Pre-training sets low-rank geometry; fine-tuning cannot fix fundamental mistakes

yunta_tsai · x · 2026-08-21

The core insight is that pre-training establishes the correct low-rank geometry, while RL fine-tunes the shape. A better low-rank geometry requires less fine-tuning. Conversely, fine-tuning cannot fix low-rank mistakes if the observations have gaping holes, as the optimizer cannot deduce the correct shape.

Related event: Pretraining sets the low-rank geometry; RL only reshapes it(2 posts)→

Original post →

More from Research

Research channel →