eNTK theory from BERT era applied to modern LLM pretraining, adopted by frontier labs

prfsanjeevarora · x · 2026-08-01

Sadhika Malladi tweeted that ideas from their BERT-era eNTK paper (arXiv:2210.05643) have been applied to modern LLM pretraining and inspired frontier labs. The paper investigates whether Neural Tangent Kernel (NTK) describes fine-tuning of pre-trained LMs, extends to Adam, and validates on 14 NLP tasks. A quoted tweet notes that linearizing a 150M model with a quadratic tracks true loss for up to 10% of training budget.

Original post →

More from Research

Research channel →