Towards Efficient Optimizer Design for LLM via Structured Fisher Approximation with a Low-Rank ExtensionJan 1, 2025ยทWenbo Gong,M Scetbon,C Ma,E Meedsยท 0 min read PDF Cite Source DocumentTypePreprintPublicationarXiv:2502.07752Last updated on Jan 1, 2025Optimization Learning Dynamics LLM AuthorsWenbo GongSenior ResearcherSenior Researcher at Microsoft Research Cambridge working on learning dynamics and optimization for foundation models, with prior work on causality and approximate inference. ← SWAN: SGD with Normalization and Whitening Enables Stateless LLM Training Jan 1, 2025Deep end-to-end causal inference Jan 1, 2024 →