Individual transformer dimensions encode concepts via signs alone, enabling training-free detection and steering without learned feature dictionaries.
Sparse autoencoders can interpret randomly initialized transformers
7 Pith papers cite this work. Polarity classification is still indexing.
7
Pith papers citing it
years
2026 7representative citing papers
Many distinct SAE features share identical explanations, with the average annotation resolving only 70% of feature identity in a large annotated dataset.
SAEs used for layer selection with raw task vectors outperform subspace projection and raise math reasoning accuracy on Gemma-3-4B-IT.
Stable SAE features dominate functional signal while unstable features concentrate in reproducible subspaces, allowing more stable SAEs via cross-seed pooling.
Sparse autoencoder analysis of PatchTST FFN activations shows sparse, stable representations with no empirical support for superposition on standard time series forecasting tasks.
Linear probes show rhetorical questions are encoded via multiple dataset-specific directions in LLM representations, with low cross-probe agreement on the same data.
A 53K-parameter weight-shared transformer generates novel valid SMILES at 95% rate on ZINC-250K and resolves constraints hierarchically via bracket, ring, and valence stages as shown by probing and ablation.
citing papers explorer
-
The Signs Were Always There: Training-Free Concept Detection and Steering in Raw Transformer Dimensions
Individual transformer dimensions encode concepts via signs alone, enabling training-free detection and steering without learned feature dictionaries.
-
Descriptive Collision in Sparse Autoencoder Auto-Interpretability: When One Explanation Describes Many Features
Many distinct SAE features share identical explanations, with the average annotation resolving only 70% of feature identity in a large annotated dataset.
-
Interpretability-Guided Layer Selection over Subspace Projection: SAEs as Stethoscopes, Not Scalpels, for Raw Task Vector Model Editing
SAEs used for layer selection with raw task vectors outperform subspace projection and raise math reasoning accuracy on Gemma-3-4B-IT.
-
Unstable Features, Reproducible Subspaces: Understanding Seed Dependence in Sparse Autoencoders
Stable SAE features dominate functional signal while unstable features concentrate in reproducible subspaces, allowing more stable SAEs via cross-seed pooling.
-
Superposition Is Not Necessary: A Mechanistic Interpretability Analysis of Transformer Representations for Time Series Forecasting
Sparse autoencoder analysis of PatchTST FFN activations shows sparse, stable representations with no empirical support for superposition on standard time series forecasting tasks.
-
Rhetorical Questions in LLM Representations: A Linear Probing Study
Linear probes show rhetorical questions are encoded via multiple dataset-specific directions in LLM representations, with low cross-probe agreement on the same data.
-
SMolLM: Small Language Models Learn Small Molecular Grammar
A 53K-parameter weight-shared transformer generates novel valid SMILES at 95% rate on ZINC-250K and resolves constraints hierarchically via bracket, ring, and valence stages as shown by probing and ablation.