By decomposing GPT-2's first-layer attention scores into token and position components, the authors show the model has a learned linear plus sinusoidal position bias toward nearby tokens and token affinities that recombine split words.
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
-
Weight-based Analysis of Detokenization in Language Models: Understanding the First Stage of Inference Without Inference
By decomposing GPT-2's first-layer attention scores into token and position components, the authors show the model has a learned linear plus sinusoidal position bias toward nearby tokens and token affinities that recombine split words.