Study quantifies how small weight updates reshape language model behavior across training stages
A new arXiv paper reports that post-training shifts a language model's weights by roughly a thousandth of their norm yet sharply changes behavior, while pretraining moves them by about a third of their norm.
Read the full story at arXiv cs.LG ↗
Timeline · 1 report
- 2026-10-30 04:00 · arXiv cs.LG
What a Weight Update Buys: Leverage Across Training Stages