Paper Detail

The Surprising Effectiveness of Deleting Weights in LLM Post-Training

Jack Lu, Zhenbang Yang, Mike Lasby, Tejas Pote, Yani Ioannou, Mengye Ren

huggingface Score 13.9

Published 2026-11-07 · First seen 2026-11-11

Research Track A · General AI

Abstract

How little of a pretrained large language model (LLM) has to change for it to acquire new capabilities? We find that zeroing fewer than 0.05% of its weights, with no other modification, matches full fine-tuning (FFT) at scale across three LLM post-training regimes: reinforcement learning with verifiable rewards, supervised fine-tuning, and on-policy distillation. Our method, Deletion Tuning, learns a binary mask over the gate projection of transformer blocks. Two practical advantages follow: the trained mask is hundreds to thousands of times smaller than the corresponding FFT checkpoint, and Deletion Tuning forgets far less, which addresses a central challenge in continual learning. Under reinforcement learning, its prior-task accuracy stays near the base level while FFT, LoRA, and random sparse fine-tuning lose up to 5.4 points. To understand why masking alone suffices, we analyze the geometry of the learned weight changes under reinforcement learning using stable rank, spectrum drift, and principal-weight overlap. Compared with LoRA and random sparse fine-tuning, Deletion Tuning is closest to FFT across the three measures.

Workflow Status

Review status
pending
Role
unreviewed
Read priority
now
Vote
Not set.
Saved
no
Collections
Not filed yet.
Next action
Not filled yet.

Reading Brief

No structured notes yet. Add `summary_sections`, `why_relevant`, `claim_impact`, or `next_action` in `papers.jsonl` to enrich this view.

Why It Surfaced

No ranking explanation is available yet.

Tags

No tags.

BibTeX

@misc{lu2026surprising,
  title = {The Surprising Effectiveness of Deleting Weights in LLM Post-Training},
  author = {Jack Lu and Zhenbang Yang and Mike Lasby and Tejas Pote and Yani Ioannou and Mengye Ren},
  year = {2026},
  abstract = {How little of a pretrained large language model (LLM) has to change for it to acquire new capabilities? We find that zeroing fewer than 0.05\% of its weights, with no other modification, matches full fine-tuning (FFT) at scale across three LLM post-training regimes: reinforcement learning with verifiable rewards, supervised fine-tuning, and on-policy distillation. Our method, Deletion Tuning, learns a binary mask over the gate projection of transformer blocks. Two practical advantages follow: the},
  url = {https://huggingface.co/papers/2611.10866},
  keywords = {code available, huggingface daily},
  eprint = {2611.10866},
  archiveprefix = {arXiv},
}

Metadata

{}