TopK Sparse Autoencoders Across Three Model Architectures: Dictionary Collapse, Dense-Feature Degeneracy, and the Limits of Activation-Pattern Feature Matching

TopK SAEs trained at matched sparsity and dictionary size on Llama-3.2-3B, Mistral-7B-v0.3 and Qwen2.5-3B. Reports abrupt dictionary collapse with partial recovery, dense-feature degeneracy invalidating frequency-based feature selection, and a null cross-architecture matching result whose detection floor is measured by planted-signal power analysis (MDES 0.95-1.00).

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-08-14
DOI
https://doi.org/10.5281/zenodo.21906712
Primary Topic
Domain Adaptation and Few-Shot Learning
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
preprint

TopK Sparse Autoencoders Across Three Model Architectures: Dictionary Collapse, Dense-Feature Degeneracy, and the Limits of Activation-Pattern Feature Matching

J. Melton
Zenodo (CERN European Organization for Nuclear Research)
Domain Adaptation and Few-Shot Learning
preprint

TopK Sparse Autoencoders Across Three Model Architectures: Dictionary Collapse, Dense-Feature Degeneracy, and the Limits of Activation-Pattern Feature Matching

J. Melton
preprint en

Abstract

TopK SAEs trained at matched sparsity and dictionary size on Llama-3.2-3B, Mistral-7B-v0.3 and Qwen2.5-3B. Reports abrupt dictionary collapse with partial recovery, dense-feature degeneracy invalidating frequency-based feature selection, and a null cross-architecture matching result whose detection floor is measured by planted-signal power analysis (MDES 0.95-1.00).

Zenodo (CERN European Organization for Nuclear Research)
American Standard (United States) (US)
Sustainable cities and communities
Domain Adaptation and Few-Shot Learning
DOIOpen Access PDFOpenAlex
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

TopK Sparse Autoencoders Across Three Model Architectures: Dictionary Collapse, Dense-Feature Degeneracy, and the Limits of Activation-Pattern Feature Matching — J. Melton · Zenodo (CERN European Organization for Nuclear Research) (2026) | TGRS Research Map | TGRS