Deletion Tuning Matches Full Fine-Tuning by Zeroing Less Than 0.05% of LLM Weights

·2026.11.11 04:44

Key point

The method learns a binary mask over gate projections, resulting in checkpoints hundreds to thousands of times smaller than full fine-tuning.

Zeroing fewer than 0.05% of pretrained LLM weights matches full fine-tuning (FFT) performance across reinforcement learning, supervised fine-tuning, and on-policy distillation.
The trained binary mask is hundreds to thousands of times smaller than the corresponding FFT checkpoint.
Deletion Tuning retains prior-task accuracy near base levels under reinforcement learning, whereas FFT, LoRA, and random sparse fine-tuning lose up to 5.4 points.
Geometric analysis using stable rank, spectrum drift, and principal-weight overlap shows Deletion Tuning is closest to FFT compared to LoRA and random sparse fine-tuning.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.