LoRA's first step isn't standard SGD: an implicit low-rank preconditioner throttles early fine-tuning
pbaylies · x · 2026-09-07
A technical thread argues that LoRA does not perform standard SGD on your weights.
- At step 0, LoRA runs gradient descent distorted by an implicit low-rank preconditioner: P = alpha^2 A^T A.
- This random bottleneck quietly throttles early-stage fine-tuning gradients.
- The takeaway: understanding LoRA requires accounting for this preconditioning structure rather than treating it as plain low-rank SGD on weights.
Related event: New Research Explains LoRA's Slow Convergence, Proposes One-Line NoRA Fix(3 posts)→
More from Research
- Meta FAIR's Self-Direction turns agent failures into reusable LoRA skills, +13.8pp on Terminal-Bench — karen_ullrich · 2026-10-27
- DONIC hits 86.79% on bimanual dexterous tracking, crushing prior SOTA's 9.43% — liuziwei7 · 2026-10-27
- Why don't more LLMs ship 10M-token context? A researcher asks about attention tradeoffs — agarwl_ · 2026-10-27
- PSNR below 15 dB in papers likely means the metric was misused, researcher warns — kwangmoo_yi · 2026-10-27
- Ortet's first health eval: 34 open-weight models, 50 tasks, generic rankings fall short — thegautamkamath · 2026-10-27
- Scaling-fit discrepancy: output-head FLOPs are ~65% of compute at the smallest rung, ~3% at 535B — dlwh · 2026-10-27