The Surprising Effectiveness of Deleting Weights in LLM Post-Training (arxiv.org)

A recent study has introduced a novel approach called Deletion Tuning, which reveals that simply zeroing out less than 0.05% of the weights in a pretrained large language model (LLM) can achieve the same effectiveness as full fine-tuning (FFT) across various post-training scenarios. This method focuses on learning a binary mask for the gate projection of transformer blocks, demonstrating impressive efficiency by significantly reducing the storage requirements for model checkpoints—by hundreds to thousands of times compared to conventional fine-tuning methods. This development is particularly significant for the AI/ML community as it offers a solution to the continual learning challenge, allowing models to maintain prior-task performance while acquiring new skills. The study also delves into the geometric analysis of weight changes, revealing that the masking strategy effectively captures relevant information without the extensive modifications required by existing fine-tuning techniques. As a result, Deletion Tuning could change how developers approach model updates, making it a promising alternative for enhancing LLM capabilities without the computational burden usually associated with training.