LoRA-TSD: Tangent-Space Spectral Descent for LoRA via Muon-Style Updates
This episode discusses the paper 'LoRA-TSD,' a new optimizer that addresses geometric disconnect in LoRA updates. The hosts explain how it uses Tangent-Space Spectral Descent and Muon-Style Updates to treat the update as movement along a specific curved path.The method is shown to be more stable and efficient than existing baselines across multiple benchmarks.