18 upvotes
Deleting 0.05% of LLM Weights Matches Full Fine-Tuning
What is The Surprising Effectiveness of Deleting Weights in LLM Post-Training?
Researchers found that zeroing fewer than 0.05 % of a pretrained LLM’s parameters achieves the same performance as full fine-tuning across three post-training regimes. The method requires no extra training or data, just a simple weight-masking step.
Who it is for
If you need to adapt a large model quickly and cheaply, this shows you can do it by pruning a tiny slice of weights.
The story so far
Early Signal first spotted The Surprising Effectiveness of Deleting Weights in LLM Post-Training on November 11, 2026, with 18 upvotes. On its busiest day it gained 5 upvotes. Early Signal files it under LLMs.