Rethinking Local Learning: A Cheaper and Faster Recipe for LLM Post-Training

TL;DR

LoPT introduces a local learning strategy for LLM post-training that reduces memory and computational costs by placing a gradient boundary at the transformer midpoint, enabling efficient task adaptation.

Contribution

The paper proposes a novel local learning approach, LoPT, which simplifies and accelerates LLM post-training by decoupling early and late layer updates with a gradient boundary.

Findings

LoPT achieves competitive performance with less memory usage.

LoPT improves training efficiency compared to full-depth backpropagation.

LoPT better preserves pretrained capabilities during post-training.

Abstract

LLM post-training typically propagates task gradients through the full depth of the model. Although this end-to-end structure is simple and general, it couples task adaptation to full-depth activation storage, long-range backward dependencies and direct task-gradient access to pretrained representations. We argue that this full-depth backward coupling can be unnecessarily expensive and intrusive, particularly when post-training supervision is much narrower than pre-training. To this end, we propose \textbf{LoPT}: Local-Learning Post-Training, a simple post-training strategy that makes gradient reach an explicit design choice. LoPT places a single gradient boundary at the transformer midpoint: the second-half block learns from the task objective, while the first-half block is updated by a lightweight feature-reconstruction objective to preserve useful representations and maintain interface compatibility. LoPT shortens the task-induced backward path while limiting direct interference from narrow task gradients on early-layer representations. Extensive experiments demonstrate that LoPT achieves competitive performance with lower memory cost, higher training efficiency and better retention of pretrained capabilities. Our code is available at: https://github.com/HumyuShi/LoPT