TopK (k-sparse) Sparse Autoencoders

methods / Dictionary Learning / TopK (k-sparse) Sparse Autoencoders

Techniqueadvanced

Replaces a standard SAE's ReLU-plus-L1-penalty with a hard TopK activation that keeps exactly the k largest pre-activations and zeroes the rest, fixing sparsity directly (removing the L1-shrinkage bias) and enabling a dead-latent-revival auxiliary loss (AuxK) that keeps dead-latent fraction low even at 16-million-latent scale, yielding clean joint scaling laws in autoencoder size and sparsity.

Used in (13 observations)

TopK SAE features on real Whisper and HuBERT encoders are seed-stable and sparsely erasable, and steering them causally cuts Whisper false-speech detections by 70%measured

structure: Linear Direction · models: Whisper small, HuBERT-base · paper: AudioSAE: Towards Understanding of Audio-Processing Models with Sparse AutoEncoders

Clinical concept centers: a residual-stream direction per ICD-9 disease chapter, built as a decoded SAE difference of means, is held-out causally steerable in eleven open Qwen and Gemma modelsmeasured

structure: Linear Direction · models: Qwen3-8B, Qwen3.5-9B-Base, Qwen3-30B-A3B-Base, Qwen3.5-27B, II-Medical-8B, Gemma 3 4B, Gemma 3 4B Instruct, Gemma 3 27B, Gemma 3 27B Instruct, MedGemma-4B-it, MedGemma-27B-text-it · paper: Clinical Concept Centers in LLMs

Same concept, different SAE directions across modalitiesmeasured

structure: Linear Direction, Anisotropy · models: CLIP ViT-B/32, MetaCLIP ViT-B/32, OpenCLIP ViT-B/32, SigLIP 2 · paper: Same Concept, Different Directions: Cross-Modal Feature Heterogeneity in Sparse Autoencoders

Dimension-level activation outliers, measured as the ratio gamma of mean norm to per-token variation, predict sparse-autoencoder feature death across 454 model-layer combinations spanning language, vision, protein and genomic modelsmeasured

structure: Anisotropy · models: GPT-2, Pythia-410M, Pythia-70M, ESM-2 (3B), ESM-2 (650M), ESM-2 (35M), ProGen2-base, DINOv3 ViT-B/16, DINOv2 ViT-B · paper: On the Relationship Between Activation Outliers and Feature Death in Sparse Autoencoders

A single SAE latent causally mediates refusal with a dose-response tradeoffmeasured

structure: Linear Direction · models: Phi-3-mini, Llama-3.1-8B-Instruct · paper: Steering Language Model Refusal with Sparse Autoencoder Features

SAE decoder cosine reveals feature splitting in dense embeddings and steers searchmeasured

structure: Linear Direction · models: text-embedding-3-small · paper: Disentangling Dense Embeddings with Sparse Autoencoders

Monosemantic Whisper SAE latents encode phonetic/lexical features and steer transcriptsmeasured

structure: Linear Direction · models: Whisper base · paper: Mechanistic Interpretability of ASR models using Sparse Autoencoders

A Top-k SAE on sentence embeddings gives near-orthogonal, steerable latentsmeasured

structure: Linear Subspace · models: E5-large-v2 · paper: Aligning Sentence Embeddings to Human Concepts via Sparse Autoencoders

TopK SAEs give clean scaling laws and mitigate dead latentsmeasured

structure: Linear Direction · models: GPT-2 Small, GPT-4 · paper: Scaling and Evaluating Sparse Autoencoders

TopK sparse-autoencoder features on a real fine-tuned ESM2 align with real protein structural sites, and steering them causally improves in-silico protein designmeasured

structure: Linear Direction · models: ESM-2 (650M) · paper: Sparse Autoencoders for Low-N Protein Function Prediction and Design

Jacobian centroids (row-sums of a sub-component's local input-output Jacobian) form cleaner, more transferable linear feature directions than raw activations across four real-model settings: ResNet50/DTD, DINOv2-vs-DINOv3 SAE dictionaries, GPT2-Large circuit discovery, and Llama-3.1-8B truth probesmeasured

structure: Linear Centroids Hypothesis · models: ResNet-50 (supervised, ImageNet), DINOv2 ViT-L/14, DINOv3 ViT-B/16, GPT-2-Large, Llama-3.1-8B · paper: The Linear Centroids Hypothesis: Features as Directions Learned by Local Experts

Sparse-autoencoder decompositions of real neural audio codecs encode accent in activation magnitude or position, depending on codec typemeasured

structure: Linear Separability · models: DAC (Descript Audio Codec), SpeechTokenizer · paper: Towards Interpretable Framework for Neural Audio Codecs via Sparse Autoencoders: A Case Study on Accent Information

Agentic-trait probe directions share R^2 yet are orthogonal, revealing one agency axismeasured

structure: Linear Direction · models: Qwen3.5-35B-A3B (MoE, ~3B active/token) · paper: Behavioral Steering in a 35B MoE Language Model via SAE-Decoded Probe Vectors: One Agency Axis, Not Five Traits