SVD initialization of LoRA accelerates RL training convergence

Trajectory Labs shows that initializing LoRA adapters with top-k SVD components of pretrained weights (PiSSA), instead of the community's common approach, speeds up RL training convergence and improves downstream performance.

2026-09-17 ~ 2026-09-17 · 2 related posts

• PiSSA-style SVD init for LoRA adapters also improves downstream RL training, Trajectory Labs reports — simonguozirui · 2026-09-17
• LoRA weights have been initialized suboptimally: SVD-based init speeds up RL training — burny_tech · 2026-09-17