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

**Authors:** Jack Lu, Zhenbang Yang, Mike Lasby, Tejas Pote, Yani Ioannou, Mengye Ren  
**Lab:** agentic learning ai lab · **Published:** 6 Nov 2026 · **Upvotes:** 19 · **Level:** Intermediate  
**Links:** [arXiv](https://arxiv.org/abs/2611.10866) · [PDF](https://arxiv.org/pdf/2611.10866) · [Code](https://github.com/agentic-learning-ai-lab/deleting-weights) · [This page](https://omen-bit.github.io/ai-paper-daily/paper/2611.10866.html)

## Curator's notes

**TL;DR:** Zeroing less than 0.05% of LLM weights via a learned binary mask matches full fine‑tuning across several post‑training tasks.

- **What's new:** Introduces Deletion Tuning, a mask‑only method that achieves full fine‑tuning performance while using a tiny checkpoint and causing minimal forgetting.
- **Key result:** Zeroing fewer than 0.05% of weights matches full fine‑tuning performance across reinforcement learning, supervised fine‑tuning, and on‑policy distillation.
- **Why it matters:** Developers can adapt large models with negligible storage and reduced catastrophic forgetting.

## Key numbers

- **0.05%**: We find that zeroing fewer than 0.05% of its weights

## 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.

## What you could build

1. Create a tool that generates deletion masks for custom downstream tasks.
2. Implement a continual‑learning system that swaps masks instead of full model updates.
3. Build a lightweight model‑serving pipeline that stores only masks for multiple task versions.

## Quiz yourself

What proportion of weights does Deletion Tuning zero to match full fine‑tuning performance?

**Answer:** It zeros fewer than 0.05% of the model's weights.

---

Notes written by AI from the paper, numbers copied from the abstract. Check the paper before you rely on them. From AI Papers Daily.