---
title: "The Surprising Effectiveness of Deleting Weights in LLM Post-Training"
slug: the-surprising-effectiveness-of-deleting-weights-in-llm-post-training
url: https://listedarticles.com/articles/the-surprising-effectiveness-of-deleting-weights-in-llm-post-training
canonical_url: https://arxiv.org/abs/2611.10866
content_type: research
language: en
published_at: 2026-11-07T00:00:00.000Z
updated_at: 2026-11-11T08:12:38.940Z
author: "Jack Lu et al."
authored_by: human
publisher: "arXiv"
publisher_url: https://arxiv.org/
topics: ["machine-learning", "llms", "research"]
license: all-rights-reserved
word_count: 208
reading_minutes: 1
citation: "Jack Lu et al., arXiv. \"The Surprising Effectiveness of Deleting Weights in LLM Post-Training.\" 7 Nov 2026. https://arxiv.org/abs/2611.10866 (all-rights-reserved)"
---

# The Surprising Effectiveness of Deleting Weights in LLM Post-Training

> NeurIPS 2026 paper introducing Deletion Tuning: zeroing fewer than 0.05% of an LLM's weights via a learned binary mask on gate projections matches full fine-tuning across RLVR, SFT and on-policy distillation, with far smaller checkpoints and much less forgetting.

# The Surprising Effectiveness of Deleting Weights in LLM Post-Training

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

arXiv:2611.10866 [cs.LG] · Submitted 7 Nov 2026 · NeurIPS 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 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.

## Links

- [Abstract page on arXiv](https://arxiv.org/abs/2611.10866)
- [PDF](https://arxiv.org/pdf/2611.10866)
- DOI: [10.48550/arXiv.2611.10866](https://doi.org/10.48550/arXiv.2611.10866)
