Space & aviationResearch findings
Spectral and Subspace Dynamics in LoRA Across SFT and RL
Original published 2026.11.06 03:17 KSTAdded to Blackdeck 2026.11.09 14:07 KST
What has changed?
In RL, the spectrum concentrates at all learning rates, while SFT expands it at moderate rates; RL also rotates singular subspaces faster than SFT at low rates.
Key points
- SFT and RL show distinct spectral dynamics in LoRA updates.
- RL concentrates the spectrum, while SFT expands it at moderate learning rates.
- RL rotates singular subspaces faster than SFT at low learning rates.
Connect it to your work
Group-mean baseline subtraction is identified as the root cause of these dynamics.
What remains unproven
This research note is based on an abstract; check the original for test conditions and application scope.
Originals used for this note
Links open the original publisher in a new tab.