Entity · paper

Sparse Autoencoders Encode Both Concepts and Functions: The Downstream Geometry of Feature Effects

paperactiveprovisionalsparse-autoencoders-encode-both-concepts-and-functions-the-downstream-geometry-of-feature-effects-9846f45c·1 events·first seen 2d ago

Aliases: Sparse Autoencoders Encode Both Concepts and Functions: The Downstream Geometry of Feature Effects

Co-occurring entities

Feature-Effect Geometry Analysis

More like this (12)

Sparse Autoencoders Feature Auto-Encoder Unstable Features, Reproducible Subspaces: Understanding Seed Dependence in Sparse Autoencoders Sparse Autoencoders (SAEs)Sparse Autoencoder Graph Neural Network Encoder Cross-seed explainability using Procrustes-conditioned Joint End-to-end Top-K Sparse Autoencoders Natural Language Autoencoders conditional variational autoencoder C²R: Cross-sample Consistency Regularization Mitigates Feature Splitting and Absorption in Sparse Autoencoders Your UnEmbedding Matrix is Secretly a Feature Lens for Text Embeddings Optimization Dynamics Imprint Semantic Specificity in Contrastive Embedding Norms

Recent events (1)

6arXiv · cs.CL·2d ago·source ↗

FEGA framework reveals SAE features encode both concepts and context-dependent functions, complicating steering

A new arXiv preprint introduces Feature-Effect Geometry Analysis (FEGA), an unsupervised framework that studies how sparse autoencoder (SAE) feature interventions affect model logits rather than examining internal feature geometry. The authors find that consistent one-dimensional causal effects are rare across SAE features, and distinguish 'value-like' features (tied to static factual attributes) from 'pointer-like' features (context-dependent operations), with the latter exhibiting diffuse, hard-to-steer effects. The key finding is that a feature can be interpretable and causally relevant without providing a stable steering direction, which has significant implications for mechanistic interpretability workflows that rely on SAE-based steering.

AI Safety Research Sparse Autoencoders Encode Both Concepts and Functions: The Downstream Geometry of Feature Effects Feature-Effect Geometry Analysis