A well-trained next-token predictor's in-context loss equals the conditional entropy of the data process, which must decrease with context for stationary data.
Transformers learn in-context by gradient descent
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
other 1
citation-polarity summary
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1roles
other 1polarities
unclear 1representative citing papers
citing papers explorer
-
Next-token pretraining implies in-context learning
A well-trained next-token predictor's in-context loss equals the conditional entropy of the data process, which must decrease with context for stationary data.