The Surprising Effectiveness of Deleting Weights in LLM Post-Training

Zeroing fewer than 0.05% of a pretrained large language model's weights, with no other modification, matches full fine-tuning at scale across three LLM post-training regimes, including reinforcement learning with verifiable rewards, according to the researchers behind the finding. The result suggests new capabilities can be acquired by changing only a tiny fraction of a model's parameters rather than retraining the whole network.

How little of a pretrained large language model LLM has to change for it to acquire new capabilities? We find that zeroing fewer than 0.05% of its weights, with no other modification, matches full fine-tuning FFT at scale across three LLM post-training regimes: reinforcement learning with verifi