A bias-correction framework for stochastic preconditioned optimizers (AdamW, Sophia, Shampoo) using cross-fitted microbatches and delta-method inversion correction yields 0.07-0.15 nat loss reductions on Qwen2.5-0.5B pretraining.
International Conference on Learning Representations , year =
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Correcting Stochastic Update Bias in Preconditioned Language Model Optimizers
A bias-correction framework for stochastic preconditioned optimizers (AdamW, Sophia, Shampoo) using cross-fitted microbatches and delta-method inversion correction yields 0.07-0.15 nat loss reductions on Qwen2.5-0.5B pretraining.