Compact deletion masks retain source accuracy alongside selected additive controls
Can deletion tuning retain source capability after a finite development search when additive controls receive comparable selection? We compare strict binary gat
AI · Language model adaptation and interpretability
Can deletion tuning retain source capability after a finite development search when additive controls receive comparable selection? We compare strict binary gat
Deletion tuning can adapt a model without changing surviving pretrained values, but its practical benefit may depend on objective recipe, update location and re
Deletion tuning adapts frozen language models via binary masks $\mathbf{W}' = \mathbf{W} \odot \mathbf{M}$ ($\mathbf{M} \in \{0,1\}^{d_{\text{out}} \times d_{\t
Deletion tuning adapts a language model by learning a binary mask over pretrained gate_proj weights. Under the supervised fine-tuning protocol supplied with r/d
Deletion tuning adapts a language model by learning a binary mask over pretrained gate_proj weights. We previously showed that its apparent ceiling for memorizi
Does adapting a model only by deleting pretrained weights cap how much it can learn? We test this directly by memorizing N synthetic key-value facts (about 20 b
Can a model that learned by deleting pretrained weights selectively reverse one acquired group while retaining others? We evaluate exact coordinate rollback in