAE feature neighborhoods (e.g., Golden Gate Bridge → San Francisco → California) provide a concrete interpretable instantiation of the categorical polytope geometry identified by Park et al.: decoder-space proximity among SAE features reflects the same subordination relations that Park et al. formalize as polytope vertex adjacency, and cross-model feature similarity suggests this geometric structure reflects shared data properties rather than being a model-specific architectural artifact. -
IN
sae-probe-layer-transferability
The SAE-discovered plane probe trained on Mistral 7B layer 8 achieves an average logit difference of approximately −2.32 when applied to layer 6, while a raw PCA circular probe achieves only ≈ 0.029 (near-zero). -
IN
sae-raw-activation-not-causal-importance
Ranking features by raw activation magnitude does not reliably identify causally important features; a feature that fires on the token 'alone' is distinct from one encoding the concept of wanting solitude. -
IN
sae-reconstruction-variance-and-sparsity
All three SAEs (1M, 4M, 34M features) achieve ≥65% variance reconstruction with fewer than 300 active features per token. -
IN
sae-scaling-law-power-law-compute
SAE loss decreases approximately as a power law with respect to compute (features × training steps), and training runs for exactly one epoch so steps map linearly to data volume. -
IN
sae-shrinkage-problem
The SAE shrinkage problem refers to sparse autoencoders under-reconstructing the original activations, with mitigations including finetuning approaches (Wright & Sharkey) and gating activation functions (Rajamanoharan et al.). -
IN
sae-topk-architecture-used
Anthropic's scaling monosemanticity work on Claude 3 Sonnet uses TopK SAEs, which enforce sparsity more strictly than L1-regularized SAEs by retaining only the top-K activations in the sparse code. -
IN
sae-topk-sparsity-constraint
Top-k SAE variant enforces a hard sparsity constraint by allowing only the top k encoder features to fire per token, improving interpretability over L1-regularized vanilla SAEs. -
IN
sae-training-objective-mse-plus-l1
The SAE training objective minimizes reconstruction error (mean squared error) plus an L1 sparsity penalty on activations, with no principled 'ground truth' for the trade-off between the two terms. -
IN
sae-transit-feature-abstract-generalization
The transit infrastructure feature (1M/3) activates on wormholes alongside physical transit entities like trains, ferries, and tunnels, suggesting it captures a higher-level transit/transport concept rather than literal entities. -
IN
sae-umap-visualization-tool
An interactive UMAP visualization tool (linked as ./umap.html) is provided for exploring feature neighborhoods in the Sonnet SAE feature space. -
IN
sae-vs-fewshot-steering-comparison
In the scaling monosemanticity paper's tests, SAE-based features outperformed few-shot probe steering vectors in 5 of 7 test cases (secrecy, sycophancy, code errors, self-improving AI, methamphetamine), while both were effective for gender bias and agreement features. -
IN
saey-11-activation-intervals-spectrum-scoring
Spectrum-aware human scoring divides the activation range into 11 evenly spaced intervals and scores each interval separately against the overall feature hypothesis. -
IN
saey-256x-expansion-131072-features
A sparse autoencoder with a 256× expansion factor applied to a 512-neuron layer yields 131,072 learned features. -
IN
saey-412-intervals-scored-162-features-neurons
The human evaluation in the global interpretability analysis scored 412 feature activation intervals across 162 features and neurons. -
IN
saey-a1-autoencoder-4096-features-dead-ultralow
The A/1 autoencoder produces 4,096 features, of which 168 are dead (never active on 100M examples) and 292 are ultralow-density (active on fewer than 1 in 1M examples), leaving approximately 3,636 for analysis. -
IN
saey-activations-collected-from-8b-datapoints
SAE training activations were collected from 8 billion data points of the one-layer transformer's MLP layer. -
IN
saey-attention-only-rewritable-without-decomposition
In one-layer attention-only networks, the network can be rewritten to eliminate hidden-state references, bypassing the need for feature decomposition; this rewriting becomes impossible once a ReLU MLP is introduced. -
IN
saey-context-models-dna-base64-arabic-hebrew
Validation contexts are modeled with character-set rules: DNA as [ATCG], base64 as [a-zA-Z0-9+/], and Arabic/Hebrew as membership in specific Unicode block ranges. -
IN
saey-dictionary-sizes-a0-a1-a2
The dictionary sizes across models are: A/0 = 512 features, A/1 = 4,096 features, A/2 = 16,384 features. -
IN
saey-feature-browsing-interface-url
A public feature-browsing interface for qualitative inspection of individual features is available at transformer-circuits.pub/2023/monosemantic-features/vis/. -
IN
saey-feature-splitting-base64-three-subfeatures
As SAE width increases, a single 'base64' feature in a small dictionary decomposes into three more specific sub-features in a larger dictionary. -
IN
saey-five-validation-criteria-monosemantic-feature
The five validation criteria for a monosemantic feature are: high specificity, high sensitivity, causal downstream effect, non-correspondence to any single neuron, and replication in an independently trained model. -
IN
saey-log-likelihood-ratio-proxy-formula
The scoring proxy for feature specificity/sensitivity is the log-likelihood ratio log(P(s|context) / P(s)). -
IN
saey-mlp-n-token-conjunctions-vs-attention-3-point
Attention heads implement interactions over approximately 3 positions, while MLP layers are positioned to implement conjunctions over arbitrarily many tokens. -
IN
saey-over-100-token-in-context-features-the-a4
Model A/4 contains over 100 token-in-context features responding to the word 'the' in different contexts. -
IN
saey-paper-rejects-sparse-architecture-approach
The Bricken et al. 2023 paper explicitly rejects the 'sparse architecture' approach (encouraging activation sparsity at training time) as insufficient to eliminate polysemanticity. -
IN
saey-scoring-rubric-four-dimensions
The human scoring rubric has four dimensions: confidence in the explanation, consistency of activations with the explanation, consistency of logit output weights with the explanation, and specificity of the explanation. -
IN
saey-trigram-features-implemented-in-mlp
Trigram features (predicting a specific token given two preceding tokens, e.g., '19' in 'COVID-19') are implemented via MLP layers even though attention heads could in principle perform this function. -
IN
safety-feature-steering-causal-not-correlational
Steering experiments (positive and negative activation scaling) demonstrate that safety-relevant SAE features causally influence model generated output, not merely correlate with unsafe content. -
IN
tanh-l1-penalty-reduces-interpretability
Replacing the standard L1 sparsity penalty with a tanh L1 penalty improved proxy metrics but made SAE features less interpretable. -
IN
task-specificity-vs-feature-universality-v2
Evaluation results suggest embedding model quality is substantially task-specific (no single model across 33 evaluated models dominates all MTEB task categories), while SAE-extracted features tend to be model-independent (more similar across architectures than to their own model's neurons), pointing toward a distinction between output-space performance differentiation and internal representational convergence. -
IN
text-sae-generalizes-to-image-activations
SAEs trained purely on text activations generalize zero-shot to image activations (off-distribution), as reported by Anthropic in the scaling monosemanticity paper. -
IN
three-models-circular-pca-structure
In GPT-2, Mistral 7B, and Llama 3 8B, days of the week and months of the year appear as circular (conical) structures in SAE-reconstructed activations, with the first PCA dimension encoding radius ('intensity') and the circle living in PCA dimensions 2–3.