Chain-of-thought reasoning in DeepSeek-R1-Distill-Qwen-7B depends on a small set of high-intensity sparse features; suppressing them breaks both reasoning and LaTeX formatting.
Finding sparse autoencoder representations of errors in cot prompting
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Thinking vs. NoThinking: Towards Interpreting Reasoning Mechanisms of Large Language Models via Sparse Autoencoders
Chain-of-thought reasoning in DeepSeek-R1-Distill-Qwen-7B depends on a small set of high-intensity sparse features; suppressing them breaks both reasoning and LaTeX formatting.