Only models trained from scratch with latent thoughts exhibit superposition; pretrained and fine-tuned models do not and instead rely on shortcuts due to pretraining bias and capacity limits.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
The Illusion of Superposition? A Principled Analysis of Latent Thinking in Language Models
Only models trained from scratch with latent thoughts exhibit superposition; pretrained and fine-tuned models do not and instead rely on shortcuts due to pretraining bias and capacity limits.