Perplexity is order-invariant by the chain rule for a true distribution, yet GPT-2 models trained on forward, backward, and permuted text deviate systematically, with permutation training producing the largest gaps.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Probability Consistency in Large Language Models: Theoretical Foundations Meet Empirical Discrepancies
Perplexity is order-invariant by the chain rule for a true distribution, yet GPT-2 models trained on forward, backward, and permuted text deviate systematically, with permutation training producing the largest gaps.