01 / AI · Claim vs. Check · Analysis
Weight deletion delivers gains on par with full fine-tuning
Researchers reported lower storage needs for training results alongside preserved existing abilities.

Researchers at New York University and Canada's University of Calgary reported improved maths reasoning and tool use after deleting less than 0.05% of model weights, according to AI Times, a Korean AI news publication.
AI Times also reported that the researchers substantially reduced the storage needed for training results. The research addressed both performance gains and the amount of material retained after training.
According to the report, the researchers achieved performance comparable to full fine-tuning, which adjusts all of a model's weights. The comparison centred on whether selective deletion could deliver similar results while requiring less storage for training outputs.
Less storage for training results
According to AI Times, the researchers selectively removed weights—the numerical parameters a model learns during training—without modifying their values. This distinguishes the approach from adjusting the model's entire set of weights.
The storage reduction described in the report concerned the space needed to save training results. It did not establish that the model as a whole became smaller.
Performance compared with full fine-tuning
AI Times reported that the joint university team achieved results comparable to full fine-tuning. The reported comparison concerned performance after training, rather than the proportion of weights deleted alone.
According to the report, the researchers also retained abilities the model had previously acquired. Taken together, the reported findings linked selective deletion with comparable performance, smaller stored training results and preservation of existing capabilities.