A new study published on arXiv systematically compares ten deep learning optimizers for language model pretraining, challenging the dominance of AdamW. The research highlights that previous comparisons were often flawed due to unequal hyperparameter tuning and limited evaluation setups. While matrix-based optimizers like Muon and Soap show promise, their speedup over AdamW diminishes significantly with increasing model scale, offering only a 1.1x advantage for models up to 1.2 billion parameters. AI
IMPACT Suggests that optimizers other than AdamW may offer speedups in language model pretraining, though the benefits decrease with model scale.
RANK_REASON Academic paper detailing a systematic study of deep learning optimizers. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
