Report
New paper examines sparse LLM parameter updates during post-training
The authors say they use Fisher information to connect theory with empirical findings across RL, SFT and on-policy distillation.
TLDR
The authors share a new paper examining why large language model parameter updates are sparse during post-training. They say it studies the pattern through Fisher information, connecting theory and empirical findings across reinforcement learning, supervised fine-tuning and on-policy distillation.
Combined views
32.2K
5 Sources, first seen ago
