Deletion Tuning adapts models by zeroing fewer than 0.05% of weights

Removing a tiny fraction of weights can help a model adapt. Deletion Tuning researchers report matching full fine-tuning in tested post-training settings by zeroing fewer than 0.05% of weights.

This is not random deletion or training-free improvement. The method learns a mask selecting which positions to zero.

It raises a counterintuitive question: can improvement sometimes come from disabling connections rather than rewriting them? The evidence concerns the evaluated settings.

Broader replication could change how adaptations are stored and distributed. It is not a universal shortcut, and does not eliminate base-model pretraining.

Source version: arXiv, November 8, 2026, 01:33:20 Beijing / November 7, 12:33:20 EST. An earlier OpenReview version exists; earliest public date remains unverified.