Sparse autoencoders frequently capture low-dimensional correlates that co-occur with reasoning rather than core reasoning mechanisms, as demonstrated by token-level interventions and falsification across 22 model-layer-dataset configurations.
First, I need to recall the secret message
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Do Sparse Autoencoders Identify Reasoning Features in Language Models?
Sparse autoencoders frequently capture low-dimensional correlates that co-occur with reasoning rather than core reasoning mechanisms, as demonstrated by token-level interventions and falsification across 22 model-layer-dataset configurations.